System

Abstract

A system is provided.SOLUTION: A data collection means for collecting financial data, audit reports, news articles, and historical embezzlement case data of a company from online and offline sources, a data preprocessing means for cleaning the collected data, complementing missing data, and normalizing the data, a machine learning model construction means for dividing the preprocessed data into a training set, a test set, and a validation set and applying a machine learning algorithm to train a model, and a fraud detection means for inputting latest financial data and comparing the latest financial data with past data to predict the possibility of fraud; A system includes a prediction means using a generated AI and a notification means for notifying an auditor of a specific suspicious point in real time together with a prediction result.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The technology of the present disclosure relates to a system. [Background technology]

[0002] Patent document 1 discloses a persona chatbot control method performed by at least one processor, the method including the steps of receiving a user utterance, adding the user utterance to a prompt including an instruction sentence related to a description of the chatbot character, encoding the prompt, and inputting the encoded prompt into a language model to generate a chatbot utterance in response to the user utterance. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2022-180282 Summary of the Invention [Problem to be solved by the invention]

[0004] Corporate accounting fraud undermines the credibility of financial reporting and is a problem that has a significant impact on investors and stakeholders. Traditional audit methods require manual analysis of vast amounts of data, which is time-consuming and labor-intensive, and there is a high possibility that signs of fraud may be overlooked. This has created a need for a system that can efficiently and accurately detect accounting fraud and respond quickly. [Means for solving the problem]

[0005] The present invention solves the aforementioned problems by providing a system that includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, and a notification means. This system collects a company's financial data, audit reports, news articles, and data on past window dressing cases, cleans the data, fills in missing data, and normalizes the data. It then uses the preprocessed data to construct a machine learning model and uses generative AI to predict the possibility of window dressing based on the latest financial statement data. Furthermore, by notifying accounting auditors of the prediction results and specific suspicious points in real time, it is possible to efficiently and accurately detect window dressing and respond quickly.

[0006] "Data collection means" refers to the ability to collect corporate financial data, audit reports, news articles, and past accounting fraud cases from online and offline sources.

[0007] "Data preprocessing means" refers to a series of processing procedures for cleaning collected data, filling in missing data, and normalizing data.

[0008] "Machine learning model building means" refers to the techniques and methods for dividing pre-processed data into training, test, and validation sets, and applying machine learning algorithms to train and evaluate models.

[0009] "Predictive methods using generative AI" refers to the generative AI technology and process used to input the latest financial statements and compare them with past data to predict the possibility of window dressing.

[0010] "Notification means" refers to communication equipment and systems for notifying the accounting auditor of specific suspicious points based on the prediction results in real time. [Brief explanation of the drawings]

[0011] [Figure 1] 1 is a conceptual diagram showing an example of the configuration of a data processing system according to a first embodiment. [Figure 2] 1 is a conceptual diagram showing an example of main functions of a data processing device and a smart device according to a first embodiment. [Figure 3] FIG. 10 is a conceptual diagram showing an example of the configuration of a data processing system according to a second embodiment. [Figure 4] FIG. 10 is a conceptual diagram showing an example of main functions of a data processing device and smart glasses according to a second embodiment. [Figure 5] FIG. 10 is a conceptual diagram showing an example of the configuration of a data processing system according to a third embodiment. [Figure 6] FIG. 11 is a conceptual diagram showing an example of main functions of a data processing device and a headset-type terminal according to a third embodiment. [Figure 7] FIG. 10 is a conceptual diagram showing an example of the configuration of a data processing system according to a fourth embodiment. [Figure 8] FIG. 10 is a conceptual diagram showing an example of main functions of a data processing device and a robot according to a fourth embodiment. [Figure 9] 1 shows an emotion map onto which multiple emotions are mapped. [Figure 10] 1 shows an emotion map onto which multiple emotions are mapped. [Figure 11] FIG. 3 is a sequence diagram showing a processing flow of the data processing system according to the first embodiment. [Figure 12] FIG. 10 is a sequence diagram showing the flow of processing in the data processing system in Application Example 1. [Figure 13] FIG. 10 is a sequence diagram showing the flow of processing in the data processing system according to the second embodiment when an emotion engine is combined. [Figure 14] FIG. 10 is a sequence diagram showing the flow of processing in the data processing system in Application Example 2 when an emotion engine is combined. DETAILED DESCRIPTION OF THE INVENTION

[0012] An example of an embodiment of a system according to the technology of the present disclosure will be described below with reference to the accompanying drawings.

[0013] First, the terms used in the following description will be explained.

[0014] In the following embodiments, a coded processor (hereinafter simply referred to as a "processor") may be a single arithmetic device or a combination of multiple arithmetic devices. Furthermore, a processor may be a single type of arithmetic device or a combination of multiple types of arithmetic devices. Examples of arithmetic devices include a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a GPGPU (General-Purpose computing on Graphics Processing Units), and an APU (Accelerated Processing Unit).

[0015] In the following embodiments, a coded RAM (Random Access Memory) is a memory in which information is temporarily stored and is used as a working memory by a processor.

[0016] In the following embodiments, the coded storage is one or more non-volatile storage devices that store various programs, various parameters, etc. Examples of non-volatile storage devices include flash memory (SSD (Solid State Drive)), magnetic disks (e.g., hard disks), and magnetic tapes.

[0017] In the following embodiments, a communication I / F (Interface) with a symbol is an interface including a communication processor, an antenna, etc. The communication I / F controls communication between multiple computers. Examples of communication standards applied to the communication I / F include wireless communication standards including 5G (5th Generation Mobile Communication System), Wi-Fi (registered trademark), Bluetooth (registered trademark), etc.

[0018] In the following embodiments, "A and / or B" is synonymous with "at least one of A and B." In other words, "A and / or B" means that it may be only A, only B, or a combination of A and B. Furthermore, in this specification, the same concept as "A and / or B" is also applied when three or more things are expressed connected by "and / or."

[0019] [First embodiment]

[0020] FIG. 1 shows an example of the configuration of a data processing system 10 according to the first embodiment.

[0021] 1, a data processing system 10 includes a data processing device 12 and a smart device 14. An example of the data processing device 12 is a server.

[0022] The data processing device 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" according to the technology of the present disclosure. The computer 22 includes a processor 28, a RAM 30, and a storage 32. The processor 28, the RAM 30, and the storage 32 are connected to a bus 34. The database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. Examples of the network 54 include a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0023] The smart device 14 includes a computer 36, a reception device 38, an output device 40, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, a RAM 48, and a storage 50. The processor 46, the RAM 48, and the storage 50 are connected to a bus 52. The reception device 38, the output device 40, and the camera 42 are also connected to the bus 52.

[0024] The reception device 38 includes a touch panel 38A, a microphone 38B, and the like, and receives user input. The touch panel 38A detects contact with an indicator (for example, a pen or a finger) to receive user input by the touch of the indicator. The microphone 38B detects the user's voice to receive user input by voice. The control unit 46A transmits data indicating the user input received by the touch panel 38A and the microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the data indicating the user input.

[0025] The output device 40 includes a display 40A and a speaker 40B, and presents data to the user 20 by outputting the data in a form of expression that the user 20 can perceive (for example, audio and / or text). The display 40A displays visible information such as text and images in accordance with instructions from the processor 46. The speaker 40B outputs audio in accordance with instructions from the processor 46. The camera 42 is a compact digital camera equipped with an optical system including a lens, aperture, and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor.

[0026] The communication I / F 44 is connected to a network 54. The communication I / Fs 44 and 26 control the exchange of various information between the processor 46 and the processor 28 via the network 54.

[0027] FIG. 2 shows an example of the main functions of the data processing device 12 and the smart device 14.

[0028] 2, in the data processing device 12, a specific process is performed by the processor 28. A specific processing program 56 is stored in the storage 32. The specific processing program 56 is an example of a "program" according to the technology of the present disclosure. The processor 28 reads the specific processing program 56 from the storage 32 and executes the read specific processing program 56 on the RAM 30. The specific process is realized by the processor 28 operating as a specific processing unit 290 in accordance with the specific processing program 56 executed on the RAM 30.

[0029] The storage 32 stores a data generation model 58 and an emotion identification model 59. The data generation model 58 and the emotion identification model 59 are used by the identification processing unit 290.

[0030] In the smart device 14, the processor 46 performs the reception output process. The storage 50 stores a reception output program 60. The reception output program 60 is used in conjunction with the specific processing program 56 by the data processing system 10. The processor 46 reads the reception output program 60 from the storage 50 and executes the read reception output program 60 on the RAM 48. The reception output process is realized by the processor 46 operating as the control unit 46A in accordance with the reception output program 60 executed on the RAM 48.

[0031] Next, a description will be given of the specific processing performed by the specific processing unit 290 of the data processing device 12. In the following description, the data processing device 12 will be referred to as a "server" and the smart device 14 will be referred to as a "terminal."

[0032] System Overview

[0033] The system of the present invention includes data collection, data preprocessing, machine learning model construction, prediction using generative AI, and notification methods. This system is designed to efficiently detect corporate window dressing and reduce the burden of accounting audits.

[0034] Data collection

[0035] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[0036] Data Preprocessing

[0037] The server preprocesses the collected data. Specifically, it removes duplicates and errors from the data and fills in missing data from other sources. It also standardizes the format of numeric data and dates and normalizes all data. This prepares a clean dataset suitable for analysis.

[0038] Building a learning model

[0039] The server uses the preprocessed data to build a machine learning model. First, it divides the data into a training set, a test set, and a validation set. Next, it uses the training set to train the model by applying techniques such as anomaly detection algorithms, random forests, and deep learning. Finally, it evaluates the model's performance using the test set, checking precision and recall.

[0040] Predictions using generative AI

[0041] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal then sends this data to the server in real time. The server then uses generative AI and a trained model to analyze the latest data and compare it with past data. This allows it to predict whether there is a possibility of window dressing.

[0042] Real-time notifications

[0043] Based on the prediction results, the server generates a notification containing specific suspicious points. This notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structures. The server sends this notification in real time to the terminal, which receives the notification and displays it to the user (accountant). This allows the user to immediately identify suspicious transactions or unusual financial data and begin detailed audit procedures.

[0044] Specific examples

[0045] Examples of data collection

[0046] The server collects Company A's financial statement data, audit reports, and related publicly available news articles for the past 10 years.

[0047] Specific examples of data preprocessing

[0048] The server removes duplicate entries from the collected data, uses statistical methods to fill in missing sales data from other relevant financial data, standardizes all numerical data to a single currency (e.g., USD), and converts dates to "YYYY-MM-DD."

[0049] Specific examples of building learning models

[0050] The server divides the preprocessed data into a training set (70%), a test set (20%), and a validation set (10%). An anomaly detection algorithm is used to extract features, such as sales spikes and cost structure fluctuations. The model is then trained using a random forest algorithm, and its accuracy is verified on the test set.

[0051] Examples of predictions using generative AI

[0052] The user (accountant) inputs the latest financial data for Company B, and the device sends the data to the server. The server analyzes the data using generative AI and detects that sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, predicting a high possibility of window dressing.

[0053] Real-time notification examples

[0054] Based on the forecast results, the server generates a notification stating specific suspicious points, such as "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." The server sends this result to the terminal, which receives it and displays it to the user (accountant). Based on this, the user can immediately start a detailed audit.

[0055] In this way, by implementing the system of the present invention, it is possible to efficiently detect signs of window dressing and respond quickly.

[0056] The processing flow will be explained below.

[0057] Step 1:

[0058] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from the internet and internal systems. This data comes from a variety of sources, including publicly available company financial statements, third-party auditor reports, and even public reports.

[0059] Step 2:

[0060] The server cleans the collected data, removing duplicates and errors, filling in missing data from other data, and standardizing the format of numeric data and dates.

[0061] Step 3:

[0062] The server uses the preprocessed data to build machine learning models by splitting the data into training, test, and validation sets and applying techniques such as anomaly detection algorithms, random forests, and deep learning to train the models.

[0063] Step 4:

[0064] The server evaluates the trained model on the test set to check precision and recall, and retrains the model as needed to optimize performance.

[0065] Step 5:

[0066] The user (accountant) inputs the latest financial statement data of the company being audited into the terminal.

[0067] Step 6:

[0068] The terminal transmits the input data to the server in real time.

[0069] Step 7:

[0070] The server uses generative AI and trained models to analyze the latest financial data it receives, comparing it with historical data to determine whether any unusual patterns exist.

[0071] Step 8:

[0072] As a result of the analysis, the server predicts the possibility of accounting fraud, including specific suspicious points such as sudden increases in sales or fluctuations in cost structure.

[0073] Step 9:

[0074] The server generates a notification containing specific suspicious points based on the prediction results.

[0075] Step 10:

[0076] The server sends this notification to the terminal in real time.

[0077] Step 11:

[0078] The terminal receives the notification and displays it to the user (accountant), who can then immediately initiate a detailed audit.

[0079] Example 1

[0080] Next, a description will be given of Example 1. In the following description, the data processing device 12 will be referred to as a "server" and the smart device 14 will be referred to as a "terminal."

[0081] The challenge is to efficiently and quickly detect accounting fraud in corporate financial data and reduce the burden of accounting audits. Conventional methods require manual analysis of large amounts of data, which is prone to errors and is time-consuming and costly. In addition, advanced algorithms to improve the accuracy of anomaly detection have not been fully utilized.

[0082] The specific processing by the specific processing unit 290 of the data processing device 12 in the first embodiment is realized by the following means.

[0083] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using a generative AI, and a real-time notification means. This enables comprehensive collection and preprocessing of corporate financial data, audit reports, news articles, data on past window dressing cases, etc., and construction of a machine learning model for highly accurate anomaly detection. Furthermore, the use of generative AI enables real-time analysis and prediction based on the latest data entered by the user, and immediate notification can be sent if an anomaly is detected, enabling a rapid response.

[0084] "Data collection means" refers to the ability to collect corporate financial data, audit reports, news articles, and past accounting fraud cases from online and offline sources.

[0085] "Data preprocessing means" refers to functions for cleansing collected data, filling in missing data, and normalizing data.

[0086] "Machine learning model construction means" refers to the function of using preprocessed data to divide the data into a training set, a test set, and a validation set, constructing a model using an anomaly detection algorithm, random forest, or deep learning, and evaluating the accuracy and recall.

[0087] "Prediction method using generative AI" refers to a function that receives the latest financial results data of a company entered by the user, analyzes the data using a learning model and generative AI, detects anomalies, and predicts the possibility of window dressing.

[0088] The "real-time notification means" refers to a function that generates a notification containing suspicious points based on the prediction results, and transmits and displays the notification to the user terminal.

[0089] The system of the present invention is composed of a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, and a real-time notification means. This system is designed to efficiently detect corporate window dressing and reduce the burden of accounting audits.

[0090] Data collection

[0091] The server collects company financial data, audit reports, news articles, and data on past accounting fraud cases via the internet and internal systems. To collect data, it uses Python's BeautifulSoup library and Selenium to perform web scraping. It also uses an ERP system to extract data from the company's internal systems.

[0092] Example: A server sends an HTTP request to retrieve Company A's financial statements from a financial information website, extracts the necessary data from the HTML using BeautifulSoup, and retrieves financial data from the company's ERP system via an API.

[0093] Data Preprocessing

[0094] The server uses data processing libraries such as Pandas and NumPy to preprocess the collected data, specifically deleting duplicate data, correcting erroneous data, filling in missing data, standardizing the currency units of numeric data, and standardizing date formats.

[0095] Example: The server removes duplicate entries from the collected dataset using Pandas' drop_duplicates() method, fills missing data with the average value using the fillna() method, converts all currency units to USD, and standardizes the date format to "YYYY-MM-DD".

[0096] Building a learning model

[0097] The server uses the preprocessed data to build a machine learning model. It divides the data into a training set, test set, and validation set, and trains the model using an anomaly detection algorithm (e.g., Isolation Forest), random forest, or a deep learning framework (e.g., TENSORFLOW (registered trademark)). It evaluates the precision and recall of the models and selects the optimal model.

[0098] Example: The server splits the preprocessed data using scikit-learn's train_test_split function, trains the model using the Isolation Forest algorithm, runs the training using the fit() method, and evaluates the test data using the predict() method.

[0099] Predictions using generative AI

[0100] The user (accountant) enters the company's latest financial data into the terminal, which then sends the data to the server. The server uses a generative AI model (e.g., GPT-4 (registered trademark)) and a trained model to analyze the data, detect anomalies, and predict the possibility of window dressing.

[0101] Example: A user enters financial data for Company B into a form on a terminal and presses the "Submit" button. The terminal sends the data in JSON format to the server as a POST request. The server inputs the received data into the GPT-4 generation AI, which generates a prompt such as "Please analyze the sudden increase in sales and fluctuations in operating profit margins for Company B."

[0102] Real-time notifications

[0103] The server generates a notification containing suspicious points based on the prediction results and sends it to the user's terminal. The user can receive this notification and perform a detailed audit based on the displayed abnormal points.

[0104] Example: Based on the prediction results, the server generates a notification such as, "Company B's second-quarter sales have increased by 50% compared to the same month last year, but there is no significant change in its operating profit margin, so there is a high possibility that this is an abnormal transaction," and sends it to the terminal. The terminal receives this notification and displays it to the user as a pop-up or email notification.

[0105] In this way, this system can efficiently detect corporate window dressing and enable rapid response.

[0106] The flow of the identification process in the first embodiment will be described with reference to FIG.

[0107] Step 1:

[0108] The server uses data collection tools to collect data from the internet and internal company systems. Specifically, it uses web scraping tools such as BeautifulSoup and Selenium to obtain company financial data, audit reports, news articles, and data on past accounting fraud cases. It also collects internal data from the company's ERP system via APIs.

[0109] Input: A list of URLs or API endpoints for data collection.

[0110] Output: HTML page and / or JSON formatted data.

[0111] What it does: The server crawls financial information website URLs and sends HTTP requests. It then parses the HTML content with BeautifulSoup to extract financial data. It also uses the API of the company's ERP system to obtain an authentication token and collect the necessary financial data in JSON format.

[0112] Step 2:

[0113] The server preprocesses the collected data using data preprocessing tools, such as data cleansing, removing duplicates, filling in missing data, standardizing numeric data, and standardizing date formats. This is done using data processing libraries such as Pandas and NumPy.

[0114] Input: Raw data such as collected financial data and audit reports.

[0115] Output: A preprocessed clean dataset.

[0116] Specific operation: The server removes duplicate data using Pandas' drop_duplicates() method, fills missing values ​​with the mean using the fillna() method, converts numeric data to a unified currency (e.g., USD), and standardizes the date format to "YYYY-MM-DD".

[0117] Step 3:

[0118] The server uses a machine learning model building tool to build a machine learning model using the preprocessed data. It splits the data into a training set, a test set, and a validation set, and trains the model using an anomaly detection algorithm, random forest, or a deep learning framework (e.g., TensorFlow).

[0119] Input: Preprocessed clean dataset.

[0120] Output: A trained machine learning model.

[0121] Specific operation: The server uses scikit-learn's train_test_split function to split the data into a 70% training set, a 20% test set, and a 10% validation set. It trains the model using the Isolation Forest algorithm or Random Forest, and builds the model using the fit() method. It also evaluates the test data using the predict() method to check precision and recall.

[0122] Step 4:

[0123] A user (accountant) inputs the company's latest financial data into a terminal, which then sends the data to a server. The server uses a generative AI model (e.g., GPT-4) and a trained model to analyze the data, detect anomalies, and predict possible window dressing.

[0124] Input: The most recent financial data entered by the user.

[0125] Output: Anomaly detection results and predicted window dressing probability.

[0126] Specific operation: The user enters the company's financial data into a form on the device and presses the "Submit" button. The device sends this data in JSON format to the server as a POST request. The server inputs the received data into the GPT-4 generation AI, which generates and analyzes the prompt: "Please analyze the sudden increase in sales and fluctuations in operating profit margins for Company B."

[0127] Step 5:

[0128] The server generates a notification containing suspicious points based on the prediction results and sends it to the user terminal in real time. The user can receive this notification and perform a detailed audit based on the displayed abnormalities.

[0129] Input: Anomaly detection results and predicted probability of window dressing.

[0130] Output: Informational message.

[0131] Specific operation: Based on the prediction results, the server generates a notification such as, "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions," and sends it to the terminal in JSON format. The terminal receives this notification and displays it to the user as a pop-up or email notification.

[0132] (Application example 1)

[0133] Next, a description will be given of Application Example 1. In the following description, the data processing device 12 will be referred to as a "server" and the smart device 14 will be referred to as a "terminal."

[0134] Currently, detecting corporate accounting fraud requires a great deal of effort and time, making it difficult to detect in real time. Furthermore, it is difficult to ensure the quality and consistency of collected data, which increases the risk of false positives and oversights. For this reason, there is a need to quickly and accurately detect corporate financial fraud and improve the efficiency of accounting audits.

[0135] The specific processing by the specific processing unit 290 of the data processing device 12 in the application example 1 is realized by the following means.

[0136] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using the generative AI, a notification means for sending a notification to a specific device, a means for collecting and analyzing data from multiple information sources of the company in real time, a means for inputting new standardized data into the generative AI model, a means for immediately notifying a user based on the anomaly detection result, and a means for detecting financial fraud at the company using the prediction result of the generative AI. This makes it possible to accurately detect financial fraud at the company in real time and significantly improve the efficiency of accounting audits.

[0137] "Data collection means" refers to means for collecting corporate financial data, audit reports, news articles, and past accounting fraud cases from online and offline sources.

[0138] "Data preprocessing means" refers to means for cleaning collected data, imputing missing data, normalizing data, and standardizing data for input into generative AI models.

[0139] A "machine learning model construction means" is a means for constructing a machine learning model using preprocessed data and using that model to perform anomaly detection and prediction.

[0140] A "predictive method using generative AI" is a method of predicting corporate financial fraud by inputting new standardized data into a generative AI model.

[0141] "Notification means" refers to a means for immediately notifying users of normal or emergency conditions based on the results of anomaly detection.

[0142] "Means for collecting and analyzing data from multiple sources of information for a company in real time" refers to means for continuously collecting and analyzing a company's financial data, audit reports, news articles, and data on past accounting fraud cases in real time.

[0143] "Notification means for sending a notification to a specific device" refers to a means for sending a notification to a specific device such as a smartphone or desktop when financial fraud at a company is detected.

[0144] "Means for inputting new standardized data into the generative AI model" refers to means for inputting newly collected data into the generative AI model and analyzing it in real time.

[0145] "Means for immediately notifying the user based on the anomaly detection results" refers to a means for immediately sending a notification to the user based on an anomaly detected by the generative AI model.

[0146] The system of the present invention aims to efficiently detect corporate window dressing and reduce the burden of accounting audits. This system includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, a notification means for sending notifications to specific devices, a means for collecting and analyzing data from multiple corporate information sources in real time, a means for inputting new standardized data into the generative AI model, a means for immediately notifying users based on the results of anomaly detection, and a means for detecting corporate financial fraud using the prediction results of the generative AI.

[0147] The server collects company financial data, audit reports, news articles, and data on past accounting fraud cases via the Internet and internal systems. This allows data to be collected from companies' published financial statements, reports from third-party auditors, and even public reports. A specific example of data collection would be obtaining Company A's financial data for the past 10 years via an API.

[0148] The server also preprocesses the collected data, removing duplicates and errors, filling in missing data from other data, and standardizing the data format. This results in a clean dataset with consistent numerical data and dates. For example, missing sales data can be filled in using statistical methods from other relevant data.

[0149] The server then uses the preprocessed data to build a machine learning model. The data is divided into a training set, a test set, and a validation set, and the model is trained using techniques such as anomaly detection algorithms, random forests, and deep learning. For example, a random forest model is trained using the training set, and the performance of the constructed model is verified using the test set.

[0150] As a prediction method using generative AI, the user inputs the latest data of the audited company into a terminal and sends the data from the terminal to the server. The server analyzes the latest data using the generative AI model and the trained model and compares it with past data. At this point, if sales have increased significantly compared to the same period last year but the operating profit margin has not changed, it predicts that there is a high possibility of window dressing.

[0151] Finally, the server uses the notification method to notify the user in real time based on the prediction results. The notification details any abnormal patterns and specific suspicious points. For example, the notification might say, "Second-quarter sales increased by 50% compared to the same month last year, but there was no significant change in the operating profit margin, so there is a high possibility of abnormal transactions."

[0152] The hardware used includes a server for collecting and processing data, and a terminal where users can input data and receive notifications. The software includes an API used for data collection and preprocessing, generative AI models and machine learning algorithms used for data analysis, and a communication protocol for notifications. Specifically, Python's pandas and scikit-learn, and deep learning libraries (such as TensorFlow and PyTorch) are used.

[0153] As an example, here is a prompt that uses a generative AI model to analyze new financial data:

[0154] Prompt statement

[0155] Analyze new financial data. This data includes information such as sales, operating profit, net profit, and debt ratio. Also, if sales increase by 50% year-on-year, but the operating profit margin remains unchanged, this may indicate window dressing. Detect and report such abnormal patterns.

[0156] As described above, by implementing the system of the present invention, it is possible to accurately detect corporate window dressing in real time, thereby significantly improving the efficiency of accounting audits.

[0157] The flow of the specific processing in the application example 1 will be described with reference to FIG.

[0158] Step 1:

[0159] The server collects corporate financial data, audit reports, news articles, and data on past accounting fraud cases via the internet and internal systems. Specifically, it obtains this information using APIs. For example, it uses financial data APIs and audit report APIs. The API endpoint and authentication information are required as input, and the collected raw data is obtained as output.

[0160] Step 2:

[0161] The server preprocesses the collected data, removing duplicates and erroneous data and imputing missing data from other data. It also standardizes and normalizes numerical data units and makes date formats consistent. Specifically, it uses the Python pandas library for data cleaning and imputation. The input is the collected raw data, and the output is a preprocessed, clean dataset.

[0162] Step 3:

[0163] The server uses the preprocessed data to build a machine learning model. Specifically, it divides the data into a training set, a test set, and a validation set, and trains it using anomaly detection algorithms, random forests, and deep learning techniques. Libraries used include scikit-learn and TensorFlow. The input is the preprocessed dataset, and the output is a trained machine learning model.

[0164] Step 4:

[0165] The user inputs the latest financial statement data of the company being audited into the terminal. The terminal sends this data to the server in real time. For example, the user inputs financial statements and related data into an input form. The input is the user's latest financial statement data, and the output is the data sent from the terminal to the server.

[0166] Step 5:

[0167] The server analyzes the latest input data using the generative AI model and the trained model. This allows it to compare the latest data with past data and detect abnormal patterns. The prompt text used is something like, "Please analyze the new financial data. This data includes information such as sales, operating profit, net profit, and debt ratio. Furthermore, if sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, this could be window dressing. Please detect and report such abnormal patterns." The input is the latest financial data, and the output is the results of anomaly detection.

[0168] Step 6:

[0169] The server generates a notification containing specific suspicious points based on the prediction results. For example, it generates a notification containing specific content such as, "Sales for the second quarter have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." The input is the anomaly detection result, and the output is the generated notification message.

[0170] Step 7:

[0171] The device displays notifications received from the server to the user in real time. Notifications are displayed as push notifications on smartphones or desktop notifications. Specifically, notifications are received and displayed on the device using push notification APIs or desktop notification APIs. The input is the notification message sent from the server, and the output is the notification that the user can receive and check.

[0172] Furthermore, an emotion engine that estimates the user's emotion may be combined. That is, the identification processing unit 290 may estimate the user's emotion using the emotion identification model 59 and perform identification processing using the user's emotion.

[0173] System Overview

[0174] The system of the present invention comprises a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, a notification means, and an emotion engine that recognizes user emotions. This system is designed not only to efficiently detect corporate window dressing and reduce the burden of accounting audits, but also to respond according to the user's emotional state.

[0175] Data collection

[0176] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[0177] Data Preprocessing

[0178] The server cleans the collected data. Specifically, it removes duplicates and errors from the data and fills in missing data from other sources. It also standardizes the format of numeric data and dates and normalizes all data. This prepares a clean dataset suitable for analysis.

[0179] Building a learning model

[0180] The server uses the preprocessed data to build a machine learning model. First, it divides the data into a training set, a test set, and a validation set. Next, it uses the training set to train the model by applying techniques such as anomaly detection algorithms, random forests, and deep learning. Finally, it evaluates the model's performance using the test set, checking precision and recall.

[0181] Predictions using generative AI

[0182] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal then sends this data to the server in real time. The server then uses generative AI and a trained model to analyze the latest data and compare it with past data. This allows it to predict whether there is a possibility of window dressing.

[0183] Emotion Engine Operation

[0184] The device's built-in emotion engine analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. For example, it analyzes the user's stress level and sense of security from changes in voice tone and facial expressions. This information is reflected when the prediction results are notified.

[0185] Real-time notifications

[0186] The server generates a notification containing specific suspicious points based on the prediction results. The notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structure. At the same time, the emotion engine determines the user's emotional state and adjusts the tone of the notification and the way detailed information is presented. For example, if the user is feeling stressed, the notification content will be concise and clear.

[0187] Logging

[0188] The server logs the user's emotional state as determined by the emotion engine and stores it in a database for later analysis and improvement. This data helps understand user emotional trends over time and further optimize the system.

[0189] Specific examples

[0190] Examples of data collection

[0191] The server collects Company A's financial statement data, audit reports, and related publicly available news articles for the past 10 years.

[0192] Specific examples of data preprocessing

[0193] The server removes duplicate entries from the collected data, uses statistical methods to fill in missing sales data from other relevant financial data, standardizes all numerical data to a single currency (e.g., USD), and converts dates to "YYYY-MM-DD."

[0194] Specific examples of building learning models

[0195] The server divides the preprocessed data into a training set (70%), a test set (20%), and a validation set (10%). An anomaly detection algorithm is used to extract features, such as sales spikes and cost structure fluctuations. The model is then trained using a random forest algorithm, and its accuracy is verified on the test set.

[0196] Examples of predictions using generative AI

[0197] The user (accountant) inputs the latest financial data for Company B, and the device sends the data to the server. The server analyzes the data using generative AI and detects that sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, predicting a high possibility of window dressing.

[0198] Examples of emotion engines

[0199] The device's emotion engine analyzes the user's (the accountant's) voice and determines that their stress level is high. Based on this information, the server generates short, concise notifications to reduce the user's stress.

[0200] Real-time notification examples

[0201] Based on the forecast results, the server generates a notification with specific suspicious points, such as "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." Because the emotion engine determined the stress level to be high, only the main points are included in the detailed information. The server sends this result to the terminal, which receives it and displays it to the user (accountant). Based on this, the user can immediately start a detailed audit.

[0202] In this way, by implementing the system of the present invention, it is possible to efficiently detect signs of window dressing and take the most appropriate action depending on the user's emotional state.

[0203] The processing flow will be explained below.

[0204] Step 1:

[0205] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[0206] Step 2:

[0207] The server cleans the collected data. Specifically, it removes duplicates and errors from the data. It also supplements missing data from other data and standardizes the format of numerical data and dates. For example, it standardizes data in different currency units and converts them to the "YYYY-MM-DD" format.

[0208] Step 3:

[0209] The server uses the preprocessed data to build a machine learning model by first splitting the data into a training set, a test set, and a validation set, and then using the training set to apply techniques such as anomaly detection algorithms, random forests, or deep learning to train the model.

[0210] Step 4:

[0211] The server evaluates the trained model on the test set to check precision and recall, and retrains the model as needed to optimize performance.

[0212] Step 5:

[0213] The user (accountant) inputs the latest financial statement data of the company being audited into the terminal, such as data on sales, operating profit, and costs.

[0214] Step 6:

[0215] The terminal transmits the input data to the server in real time.

[0216] Step 7:

[0217] The server uses generative AI and trained models to analyze the latest financial data it receives, comparing the input data with past data to detect any abnormal patterns, such as sudden increases in sales or cost inconsistencies.

[0218] Step 8:

[0219] Based on the analysis results, the server predicts the possibility of window dressing, including specific suspicious points such as sudden increases in sales or fluctuations in cost structure.

[0220] Step 9:

[0221] The device's built-in emotion engine analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. For example, it can determine their stress level based on changes in their voice tone and facial expressions.

[0222] Step 10:

[0223] Based on the emotion engine's findings, the server adjusts the content of the notification. For example, if the user is feeling stressed, the notification will be brief and clear.

[0224] Step 11:

[0225] Based on the predictions, the server generates specific suspicious alerts, including details such as unusual sales spikes or fluctuations in cost structures.

[0226] Step 12:

[0227] The server sends this notification to the terminal in real time.

[0228] Step 13:

[0229] The terminal receives the notification and displays it to the user (accountant), who can then immediately begin a detailed audit, for example by gathering additional information on specific items or preparing questions for the company.

[0230] Step 14:

[0231] The server logs the user's emotional state as determined by the emotion engine and stores it in a database for later analysis and improvement, for example, to understand the user's emotional trends over a long period of time and to help further optimize the system.

[0232] Example 2

[0233] Next, a description will be given of Example 2. In the following description, the data processing device 12 will be referred to as a "server" and the smart device 14 will be referred to as a "terminal."

[0234] In audit work, it is important to quickly and accurately detect corporate accounting fraud, but conventional methods require a great deal of time and effort from data collection to analysis. Furthermore, notification methods that do not take into account the stress and emotional state of accountants may reduce the efficiency of audit work. To solve these problems, efficient data processing and prediction methods, as well as appropriate responses according to the user's emotional state, are required.

[0235] The specific processing by the specific processing unit 290 of the data processing device 12 in the second embodiment is realized by the following means.

[0236] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, a notification means, and a log recording means, which enable fast and accurate detection of corporate window dressing and provide an optimal notification method according to the user's emotional state.

[0237] "Data collection methods" are methods for collecting corporate financial information, audit reports, news articles, and information on past accounting fraud cases from online and offline sources.

[0238] "Data preprocessing means" refers to means for cleaning collected information, filling in missing data, and normalizing information.

[0239] A "machine learning model building means" is a means for training and building a machine learning model using preprocessed data.

[0240] "Predictive methods using generative AI" are methods that analyze a company's financial information based on machine learning models and predict the possibility of accounting fraud.

[0241] The "emotion recognition means" is a means for analyzing the user's voice or facial expression and determining the user's emotional state.

[0242] The "notification means" is a means for adjusting and providing notification content according to the prediction result and the emotional state determined by the emotion recognition means.

[0243] The "log recording means" is a means for recording the emotional state of the user determined by the emotion recognition means and storing it in a database.

[0244] The system of the present invention includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, a notification means, and a log recording means. This system is designed to efficiently detect corporate window dressing, reduce the burden of accounting audits, and respond according to the user's emotional state.

[0245] First, the server uses data collection tools to collect company financial information, audit reports, news articles, and information on past accounting fraud cases from the Internet and internal systems. Specifically, it efficiently collects the required information using Python libraries (such as BeautifulSoup and Selenium) and APIs (such as FinAPI). For example, it can collect Company A's financial statement data for the past 10 years, audit reports, and publicly available related news articles.

[0246] Next, the server uses data preprocessing tools to clean the collected data, impute missing data, and normalize the data. Specifically, it removes duplicate data using the drop_duplicates() function in the pandas library, detects missing values ​​using the isnull() function, and imputes them using statistical methods. It also converts numeric data to a unified currency unit (e.g., USD) and standardizes date formats to "YYYY-MM-DD" using the datetime module.

[0247] The server then uses the preprocessed data to build a machine learning model using a machine learning model building method. The data set is split into a training set, a test set, and a validation set, and the train_test_split function in the sklearn library is used. The server also uses Isolation Forest or Random Forest Classifier for feature extraction to train and build the model. For example, an anomaly detection algorithm is used to extract features using indicators such as sudden sales increases or fluctuations in cost structure.

[0248] The user (accountant) inputs the latest financial statement data of the company being audited into the terminal. The terminal sends the data to the server in real time. The server analyzes the input data using a predictive method that uses generative AI, and compares it with past data to predict the possibility of window dressing. The generative AI uses a model such as OpenAI (registered trademark)'s GPT-4. For example, a prompt statement could be entered as follows: "Enter the latest financial statement data of Company B, have the server analyze it, and detect that sales have increased by 50% compared to the same period last year, but the operating profit margin has remained unchanged, and predict that there is a high possibility of window dressing."

[0249] Furthermore, the device's built-in emotion recognition means analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. It uses the librosa library for voice analysis and the OpenCV and dlib libraries for facial expression recognition. For example, it detects changes in voice tone or abnormal facial expressions to determine whether the user is in a high-stress state. This information is reflected in the notification content sent by the prediction result notification means.

[0250] The server uses a real-time notification mechanism to generate a notification containing specific suspicious points based on the prediction results. This notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structure. The notification content is adjusted to be concise and clear, taking into account the user's emotional information determined by the emotion recognition mechanism. For example, a notification may be generated that states, "Sales in the second quarter have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." This notification is sent to the terminal and displayed to the user (accountant).

[0251] Finally, the server uses the logging means to log the user's emotional state as determined by the emotion recognition means and stores the log in a database. This data is used to understand the user's emotional trends over a long period of time and to further optimize the system. For example, if a user is in a state of high stress under certain conditions, this data can be used as feedback to revise the system's interface and notification methods.

[0252] The above configuration and processing make it possible to quickly and accurately detect corporate window dressing and respond optimally according to the user's emotional state.

[0253] The flow of the identification process in the second embodiment will be described with reference to FIG.

[0254] Processing flow

[0255] Step 1: Collect data

[0256] The server collects corporate financial information, audit reports, news articles, and information on past accounting fraud cases via the internet and internal systems. It efficiently collects the necessary data using libraries such as Python's BeautifulSoup and Selenium, and APIs such as FinAPI. Specifically, it retrieves data from the API using the requests library, and parses HTML with BeautifulSoup to extract financial information.

[0257] Input: URL or API endpoint to collect data

[0258] Output: Dataset of company financial information, audit reports, press articles, and information on past accounting fraud cases

[0259] Step 2: Preprocessing the data

[0260] The server cleans the collected data, imputes missing data, and normalizes the data. Specifically, it uses the pandas library to remove duplicate data, the isnull() function to detect missing values, and statistical methods to impute them. It also standardizes date formats using the datetime module and converts numeric data to a uniform currency unit.

[0261] Input: Dataset of company financial information, audit reports, news articles, and information on past accounting fraud cases

[0262] Output: A clean and normalized dataset

[0263] Step 3: Building a learning model

[0264] The server builds a machine learning model using the preprocessed data. It splits the dataset into training, test, and validation sets using the train_test_split function in the sklearn library, and uses IsolationForest and RandomForestClassifier for feature extraction. After training the model, it evaluates its accuracy on the test set.

[0265] Input: Clean and normalized dataset

[0266] Output: A trained machine learning model

[0267] Step 4: Predict using generative AI

[0268] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal sends the data in real time to the server. The server uses generative AI and a trained model to analyze the input data and predict the possibility of window dressing. The generative AI uses a model such as GPT-4.

[0269] Input: Latest financial data

[0270] Output: Prediction result regarding the possibility of window dressing

[0271] Step 5: Emotion Engine in Action

[0272] The emotion engine installed on the device analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. Changes in voice tone are analyzed using librosa, and facial expressions are evaluated using OpenCV and dlib. This determines stress levels and emotional states.

[0273] Input: User's voice data, facial expression data

[0274] Output: User's emotional state

[0275] Step 6: Real-time notifications

[0276] The server generates a notification based on the prediction results, including specific suspicious points, such as sudden sales increases or fluctuations in cost structure, and adjusts the content and tone of the notification based on the emotional information determined by the emotion recognition means.

[0277] Input: Prediction result, user's emotional state

[0278] Output: Adjusted notification content

[0279] Step 7: Logging

[0280] The server records the user's emotional state as determined by the emotion engine as a log and stores it in a database. The log data is used to improve the system and optimize user responses.

[0281] Input: User's emotional state

[0282] Output: Saved log data

[0283] (Application example 2)

[0284] Next, a description will be given of Application Example 2. In the following description, the data processing device 12 will be referred to as a "server" and the smart device 14 will be referred to as a "terminal."

[0285] In modern business activities, window dressing remains a serious problem, placing a heavy burden on accountants and auditors. Audit work also creates significant stress and strain. To address this issue, a system is needed that efficiently detects abnormal transactions and provides appropriate notifications based on the accountant's emotional state. The present invention aims to improve corporate reliability and reduce the burden of accounting audits by solving these problems.

[0286] The specific processing by the specific processing unit 290 of the data processing device 12 in the application example 2 is realized by the following means.

[0287] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, and a real-time notification means, which enable efficient detection of anomalies in a company's financial data and appropriate notification according to the user's emotional state.

[0288] "Data collection methods" are methods for collecting financial data, audit reports, information articles, and historical anomaly data from online or offline sources.

[0289] "Data preprocessing means" refers to methods for cleaning collected data, filling in missing data, and normalizing the data into a format that is easy to analyze.

[0290] "Machine learning model building means" refers to a method of using pre-processed data to apply anomaly detection algorithms and other machine learning techniques to build a model for detecting anomalous transactions.

[0291] "Prediction methods using generative AI" are methods that use trained machine learning models and generative AI to analyze newly obtained data and predict the possibility of abnormal transactions.

[0292] "Emotion recognition means" is a method for determining the user's emotional state by analyzing the user's voice and facial expressions, and adjusting the system's response based on that information.

[0293] "Real-time notification means" is a method of generating a notification based on the results of abnormal transaction detection, with content that corresponds to the user's emotional state, and quickly notifying the user.

[0294] The present invention is a system for monitoring and detecting anomalies in corporate financial data, and is intended to be used on end-user devices such as smartphone applications. Specific procedures for implementing the invention are described below.

[0295] Hardware and software used

[0296] Hardware: Smartphones, cloud servers

[0297] Software: Python, Scikit-learn, Keras, EmotionEngine, NotificationService

[0298] System configuration

[0299] 1. Data collection methods:

[0300] The server collects financial data, audit reports, news articles, and past anomaly data from online and offline sources. For example, financial data is collected through APIs, and audit reports and news articles are collected using a web crawling system.

[0301] 2. Data preprocessing methods:

[0302] Once collected, the data is filled in with missing values, errors corrected, and formatted in a consistent manner (for example, missing sales data is filled in from historical data, and date formats are standardized).

[0303] 3. Machine learning model building methods:

[0304] The preprocessed data is used to apply an anomaly detection algorithm and build a model. Specifically, the data is divided into a training set and a test set, and training is performed using a random forest or deep learning model.

[0305] 4. Generative AI prediction methods:

[0306] When a user inputs their latest financial data, the data is sent to the server, where generative AI and trained models are used to predict anomalies. For example, if sales have increased significantly but the operating profit margin remains unchanged, the system predicts possible anomalies.

[0307] 5. Emotion recognition means:

[0308] Analyzes the user's voice and facial expressions to determine their emotional state. Emotion Engine is used to analyze voice tone and facial expression changes to determine the user's stress level.

[0309] 6. Real-time notification methods:

[0310] Based on the anomaly detection results, notifications are generated with content that corresponds to the user's emotional state. For example, if the user is particularly stressed, the notification content will be brief and only the main points will be presented.

[0311] Specific examples of realization

[0312] 1. Examples of data collection:

[0313] Use APIs to retrieve company financial data and collect audit reports for the past 10 years.

[0314] 2. Examples of data preprocessing:

[0315] Remove duplicate entries from collected data, impute missing data using statistical methods, convert all numerical data to a single currency, and standardize date formats.

[0316] 3. Specific examples of building learning models:

[0317] The preprocessed data is used to train a model using the random forest algorithm, and the performance of the anomaly detection model is verified.

[0318] 4. Specific examples of predictions using generative AI:

[0319] Users enter their company's latest financial data, which is then analyzed on the server, and if there is a high possibility of an abnormality, an alert is issued immediately.

[0320] 5. Examples of emotion recognition:

[0321] The system analyzes the user's voice and, if it determines that the stress level is high, simplifies the notification content.

[0322] 6. Examples of real-time notifications:

[0323] "Second-quarter sales have surged 50% compared to the same month last year, but there has been no significant change in operating profit margins, so there is a high possibility that this is an abnormal transaction," the notice said.

[0324] Prompt Sentence Examples

[0325] "Evaluate a company's latest financial data and detect anomalies. Write a program that adjusts the tone of notifications based on the user's stress level."

[0326] The flow of the specific processing in the application example 2 will be described with reference to FIG.

[0327] Step 1:

[0328] Data collection by data collection method

[0329] The server collects financial data, audit reports, information articles, and historical anomaly data from online and offline sources. Specifically, it retrieves financial data using APIs and collects news articles using a web crawling system. It uses the URL or API endpoint of each source as input and obtains the collected raw data as output.

[0330] Step 2:

[0331] Data cleaning through data preprocessing methods

[0332] The server cleans the collected raw data, imputes missing data, and normalizes the data into a format that is easy to analyze. Specifically, it removes duplicate data, imputes missing data (e.g., imputation using statistical methods), and standardizes date formats. It uses the collected raw data as input and obtains a cleaned dataset as output.

[0333] Step 3:

[0334] Building models using machine learning model building methods

[0335] The server uses the preprocessed data to apply an anomaly detection algorithm and build a model. Specifically, it splits the data into a training set and a test set, and trains the model using a random forest or deep learning algorithm. The preprocessed dataset is used as input, and the trained model is obtained as output.

[0336] Step 4:

[0337] Predicting anomalies using generative AI-based predictive methods

[0338] When a user inputs their latest financial data, the device sends the data to the server. The server then uses generative AI and trained models to analyze this data and predict anomalies. Specifically, it detects sudden increases in sales or inconsistencies in cost structures. The latest financial data provided by the user is used as input, and a prediction result indicating whether or not anomalies exist is obtained as output.

[0339] Step 5:

[0340] Emotional state analysis using emotion recognition tools

[0341] The device analyzes the user's voice and facial expressions in real time to determine their emotional state. Specifically, it uses the Emotion Engine to analyze changes in voice tone and facial expressions to determine stress levels and feelings of relief. The device uses the user's voice and facial expression data as input and outputs the determined emotional state.

[0342] Step 6:

[0343] Generate and send notifications via real-time notification methods

[0344] The server generates a notification based on the anomaly detection results, with content appropriate to the user's emotional state. Specifically, if an anomaly is detected, a detailed explanation is provided, and if the user is in a stressful state, a brief notification is provided. The anomaly detection results and the user's emotional state are used as inputs, and the generated notification message is obtained as output.

[0345] The specific processing unit 290 transmits the result of the specific processing to the smart device 14. In the smart device 14, the control unit 46A causes the output device 40 to output the result of the specific processing. The microphone 38B acquires audio indicating a user input regarding the result of the specific processing. The control unit 46A transmits audio data indicating the user input acquired by the microphone 38B to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the audio data.

[0346] The data generation model 58 is a so-called generative AI (Artificial Intelligence). An example of the data generation model 58 is ChatGPT (registered trademark) (Internet search engine).<URL: https: / / openai.com / blog / chatgpt> ), Gemini (registered trademark) (Internet search <url: https: gemini.google.com ?hl="ja">) and other generation AIs. The data generation model 58 is obtained by performing deep learning on a neural network. A prompt including an instruction is input to the data generation model 58, and inference data such as voice data indicating voice, text data indicating text, and image data indicating an image is also input. The data generation model 58 performs inference on the input inference data in accordance with the instruction indicated by the prompt, and outputs the inference result in a data format such as voice data and text data. Here, inference refers to, for example, analysis, classification, prediction, and / or summarization.

[0347] In the above embodiment, an example in which the specific process is performed by the data processing device 12 has been given, but the technology of the present disclosure is not limited to this, and the specific process may be performed by the smart device 14.

[0348] [Second embodiment]

[0349] FIG. 3 shows an example of the configuration of a data processing system 210 according to the second embodiment.

[0350] 3, the data processing system 210 includes the data processing device 12 and smart glasses 214. An example of the data processing device 12 is a server.

[0351] The data processing device 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" according to the technology of the present disclosure. The computer 22 includes a processor 28, a RAM 30, and a storage 32. The processor 28, the RAM 30, and the storage 32 are connected to a bus 34. The database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. Examples of the network 54 include a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0352] The smart glasses 214 include a computer 36, a microphone 238, a speaker 240, a camera 42, and a communication I / F 44. The computer 36 includes a processor 46, a RAM 48, and a storage 50. The processor 46, the RAM 48, and the storage 50 are connected to a bus 52. The microphone 238, the speaker 240, and the camera 42 are also connected to the bus 52.

[0353] The microphone 238 receives instructions and the like from the user 20 by receiving voice uttered by the user 20. The microphone 238 captures the voice uttered by the user 20, converts the captured voice into audio data, and outputs it to the processor 46. The speaker 240 outputs audio in accordance with instructions from the processor 46.

[0354] Camera 42 is a small digital camera equipped with an optical system including a lens, aperture, and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, and captures images of the surroundings of user 20 (for example, an imaging range defined by an angle of view equivalent to the field of vision of a typical healthy person).

[0355] The communication I / F 44 is connected to a network 54. The communication I / Fs 44 and 26 are responsible for the exchange of various information between the processor 46 and the processor 28 via the network 54. The exchange of various information between the processor 46 and the processor 28 using the communication I / Fs 44 and 26 is carried out in a secure state.

[0356] Fig. 4 shows an example of the main functions of the data processing device 12 and the smart glasses 214. As shown in Fig. 4, in the data processing device 12, a specific process is performed by the processor 28. A specific process program 56 is stored in the storage 32.

[0357] The specific processing program 56 is an example of a "program" according to the technology of the present disclosure. The processor 28 reads the specific processing program 56 from the storage 32 and executes the read specific processing program 56 on the RAM 30. The specific processing is realized by the processor 28 operating as a specific processing unit 290 in accordance with the specific processing program 56 executed on the RAM 30.

[0358] The storage 32 stores a data generation model 58 and an emotion identification model 59. The data generation model 58 and the emotion identification model 59 are used by the identification processing unit 290.

[0359] In the smart glasses 214, the reception output process is performed by the processor 46. A reception output program 60 is stored in the storage 50. The processor 46 reads the reception output program 60 from the storage 50 and executes the read reception output program 60 on the RAM 48. The reception output process is realized by the processor 46 operating as the control unit 46A in accordance with the reception output program 60 executed on the RAM 48.

[0360] Next, a description will be given of the identification process performed by the identification processing unit 290 of the data processing device 12. In the following description, the data processing device 12 will be referred to as the "server" and the smart glasses 214 will be referred to as the "terminal."

[0361] System Overview

[0362] The system of the present invention includes data collection, data preprocessing, machine learning model construction, prediction using generative AI, and notification methods. This system is designed to efficiently detect corporate window dressing and reduce the burden of accounting audits.

[0363] Data collection

[0364] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[0365] Data Preprocessing

[0366] The server preprocesses the collected data. Specifically, it removes duplicates and errors from the data and fills in missing data from other sources. It also standardizes the format of numeric data and dates and normalizes all data. This prepares a clean dataset suitable for analysis.

[0367] Building a learning model

[0368] The server uses the preprocessed data to build a machine learning model. First, it divides the data into a training set, a test set, and a validation set. Next, it uses the training set to train the model by applying techniques such as anomaly detection algorithms, random forests, and deep learning. Finally, it evaluates the model's performance using the test set, checking precision and recall.

[0369] Predictions using generative AI

[0370] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal then sends this data to the server in real time. The server then uses generative AI and a trained model to analyze the latest data and compare it with past data. This allows it to predict whether there is a possibility of window dressing.

[0371] Real-time notifications

[0372] Based on the prediction results, the server generates a notification containing specific suspicious points. This notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structures. The server sends this notification in real time to the terminal, which receives the notification and displays it to the user (accountant). This allows the user to immediately identify suspicious transactions or unusual financial data and begin detailed audit procedures.

[0373] Specific examples

[0374] Examples of data collection

[0375] The server collects Company A's financial statement data, audit reports, and related publicly available news articles for the past 10 years.

[0376] Specific examples of data preprocessing

[0377] The server removes duplicate entries from the collected data, uses statistical methods to fill in missing sales data from other relevant financial data, standardizes all numerical data to a single currency (e.g., USD), and converts dates to "YYYY-MM-DD."

[0378] Specific examples of building learning models

[0379] The server divides the preprocessed data into a training set (70%), a test set (20%), and a validation set (10%). An anomaly detection algorithm is used to extract features, such as sales spikes and cost structure fluctuations. The model is then trained using a random forest algorithm, and its accuracy is verified on the test set.

[0380] Examples of predictions using generative AI

[0381] The user (accountant) inputs the latest financial data for Company B, and the device sends the data to the server. The server analyzes the data using generative AI and detects that sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, predicting a high possibility of window dressing.

[0382] Real-time notification examples

[0383] Based on the forecast results, the server generates a notification stating specific suspicious points, such as "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." The server sends this result to the terminal, which receives it and displays it to the user (accountant). Based on this, the user can immediately start a detailed audit.

[0384] In this way, by implementing the system of the present invention, it is possible to efficiently detect signs of window dressing and respond quickly.

[0385] The processing flow will be explained below.

[0386] Step 1:

[0387] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from the internet and internal systems. This data comes from a variety of sources, including publicly available company financial statements, third-party auditor reports, and even public reports.

[0388] Step 2:

[0389] The server cleans the collected data, removing duplicates and errors, filling in missing data from other data, and standardizing the format of numeric data and dates.

[0390] Step 3:

[0391] The server uses the preprocessed data to build machine learning models by splitting the data into training, test, and validation sets and applying techniques such as anomaly detection algorithms, random forests, and deep learning to train the models.

[0392] Step 4:

[0393] The server evaluates the trained model on the test set to check precision and recall, and retrains the model as needed to optimize performance.

[0394] Step 5:

[0395] The user (accountant) inputs the latest financial statement data of the company being audited into the terminal.

[0396] Step 6:

[0397] The terminal transmits the input data to the server in real time.

[0398] Step 7:

[0399] The server uses generative AI and trained models to analyze the latest financial data it receives, comparing it with historical data to determine whether any unusual patterns exist.

[0400] Step 8:

[0401] As a result of the analysis, the server predicts the possibility of accounting fraud, including specific suspicious points such as sudden increases in sales or fluctuations in cost structure.

[0402] Step 9:

[0403] The server generates a notification containing specific suspicious points based on the prediction results.

[0404] Step 10:

[0405] The server sends this notification to the terminal in real time.

[0406] Step 11:

[0407] The terminal receives the notification and displays it to the user (accountant), who can then immediately initiate a detailed audit.

[0408] Example 1

[0409] Next, a description will be given of Example 1. In the following description, the data processing device 12 will be referred to as a "server" and the smart glasses 214 will be referred to as a "terminal."

[0410] The challenge is to efficiently and quickly detect accounting fraud in corporate financial data and reduce the burden of accounting audits. Conventional methods require manual analysis of large amounts of data, which is prone to errors and is time-consuming and costly. In addition, advanced algorithms to improve the accuracy of anomaly detection have not been fully utilized.

[0411] The specific processing by the specific processing unit 290 of the data processing device 12 in the first embodiment is realized by the following means.

[0412] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using a generative AI, and a real-time notification means. This enables comprehensive collection and preprocessing of corporate financial data, audit reports, news articles, data on past window dressing cases, etc., and construction of a machine learning model for highly accurate anomaly detection. Furthermore, the use of generative AI enables real-time analysis and prediction based on the latest data entered by the user, and immediate notification can be sent if an anomaly is detected, enabling a rapid response.

[0413] "Data collection means" refers to the ability to collect corporate financial data, audit reports, news articles, and past accounting fraud cases from online and offline sources.

[0414] "Data preprocessing means" refers to functions for cleansing collected data, filling in missing data, and normalizing data.

[0415] "Machine learning model construction means" refers to the function of using preprocessed data to divide the data into a training set, a test set, and a validation set, constructing a model using an anomaly detection algorithm, random forest, or deep learning, and evaluating the accuracy and recall.

[0416] "Prediction method using generative AI" refers to a function that receives the latest financial results data of a company entered by the user, analyzes the data using a learning model and generative AI, detects anomalies, and predicts the possibility of window dressing.

[0417] The "real-time notification means" refers to a function that generates a notification containing suspicious points based on the prediction results, and transmits and displays the notification to the user terminal.

[0418] The system of the present invention is composed of a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, and a real-time notification means. This system is designed to efficiently detect corporate window dressing and reduce the burden of accounting audits.

[0419] Data collection

[0420] The server collects company financial data, audit reports, news articles, and data on past accounting fraud cases via the internet and internal systems. To collect data, it uses Python's BeautifulSoup library and Selenium to perform web scraping. It also uses an ERP system to extract data from the company's internal systems.

[0421] Example: A server sends an HTTP request to retrieve Company A's financial statements from a financial information website, extracts the necessary data from the HTML using BeautifulSoup, and retrieves financial data from the company's ERP system via an API.

[0422] Data Preprocessing

[0423] The server uses data processing libraries such as Pandas and NumPy to preprocess the collected data, specifically deleting duplicate data, correcting erroneous data, filling in missing data, standardizing the currency units of numeric data, and standardizing date formats.

[0424] Example: The server removes duplicate entries from the collected dataset using Pandas' drop_duplicates() method, fills missing data with the average value using the fillna() method, converts all currency units to USD, and standardizes the date format to "YYYY-MM-DD".

[0425] Building a learning model

[0426] The server builds a machine learning model using the preprocessed data. It divides the data into a training set, a test set, and a validation set, and trains the model using an anomaly detection algorithm (e.g., Isolation Forest), a random forest, or a deep learning framework (e.g., TensorFlow). It evaluates the precision and recall of the models and selects the best model.

[0427] Example: The server splits the preprocessed data using scikit-learn's train_test_split function, trains the model using the Isolation Forest algorithm, runs the training using the fit() method, and evaluates the test data using the predict() method.

[0428] Predictions using generative AI

[0429] A user (accountant) inputs the company's latest financial data into a terminal, which then sends the data to a server. The server uses a generative AI model (e.g., GPT-4) and a trained model to analyze the data, detect anomalies, and predict possible window dressing.

[0430] Example: A user enters financial data for Company B into a form on a terminal and presses the "Submit" button. The terminal sends the data in JSON format to the server as a POST request. The server inputs the received data into the GPT-4 generation AI, which generates a prompt such as "Please analyze the sudden increase in sales and fluctuations in operating profit margins for Company B."

[0431] Real-time notifications

[0432] The server generates a notification containing suspicious points based on the prediction results and sends it to the user's terminal. The user can receive this notification and perform a detailed audit based on the displayed abnormal points.

[0433] Example: Based on the prediction results, the server generates a notification such as, "Company B's second-quarter sales have increased by 50% compared to the same month last year, but there is no significant change in its operating profit margin, so there is a high possibility that this is an abnormal transaction," and sends it to the terminal. The terminal receives this notification and displays it to the user as a pop-up or email notification.

[0434] In this way, this system can efficiently detect corporate window dressing and enable rapid response.

[0435] The flow of the identification process in the first embodiment will be described with reference to FIG.

[0436] Step 1:

[0437] The server uses data collection tools to collect data from the internet and internal company systems. Specifically, it uses web scraping tools such as BeautifulSoup and Selenium to obtain company financial data, audit reports, news articles, and data on past accounting fraud cases. It also collects internal data from the company's ERP system via APIs.

[0438] Input: A list of URLs or API endpoints for data collection.

[0439] Output: HTML page and / or JSON formatted data.

[0440] What it does: The server crawls financial information website URLs and sends HTTP requests. It then parses the HTML content with BeautifulSoup to extract financial data. It also uses the API of the company's ERP system to obtain an authentication token and collect the necessary financial data in JSON format.

[0441] Step 2:

[0442] The server preprocesses the collected data using data preprocessing tools, such as data cleansing, removing duplicates, filling in missing data, standardizing numeric data, and standardizing date formats. This is done using data processing libraries such as Pandas and NumPy.

[0443] Input: Raw data such as collected financial data and audit reports.

[0444] Output: A preprocessed clean dataset.

[0445] Specific operation: The server removes duplicate data using Pandas' drop_duplicates() method, fills missing values ​​with the mean using the fillna() method, converts numeric data to a unified currency (e.g., USD), and standardizes the date format to "YYYY-MM-DD".

[0446] Step 3:

[0447] The server uses a machine learning model building tool to build a machine learning model using the preprocessed data. It splits the data into a training set, a test set, and a validation set, and trains the model using an anomaly detection algorithm, random forest, or a deep learning framework (e.g., TensorFlow).

[0448] Input: Preprocessed clean dataset.

[0449] Output: A trained machine learning model.

[0450] Specific operation: The server uses scikit-learn's train_test_split function to split the data into a 70% training set, a 20% test set, and a 10% validation set. It trains the model using the Isolation Forest algorithm or Random Forest, and builds the model using the fit() method. It also evaluates the test data using the predict() method to check precision and recall.

[0451] Step 4:

[0452] A user (accountant) inputs the company's latest financial data into a terminal, which then sends the data to a server. The server uses a generative AI model (e.g., GPT-4) and a trained model to analyze the data, detect anomalies, and predict possible window dressing.

[0453] Input: The most recent financial data entered by the user.

[0454] Output: Anomaly detection results and predicted window dressing probability.

[0455] Specific operation: The user enters the company's financial data into a form on the device and presses the "Submit" button. The device sends this data in JSON format to the server as a POST request. The server inputs the received data into the GPT-4 generation AI, which generates and analyzes the prompt: "Please analyze the sudden increase in sales and fluctuations in operating profit margins for Company B."

[0456] Step 5:

[0457] The server generates a notification containing suspicious points based on the prediction results and sends it to the user terminal in real time. The user can receive this notification and perform a detailed audit based on the displayed abnormalities.

[0458] Input: Anomaly detection results and predicted probability of window dressing.

[0459] Output: Informational message.

[0460] Specific operation: Based on the prediction results, the server generates a notification such as, "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions," and sends it to the terminal in JSON format. The terminal receives this notification and displays it to the user as a pop-up or email notification.

[0461] (Application example 1)

[0462] Next, a description will be given of Application Example 1. In the following description, the data processing device 12 will be referred to as a "server" and the smart glasses 214 will be referred to as a "terminal."

[0463] Currently, detecting corporate accounting fraud requires a great deal of effort and time, making it difficult to detect in real time. Furthermore, it is difficult to ensure the quality and consistency of collected data, which increases the risk of false positives and oversights. For this reason, there is a need to quickly and accurately detect corporate financial fraud and improve the efficiency of accounting audits.

[0464] The specific processing by the specific processing unit 290 of the data processing device 12 in the application example 1 is realized by the following means.

[0465] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using the generative AI, a notification means for sending a notification to a specific device, a means for collecting and analyzing data from multiple information sources of the company in real time, a means for inputting new standardized data into the generative AI model, a means for immediately notifying a user based on the anomaly detection result, and a means for detecting financial fraud at the company using the prediction result of the generative AI. This makes it possible to accurately detect financial fraud at the company in real time and significantly improve the efficiency of accounting audits.

[0466] "Data collection means" refers to means for collecting corporate financial data, audit reports, news articles, and past accounting fraud cases from online and offline sources.

[0467] "Data preprocessing means" refers to means for cleaning collected data, imputing missing data, normalizing data, and standardizing data for input into generative AI models.

[0468] A "machine learning model construction means" is a means for constructing a machine learning model using preprocessed data and using that model to perform anomaly detection and prediction.

[0469] A "predictive method using generative AI" is a method of predicting corporate financial fraud by inputting new standardized data into a generative AI model.

[0470] "Notification means" refers to a means for immediately notifying users of normal or emergency conditions based on the results of anomaly detection.

[0471] "Means for collecting and analyzing data from multiple sources of information for a company in real time" refers to means for continuously collecting and analyzing a company's financial data, audit reports, news articles, and data on past accounting fraud cases in real time.

[0472] "Notification means for sending a notification to a specific device" refers to a means for sending a notification to a specific device such as a smartphone or desktop when financial fraud at a company is detected.

[0473] "Means for inputting new standardized data into the generative AI model" refers to means for inputting newly collected data into the generative AI model and analyzing it in real time.

[0474] "Means for immediately notifying the user based on the anomaly detection results" refers to a means for immediately sending a notification to the user based on an anomaly detected by the generative AI model.

[0475] The system of the present invention aims to efficiently detect corporate window dressing and reduce the burden of accounting audits. This system includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, a notification means for sending notifications to specific devices, a means for collecting and analyzing data from multiple corporate information sources in real time, a means for inputting new standardized data into the generative AI model, a means for immediately notifying users based on the results of anomaly detection, and a means for detecting corporate financial fraud using the prediction results of the generative AI.

[0476] The server collects company financial data, audit reports, news articles, and data on past accounting fraud cases via the Internet and internal systems. This allows data to be collected from companies' published financial statements, reports from third-party auditors, and even public reports. A specific example of data collection would be obtaining Company A's financial data for the past 10 years via an API.

[0477] The server also preprocesses the collected data, removing duplicates and errors, filling in missing data from other data, and standardizing the data format. This results in a clean dataset with consistent numerical data and dates. For example, missing sales data can be filled in using statistical methods from other relevant data.

[0478] The server then uses the preprocessed data to build a machine learning model. The data is divided into a training set, a test set, and a validation set, and the model is trained using techniques such as anomaly detection algorithms, random forests, and deep learning. For example, a random forest model is trained using the training set, and the performance of the constructed model is verified using the test set.

[0479] As a prediction method using generative AI, the user inputs the latest data of the audited company into a terminal and sends the data from the terminal to the server. The server analyzes the latest data using the generative AI model and the trained model and compares it with past data. At this point, if sales have increased significantly compared to the same period last year but the operating profit margin has not changed, it predicts that there is a high possibility of window dressing.

[0480] Finally, the server uses the notification method to notify the user in real time based on the prediction results. The notification details any abnormal patterns and specific suspicious points. For example, the notification might say, "Second-quarter sales increased by 50% compared to the same month last year, but there was no significant change in the operating profit margin, so there is a high possibility of abnormal transactions."

[0481] The hardware used includes a server for collecting and processing data, and a terminal where users can input data and receive notifications. The software includes an API used for data collection and preprocessing, generative AI models and machine learning algorithms used for data analysis, and a communication protocol for notifications. Specifically, Python's pandas and scikit-learn, and deep learning libraries (such as TensorFlow and PyTorch) are used.

[0482] As an example, here is a prompt that uses a generative AI model to analyze new financial data:

[0483] Prompt statement

[0484] Analyze new financial data. This data includes information such as sales, operating profit, net profit, and debt ratio. Also, if sales increase by 50% year-on-year, but the operating profit margin remains unchanged, this may indicate window dressing. Detect and report such abnormal patterns.

[0485] As described above, by implementing the system of the present invention, it is possible to accurately detect corporate window dressing in real time, thereby significantly improving the efficiency of accounting audits.

[0486] The flow of the specific processing in the application example 1 will be described with reference to FIG.

[0487] Step 1:

[0488] The server collects corporate financial data, audit reports, news articles, and data on past accounting fraud cases via the internet and internal systems. Specifically, it obtains this information using APIs. For example, it uses financial data APIs and audit report APIs. The API endpoint and authentication information are required as input, and the collected raw data is obtained as output.

[0489] Step 2:

[0490] The server preprocesses the collected data, removing duplicates and erroneous data and imputing missing data from other data. It also standardizes and normalizes numerical data units and makes date formats consistent. Specifically, it uses the Python pandas library for data cleaning and imputation. The input is the collected raw data, and the output is a preprocessed, clean dataset.

[0491] Step 3:

[0492] The server uses the preprocessed data to build a machine learning model. Specifically, it divides the data into a training set, a test set, and a validation set, and trains it using anomaly detection algorithms, random forests, and deep learning techniques. Libraries used include scikit-learn and TensorFlow. The input is the preprocessed dataset, and the output is a trained machine learning model.

[0493] Step 4:

[0494] The user inputs the latest financial statement data of the company being audited into the terminal. The terminal sends this data to the server in real time. For example, the user inputs financial statements and related data into an input form. The input is the user's latest financial statement data, and the output is the data sent from the terminal to the server.

[0495] Step 5:

[0496] The server analyzes the latest input data using the generative AI model and the trained model. This allows it to compare the latest data with past data and detect abnormal patterns. The prompt text used is something like, "Please analyze the new financial data. This data includes information such as sales, operating profit, net profit, and debt ratio. Furthermore, if sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, this could be window dressing. Please detect and report such abnormal patterns." The input is the latest financial data, and the output is the results of anomaly detection.

[0497] Step 6:

[0498] The server generates a notification containing specific suspicious points based on the prediction results. For example, it generates a notification containing specific content such as, "Sales for the second quarter have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." The input is the anomaly detection result, and the output is the generated notification message.

[0499] Step 7:

[0500] The device displays notifications received from the server to the user in real time. Notifications are displayed as push notifications on smartphones or desktop notifications. Specifically, notifications are received and displayed on the device using push notification APIs or desktop notification APIs. The input is the notification message sent from the server, and the output is the notification that the user can receive and check.

[0501] Furthermore, an emotion engine that estimates the user's emotion may be further combined. That is, the identification processing unit 290 may estimate the user's emotion using the emotion identification model 59, and perform identification processing using the user's emotion.

[0502] System Overview

[0503] The system of the present invention comprises a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, a notification means, and an emotion engine that recognizes user emotions. This system is designed not only to efficiently detect corporate window dressing and reduce the burden of accounting audits, but also to respond according to the user's emotional state.

[0504] Data collection

[0505] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[0506] Data Preprocessing

[0507] The server cleans the collected data. Specifically, it removes duplicates and errors from the data and fills in missing data from other sources. It also standardizes the format of numeric data and dates and normalizes all data. This prepares a clean dataset suitable for analysis.

[0508] Building a learning model

[0509] The server uses the preprocessed data to build a machine learning model. First, it divides the data into a training set, a test set, and a validation set. Next, it uses the training set to train the model by applying techniques such as anomaly detection algorithms, random forests, and deep learning. Finally, it evaluates the model's performance using the test set, checking precision and recall.

[0510] Predictions using generative AI

[0511] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal then sends this data to the server in real time. The server then uses generative AI and a trained model to analyze the latest data and compare it with past data. This allows it to predict whether there is a possibility of window dressing.

[0512] Emotion Engine Operation

[0513] The device's built-in emotion engine analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. For example, it analyzes the user's stress level and sense of security from changes in voice tone and facial expressions. This information is reflected when the prediction results are notified.

[0514] Real-time notifications

[0515] The server generates a notification containing specific suspicious points based on the prediction results. The notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structure. At the same time, the emotion engine determines the user's emotional state and adjusts the tone of the notification and the way detailed information is presented. For example, if the user is feeling stressed, the notification content will be concise and clear.

[0516] Logging

[0517] The server logs the user's emotional state as determined by the emotion engine and stores it in a database for later analysis and improvement. This data helps understand user emotional trends over time and further optimize the system.

[0518] Specific examples

[0519] Examples of data collection

[0520] The server collects Company A's financial statement data, audit reports, and related publicly available news articles for the past 10 years.

[0521] Specific examples of data preprocessing

[0522] The server removes duplicate entries from the collected data, uses statistical methods to fill in missing sales data from other relevant financial data, standardizes all numerical data to a single currency (e.g., USD), and converts dates to "YYYY-MM-DD."

[0523] Specific examples of building learning models

[0524] The server divides the preprocessed data into a training set (70%), a test set (20%), and a validation set (10%). An anomaly detection algorithm is used to extract features, such as sales spikes and cost structure fluctuations. The model is then trained using a random forest algorithm, and its accuracy is verified on the test set.

[0525] Examples of predictions using generative AI

[0526] The user (accountant) inputs the latest financial data for Company B, and the device sends the data to the server. The server analyzes the data using generative AI and detects that sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, predicting a high possibility of window dressing.

[0527] Examples of emotion engines

[0528] The device's emotion engine analyzes the user's (the accountant's) voice and determines that their stress level is high. Based on this information, the server generates short, concise notifications to reduce the user's stress.

[0529] Real-time notification examples

[0530] Based on the forecast results, the server generates a notification with specific suspicious points, such as "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." Because the emotion engine determined the stress level to be high, only the main points are included in the detailed information. The server sends this result to the terminal, which receives it and displays it to the user (accountant). Based on this, the user can immediately start a detailed audit.

[0531] In this way, by implementing the system of the present invention, it is possible to efficiently detect signs of window dressing and take the most appropriate action depending on the user's emotional state.

[0532] The processing flow will be explained below.

[0533] Step 1:

[0534] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[0535] Step 2:

[0536] The server cleans the collected data. Specifically, it removes duplicates and errors from the data. It also supplements missing data from other data and standardizes the format of numerical data and dates. For example, it standardizes data in different currency units and converts them to the "YYYY-MM-DD" format.

[0537] Step 3:

[0538] The server uses the preprocessed data to build a machine learning model by first splitting the data into a training set, a test set, and a validation set, and then using the training set to apply techniques such as anomaly detection algorithms, random forests, or deep learning to train the model.

[0539] Step 4:

[0540] The server evaluates the trained model on the test set to check precision and recall, and retrains the model as needed to optimize performance.

[0541] Step 5:

[0542] The user (accountant) inputs the latest financial statement data of the company being audited into the terminal, such as data on sales, operating profit, and costs.

[0543] Step 6:

[0544] The terminal transmits the input data to the server in real time.

[0545] Step 7:

[0546] The server uses generative AI and trained models to analyze the latest financial data it receives, comparing the input data with past data to detect any abnormal patterns, such as sudden increases in sales or cost inconsistencies.

[0547] Step 8:

[0548] Based on the analysis results, the server predicts the possibility of window dressing, including specific suspicious points such as sudden increases in sales or fluctuations in cost structure.

[0549] Step 9:

[0550] The device's built-in emotion engine analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. For example, it can determine their stress level based on changes in their voice tone and facial expressions.

[0551] Step 10:

[0552] Based on the emotion engine's findings, the server adjusts the content of the notification. For example, if the user is feeling stressed, the notification will be brief and clear.

[0553] Step 11:

[0554] Based on the predictions, the server generates specific suspicious alerts, including details such as unusual sales spikes or fluctuations in cost structures.

[0555] Step 12:

[0556] The server sends this notification to the terminal in real time.

[0557] Step 13:

[0558] The terminal receives the notification and displays it to the user (accountant), who can then immediately begin a detailed audit, for example by gathering additional information on specific items or preparing questions for the company.

[0559] Step 14:

[0560] The server logs the user's emotional state as determined by the emotion engine and stores it in a database for later analysis and improvement, for example, to understand the user's emotional trends over a long period of time and to help further optimize the system.

[0561] Example 2

[0562] Next, a description will be given of Example 2. In the following description, the data processing device 12 will be referred to as a "server" and the smart glasses 214 will be referred to as a "terminal."

[0563] In audit work, it is important to quickly and accurately detect corporate accounting fraud, but conventional methods require a great deal of time and effort from data collection to analysis. Furthermore, notification methods that do not take into account the stress and emotional state of accountants may reduce the efficiency of audit work. To solve these problems, efficient data processing and prediction methods, as well as appropriate responses according to the user's emotional state, are required.

[0564] The specific processing by the specific processing unit 290 of the data processing device 12 in the second embodiment is realized by the following means.

[0565] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, a notification means, and a log recording means, which enable fast and accurate detection of corporate window dressing and provide an optimal notification method according to the user's emotional state.

[0566] "Data collection methods" are methods for collecting corporate financial information, audit reports, news articles, and information on past accounting fraud cases from online and offline sources.

[0567] "Data preprocessing means" refers to means for cleaning collected information, filling in missing data, and normalizing information.

[0568] A "machine learning model building means" is a means for training and building a machine learning model using preprocessed data.

[0569] "Predictive methods using generative AI" are methods that analyze a company's financial information based on machine learning models and predict the possibility of accounting fraud.

[0570] The "emotion recognition means" is a means for analyzing the user's voice or facial expression and determining the user's emotional state.

[0571] The "notification means" is a means for adjusting and providing notification content according to the prediction result and the emotional state determined by the emotion recognition means.

[0572] The "log recording means" is a means for recording the emotional state of the user determined by the emotion recognition means and storing it in a database.

[0573] The system of the present invention includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, a notification means, and a log recording means. This system is designed to efficiently detect corporate window dressing, reduce the burden of accounting audits, and respond according to the user's emotional state.

[0574] First, the server uses data collection tools to collect company financial information, audit reports, news articles, and information on past accounting fraud cases from the Internet and internal systems. Specifically, it efficiently collects the required information using Python libraries (such as BeautifulSoup and Selenium) and APIs (such as FinAPI). For example, it can collect Company A's financial statement data for the past 10 years, audit reports, and publicly available related news articles.

[0575] Next, the server uses data preprocessing tools to clean the collected data, impute missing data, and normalize the data. Specifically, it removes duplicate data using the drop_duplicates() function in the pandas library, detects missing values ​​using the isnull() function, and imputes them using statistical methods. It also converts numeric data to a unified currency unit (e.g., USD) and standardizes date formats to "YYYY-MM-DD" using the datetime module.

[0576] The server then uses the preprocessed data to build a machine learning model using a machine learning model building method. The data set is split into a training set, a test set, and a validation set, and the train_test_split function in the sklearn library is used. The server also uses Isolation Forest or Random Forest Classifier for feature extraction to train and build the model. For example, an anomaly detection algorithm is used to extract features using indicators such as sudden sales increases or fluctuations in cost structure.

[0577] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal sends the data to the server in real time. The server analyzes the input data using a predictive method that uses generative AI, and compares it with past data to predict the possibility of window dressing. The generative AI uses a model such as OpenAI's GPT-4. For example, a prompt statement could be entered as follows: "Enter the latest financial data of Company B, have the server analyze it, and detect that sales have increased by 50% compared to the same period last year, but the operating profit margin has remained unchanged, and predict that there is a high possibility of window dressing."

[0578] Furthermore, the device's built-in emotion recognition means analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. It uses the librosa library for voice analysis and the OpenCV and dlib libraries for facial expression recognition. For example, it detects changes in voice tone or abnormal facial expressions to determine whether the user is in a high-stress state. This information is reflected in the notification content sent by the prediction result notification means.

[0579] The server uses a real-time notification mechanism to generate a notification containing specific suspicious points based on the prediction results. This notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structure. The notification content is adjusted to be concise and clear, taking into account the user's emotional information determined by the emotion recognition mechanism. For example, a notification may be generated that states, "Sales in the second quarter have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." This notification is sent to the terminal and displayed to the user (accountant).

[0580] Finally, the server uses the logging means to log the user's emotional state as determined by the emotion recognition means and stores the log in a database. This data is used to understand the user's emotional trends over a long period of time and to further optimize the system. For example, if a user is in a state of high stress under certain conditions, this data can be used as feedback to revise the system's interface and notification methods.

[0581] The above configuration and processing make it possible to quickly and accurately detect corporate window dressing and respond optimally according to the user's emotional state.

[0582] The flow of the identification process in the second embodiment will be described with reference to FIG.

[0583] Processing flow

[0584] Step 1: Collect data

[0585] The server collects corporate financial information, audit reports, news articles, and information on past accounting fraud cases via the internet and internal systems. It efficiently collects the necessary data using libraries such as Python's BeautifulSoup and Selenium, and APIs such as FinAPI. Specifically, it retrieves data from the API using the requests library, and parses HTML with BeautifulSoup to extract financial information.

[0586] Input: URL or API endpoint to collect data

[0587] Output: Dataset of company financial information, audit reports, press articles, and information on past accounting fraud cases

[0588] Step 2: Preprocessing the data

[0589] The server cleans the collected data, imputes missing data, and normalizes the data. Specifically, it uses the pandas library to remove duplicate data, the isnull() function to detect missing values, and statistical methods to impute them. It also standardizes date formats using the datetime module and converts numeric data to a uniform currency unit.

[0590] Input: Dataset of company financial information, audit reports, news articles, and information on past accounting fraud cases

[0591] Output: A clean and normalized dataset

[0592] Step 3: Building a learning model

[0593] The server builds a machine learning model using the preprocessed data. It splits the dataset into training, test, and validation sets using the train_test_split function in the sklearn library, and uses IsolationForest and RandomForestClassifier for feature extraction. After training the model, it evaluates its accuracy on the test set.

[0594] Input: Clean and normalized dataset

[0595] Output: A trained machine learning model

[0596] Step 4: Predict using generative AI

[0597] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal sends the data in real time to the server. The server uses generative AI and a trained model to analyze the input data and predict the possibility of window dressing. The generative AI uses a model such as GPT-4.

[0598] Input: Latest financial data

[0599] Output: Prediction result regarding the possibility of window dressing

[0600] Step 5: Emotion Engine in Action

[0601] The emotion engine installed on the device analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. Changes in voice tone are analyzed using librosa, and facial expressions are evaluated using OpenCV and dlib. This determines stress levels and emotional states.

[0602] Input: User's voice data, facial expression data

[0603] Output: User's emotional state

[0604] Step 6: Real-time notifications

[0605] The server generates a notification based on the prediction results, including specific suspicious points, such as sudden sales increases or fluctuations in cost structure, and adjusts the content and tone of the notification based on the emotional information determined by the emotion recognition means.

[0606] Input: Prediction result, user's emotional state

[0607] Output: Adjusted notification content

[0608] Step 7: Logging

[0609] The server records the user's emotional state as determined by the emotion engine as a log and stores it in a database. The log data is used to improve the system and optimize user responses.

[0610] Input: User's emotional state

[0611] Output: Saved log data

[0612] (Application example 2)

[0613] Next, a description will be given of Application Example 2. In the following description, the data processing device 12 will be referred to as a "server" and the smart glasses 214 will be referred to as a "terminal."

[0614] In modern business activities, window dressing remains a serious problem, placing a heavy burden on accountants and auditors. Audit work also creates significant stress and strain. To address this issue, a system is needed that efficiently detects abnormal transactions and provides appropriate notifications based on the accountant's emotional state. The present invention aims to improve corporate reliability and reduce the burden of accounting audits by solving these problems.

[0615] The specific processing by the specific processing unit 290 of the data processing device 12 in the application example 2 is realized by the following means.

[0616] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, and a real-time notification means, which enable efficient detection of anomalies in a company's financial data and appropriate notification according to the user's emotional state.

[0617] "Data collection methods" are methods for collecting financial data, audit reports, information articles, and historical anomaly data from online or offline sources.

[0618] "Data preprocessing means" refers to methods for cleaning collected data, filling in missing data, and normalizing the data into a format that is easy to analyze.

[0619] "Machine learning model building means" refers to a method of using pre-processed data to apply anomaly detection algorithms and other machine learning techniques to build a model for detecting anomalous transactions.

[0620] "Prediction methods using generative AI" are methods that use trained machine learning models and generative AI to analyze newly obtained data and predict the possibility of abnormal transactions.

[0621] "Emotion recognition means" is a method for determining the user's emotional state by analyzing the user's voice and facial expressions, and adjusting the system's response based on that information.

[0622] "Real-time notification means" is a method of generating a notification based on the results of abnormal transaction detection, with content that corresponds to the user's emotional state, and quickly notifying the user.

[0623] The present invention is a system for monitoring and detecting anomalies in corporate financial data, and is intended to be used on end-user devices such as smartphone applications. Specific procedures for implementing the invention are described below.

[0624] Hardware and software used

[0625] Hardware: Smartphones, cloud servers

[0626] Software: Python, Scikit-learn, Keras, EmotionEngine, NotificationService

[0627] System configuration

[0628] 1. Data collection methods:

[0629] The server collects financial data, audit reports, news articles, and past anomaly data from online and offline sources. For example, financial data is collected through APIs, and audit reports and news articles are collected using a web crawling system.

[0630] 2. Data preprocessing methods:

[0631] Once collected, the data is filled in with missing values, errors corrected, and formatted in a consistent manner (for example, missing sales data is filled in from historical data, and date formats are standardized).

[0632] 3. Machine learning model building methods:

[0633] The preprocessed data is used to apply an anomaly detection algorithm and build a model. Specifically, the data is divided into a training set and a test set, and training is performed using a random forest or deep learning model.

[0634] 4. Generative AI prediction methods:

[0635] When a user inputs their latest financial data, the data is sent to the server, where generative AI and trained models are used to predict anomalies. For example, if sales have increased significantly but the operating profit margin remains unchanged, the system predicts possible anomalies.

[0636] 5. Emotion recognition means:

[0637] Analyzes the user's voice and facial expressions to determine their emotional state. Emotion Engine is used to analyze voice tone and facial expression changes to determine the user's stress level.

[0638] 6. Real-time notification methods:

[0639] Based on the anomaly detection results, notifications are generated with content that corresponds to the user's emotional state. For example, if the user is particularly stressed, the notification content will be brief and only the main points will be presented.

[0640] Specific examples of realization

[0641] 1. Examples of data collection:

[0642] Use APIs to retrieve company financial data and collect audit reports for the past 10 years.

[0643] 2. Examples of data preprocessing:

[0644] Remove duplicate entries from collected data, impute missing data using statistical methods, convert all numerical data to a single currency, and standardize date formats.

[0645] 3. Specific examples of building learning models:

[0646] The preprocessed data is used to train a model using the random forest algorithm, and the performance of the anomaly detection model is verified.

[0647] 4. Specific examples of predictions using generative AI:

[0648] Users enter their company's latest financial data, which is then analyzed on the server, and if there is a high possibility of an abnormality, an alert is issued immediately.

[0649] 5. Examples of emotion recognition:

[0650] The system analyzes the user's voice and, if it determines that the stress level is high, simplifies the notification content.

[0651] 6. Examples of real-time notifications:

[0652] "Second-quarter sales have surged 50% compared to the same month last year, but there has been no significant change in operating profit margins, so there is a high possibility that this is an abnormal transaction," the notice said.

[0653] Prompt Sentence Examples

[0654] "Evaluate a company's latest financial data and detect anomalies. Write a program that adjusts the tone of notifications based on the user's stress level."

[0655] The flow of the specific processing in the application example 2 will be described with reference to FIG.

[0656] Step 1:

[0657] Data collection by data collection method

[0658] The server collects financial data, audit reports, information articles, and historical anomaly data from online and offline sources. Specifically, it retrieves financial data using APIs and collects news articles using a web crawling system. It uses the URL or API endpoint of each source as input and obtains the collected raw data as output.

[0659] Step 2:

[0660] Data cleaning through data preprocessing methods

[0661] The server cleans the collected raw data, imputes missing data, and normalizes the data into a format that is easy to analyze. Specifically, it removes duplicate data, imputes missing data (e.g., imputation using statistical methods), and standardizes date formats. It uses the collected raw data as input and obtains a cleaned dataset as output.

[0662] Step 3:

[0663] Building models using machine learning model building methods

[0664] The server uses the preprocessed data to apply an anomaly detection algorithm and build a model. Specifically, it splits the data into a training set and a test set, and trains the model using a random forest or deep learning algorithm. The preprocessed dataset is used as input, and the trained model is obtained as output.

[0665] Step 4:

[0666] Predicting anomalies using generative AI-based predictive methods

[0667] When a user inputs their latest financial data, the device sends the data to the server. The server then uses generative AI and trained models to analyze this data and predict anomalies. Specifically, it detects sudden increases in sales or inconsistencies in cost structures. The latest financial data provided by the user is used as input, and a prediction result indicating whether or not anomalies exist is obtained as output.

[0668] Step 5:

[0669] Emotional state analysis using emotion recognition tools

[0670] The device analyzes the user's voice and facial expressions in real time to determine their emotional state. Specifically, it uses the Emotion Engine to analyze changes in voice tone and facial expressions to determine stress levels and feelings of relief. The device uses the user's voice and facial expression data as input and outputs the determined emotional state.

[0671] Step 6:

[0672] Generate and send notifications via real-time notification methods

[0673] The server generates a notification based on the anomaly detection results, with content appropriate to the user's emotional state. Specifically, if an anomaly is detected, a detailed explanation is provided, and if the user is in a stressful state, a brief notification is provided. The anomaly detection results and the user's emotional state are used as inputs, and the generated notification message is obtained as output.

[0674] The specific processing unit 290 transmits the result of the specific processing to the smart glasses 214. In the smart glasses 214, the control unit 46A causes the speaker 240 to output the result of the specific processing. The microphone 238 acquires audio indicating a user input regarding the result of the specific processing. The control unit 46A transmits audio data indicating the user input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the audio data.

[0675] The data generation model 58 is a so-called generative AI (Artificial Intelligence). An example of the data generation model 58 is ChatGPT (Internet Search<URL: https: / / openai.com / blog / chatgpt> ), Gemini (Internet search <url: https: gemini.google.com ?hl="ja">) and other generation AIs. The data generation model 58 is obtained by performing deep learning on a neural network. A prompt including an instruction is input to the data generation model 58, and inference data such as voice data indicating voice, text data indicating text, and image data indicating an image is also input. The data generation model 58 performs inference on the input inference data in accordance with the instruction indicated by the prompt, and outputs the inference result in a data format such as voice data and text data. Here, inference refers to, for example, analysis, classification, prediction, and / or summarization.

[0676] In the above embodiment, an example in which the specific processing is performed by the data processing device 12 has been given, but the technology of the present disclosure is not limited to this, and the specific processing may be performed by the smart glasses 214.

[0677] [Third embodiment]

[0678] FIG. 5 shows an example of the configuration of a data processing system 310 according to the third embodiment.

[0679] 5, the data processing system 310 includes the data processing device 12 and a headset terminal 314. An example of the data processing device 12 is a server.

[0680] The data processing device 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" according to the technology of the present disclosure. The computer 22 includes a processor 28, a RAM 30, and a storage 32. The processor 28, the RAM 30, and the storage 32 are connected to a bus 34. The database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. Examples of the network 54 include a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[0681] The headset type terminal 314 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a display 343. The computer 36 includes a processor 46, a RAM 48, and a storage 50. The processor 46, the RAM 48, and the storage 50 are connected to a bus 52. The microphone 238, the speaker 240, the camera 42, and the display 343 are also connected to the bus 52.

[0682] The microphone 238 receives instructions and the like from the user 20 by receiving voice uttered by the user 20. The microphone 238 captures the voice uttered by the user 20, converts the captured voice into audio data, and outputs it to the processor 46. The speaker 240 outputs audio in accordance with instructions from the processor 46.

[0683] Camera 42 is a small digital camera equipped with an optical system including a lens, aperture, and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, and captures images of the surroundings of user 20 (for example, an imaging range defined by an angle of view equivalent to the field of vision of a typical healthy person).

[0684] The communication I / F 44 is connected to a network 54. The communication I / Fs 44 and 26 are responsible for the exchange of various information between the processor 46 and the processor 28 via the network 54. The exchange of various information between the processor 46 and the processor 28 using the communication I / Fs 44 and 26 is carried out in a secure state.

[0685] Fig. 6 shows an example of the main functions of the data processing device 12 and the headset type terminal 314. As shown in Fig. 6, in the data processing device 12, a specific process is performed by the processor 28. A specific process program 56 is stored in the storage 32.

[0686] The specific processing program 56 is an example of a "program" according to the technology of the present disclosure. The processor 28 reads the specific processing program 56 from the storage 32 and executes the read specific processing program 56 on the RAM 30. The specific processing is realized by the processor 28 operating as a specific processing unit 290 in accordance with the specific processing program 56 executed on the RAM 30.

[0687] The storage 32 stores a data generation model 58 and an emotion identification model 59. The data generation model 58 and the emotion identification model 59 are used by the identification processing unit 290.

[0688] In the headset type terminal 314, a reception output process is performed by the processor 46. A reception output program 60 is stored in the storage 50. The processor 46 reads the reception output program 60 from the storage 50 and executes the read reception output program 60 on the RAM 48. The reception output process is realized by the processor 46 operating as the control unit 46A in accordance with the reception output program 60 executed on the RAM 48.

[0689] Next, a description will be given of the identification process performed by the identification processing unit 290 of the data processing device 12. In the following description, the data processing device 12 will be referred to as the "server" and the headset type terminal 314 will be referred to as the "terminal."

[0690] System Overview

[0691] The system of the present invention includes data collection, data preprocessing, machine learning model construction, prediction using generative AI, and notification methods. This system is designed to efficiently detect corporate window dressing and reduce the burden of accounting audits.

[0692] Data collection

[0693] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[0694] Data Preprocessing

[0695] The server preprocesses the collected data. Specifically, it removes duplicates and errors from the data and fills in missing data from other sources. It also standardizes the format of numeric data and dates and normalizes all data. This prepares a clean dataset suitable for analysis.

[0696] Building a learning model

[0697] The server uses the preprocessed data to build a machine learning model. First, it divides the data into a training set, a test set, and a validation set. Next, it uses the training set to train the model by applying techniques such as anomaly detection algorithms, random forests, and deep learning. Finally, it evaluates the model's performance using the test set, checking precision and recall.

[0698] Predictions using generative AI

[0699] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal then sends this data to the server in real time. The server then uses generative AI and a trained model to analyze the latest data and compare it with past data. This allows it to predict whether there is a possibility of window dressing.

[0700] Real-time notifications

[0701] Based on the prediction results, the server generates a notification containing specific suspicious points. This notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structures. The server sends this notification in real time to the terminal, which receives the notification and displays it to the user (accountant). This allows the user to immediately identify suspicious transactions or unusual financial data and begin detailed audit procedures.

[0702] Specific examples

[0703] Examples of data collection

[0704] The server collects Company A's financial statement data, audit reports, and related publicly available news articles for the past 10 years.

[0705] Specific examples of data preprocessing

[0706] The server removes duplicate entries from the collected data, uses statistical methods to fill in missing sales data from other relevant financial data, standardizes all numerical data to a single currency (e.g., USD), and converts dates to "YYYY-MM-DD."

[0707] Specific examples of building learning models

[0708] The server divides the preprocessed data into a training set (70%), a test set (20%), and a validation set (10%). An anomaly detection algorithm is used to extract features, such as sales spikes and cost structure fluctuations. The model is then trained using a random forest algorithm, and its accuracy is verified on the test set.

[0709] Examples of predictions using generative AI

[0710] The user (accountant) inputs the latest financial data for Company B, and the device sends the data to the server. The server analyzes the data using generative AI and detects that sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, predicting a high possibility of window dressing.

[0711] Real-time notification examples

[0712] Based on the forecast results, the server generates a notification stating specific suspicious points, such as "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." The server sends this result to the terminal, which receives it and displays it to the user (accountant). Based on this, the user can immediately start a detailed audit.

[0713] In this way, by implementing the system of the present invention, it is possible to efficiently detect signs of window dressing and respond quickly.

[0714] The processing flow will be explained below.

[0715] Step 1:

[0716] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from the internet and internal systems. This data comes from a variety of sources, including publicly available company financial statements, third-party auditor reports, and even public reports.

[0717] Step 2:

[0718] The server cleans the collected data, removing duplicates and errors, filling in missing data from other data, and standardizing the format of numeric data and dates.

[0719] Step 3:

[0720] The server uses the preprocessed data to build machine learning models by splitting the data into training, test, and validation sets and applying techniques such as anomaly detection algorithms, random forests, and deep learning to train the models.

[0721] Step 4:

[0722] The server evaluates the trained model on the test set to check precision and recall, and retrains the model as needed to optimize performance.

[0723] Step 5:

[0724] The user (accountant) inputs the latest financial statement data of the company being audited into the terminal.

[0725] Step 6:

[0726] The terminal transmits the input data to the server in real time.

[0727] Step 7:

[0728] The server uses generative AI and trained models to analyze the latest financial data it receives, comparing it with historical data to determine whether any unusual patterns exist.

[0729] Step 8:

[0730] As a result of the analysis, the server predicts the possibility of accounting fraud, including specific suspicious points such as sudden increases in sales or fluctuations in cost structure.

[0731] Step 9:

[0732] The server generates a notification containing specific suspicious points based on the prediction results.

[0733] Step 10:

[0734] The server sends this notification to the terminal in real time.

[0735] Step 11:

[0736] The terminal receives the notification and displays it to the user (accountant), who can then immediately initiate a detailed audit.

[0737] Example 1

[0738] Next, a description will be given of Example 1. In the following description, the data processing device 12 will be referred to as a "server" and the headset type terminal 314 will be referred to as a "terminal."

[0739] The challenge is to efficiently and quickly detect accounting fraud in corporate financial data and reduce the burden of accounting audits. Conventional methods require manual analysis of large amounts of data, which is prone to errors and is time-consuming and costly. In addition, advanced algorithms to improve the accuracy of anomaly detection have not been fully utilized.

[0740] The specific processing by the specific processing unit 290 of the data processing device 12 in the first embodiment is realized by the following means.

[0741] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using a generative AI, and a real-time notification means. This enables comprehensive collection and preprocessing of corporate financial data, audit reports, news articles, data on past window dressing cases, etc., and construction of a machine learning model for highly accurate anomaly detection. Furthermore, the use of generative AI enables real-time analysis and prediction based on the latest data entered by the user, and immediate notification can be sent if an anomaly is detected, enabling a rapid response.

[0742] "Data collection means" refers to the ability to collect corporate financial data, audit reports, news articles, and past accounting fraud cases from online and offline sources.

[0743] "Data preprocessing means" refers to functions for cleansing collected data, filling in missing data, and normalizing data.

[0744] "Machine learning model construction means" refers to the function of using preprocessed data to divide the data into a training set, a test set, and a validation set, constructing a model using an anomaly detection algorithm, random forest, or deep learning, and evaluating the accuracy and recall.

[0745] "Prediction method using generative AI" refers to a function that receives the latest financial results data of a company entered by the user, analyzes the data using a learning model and generative AI, detects anomalies, and predicts the possibility of window dressing.

[0746] The "real-time notification means" refers to a function that generates a notification containing suspicious points based on the prediction results, and transmits and displays the notification to the user terminal.

[0747] The system of the present invention is composed of a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, and a real-time notification means. This system is designed to efficiently detect corporate window dressing and reduce the burden of accounting audits.

[0748] Data collection

[0749] The server collects company financial data, audit reports, news articles, and data on past accounting fraud cases via the internet and internal systems. To collect data, it uses Python's BeautifulSoup library and Selenium to perform web scraping. It also uses an ERP system to extract data from the company's internal systems.

[0750] Example: A server sends an HTTP request to retrieve Company A's financial statements from a financial information website, extracts the necessary data from the HTML using BeautifulSoup, and retrieves financial data from the company's ERP system via an API.

[0751] Data Preprocessing

[0752] The server uses data processing libraries such as Pandas and NumPy to preprocess the collected data, specifically deleting duplicate data, correcting erroneous data, filling in missing data, standardizing the currency units of numeric data, and standardizing date formats.

[0753] Example: The server removes duplicate entries from the collected dataset using Pandas' drop_duplicates() method, fills missing data with the average value using the fillna() method, converts all currency units to USD, and standardizes the date format to "YYYY-MM-DD".

[0754] Building a learning model

[0755] The server builds a machine learning model using the preprocessed data. It divides the data into a training set, a test set, and a validation set, and trains the model using an anomaly detection algorithm (e.g., Isolation Forest), a random forest, or a deep learning framework (e.g., TensorFlow). It evaluates the precision and recall of the models and selects the best model.

[0756] Example: The server splits the preprocessed data using scikit-learn's train_test_split function, trains the model using the Isolation Forest algorithm, runs the training using the fit() method, and evaluates the test data using the predict() method.

[0757] Predictions using generative AI

[0758] A user (accountant) inputs the company's latest financial data into a terminal, which then sends the data to a server. The server uses a generative AI model (e.g., GPT-4) and a trained model to analyze the data, detect anomalies, and predict possible window dressing.

[0759] Example: A user enters financial data for Company B into a form on a terminal and presses the "Submit" button. The terminal sends the data in JSON format to the server as a POST request. The server inputs the received data into the GPT-4 generation AI, which generates a prompt such as "Please analyze the sudden increase in sales and fluctuations in operating profit margins for Company B."

[0760] Real-time notifications

[0761] The server generates a notification containing suspicious points based on the prediction results and sends it to the user's terminal. The user can receive this notification and perform a detailed audit based on the displayed abnormal points.

[0762] Example: Based on the prediction results, the server generates a notification such as, "Company B's second-quarter sales have increased by 50% compared to the same month last year, but there is no significant change in its operating profit margin, so there is a high possibility that this is an abnormal transaction," and sends it to the terminal. The terminal receives this notification and displays it to the user as a pop-up or email notification.

[0763] In this way, this system can efficiently detect corporate window dressing and enable rapid response.

[0764] The flow of the identification process in the first embodiment will be described with reference to FIG.

[0765] Step 1:

[0766] The server uses data collection tools to collect data from the internet and internal company systems. Specifically, it uses web scraping tools such as BeautifulSoup and Selenium to obtain company financial data, audit reports, news articles, and data on past accounting fraud cases. It also collects internal data from the company's ERP system via APIs.

[0767] Input: A list of URLs or API endpoints for data collection.

[0768] Output: HTML page and / or JSON formatted data.

[0769] What it does: The server crawls financial information website URLs and sends HTTP requests. It then parses the HTML content with BeautifulSoup to extract financial data. It also uses the API of the company's ERP system to obtain an authentication token and collect the necessary financial data in JSON format.

[0770] Step 2:

[0771] The server preprocesses the collected data using data preprocessing tools, such as data cleansing, removing duplicates, filling in missing data, standardizing numeric data, and standardizing date formats. This is done using data processing libraries such as Pandas and NumPy.

[0772] Input: Raw data such as collected financial data and audit reports.

[0773] Output: A preprocessed clean dataset.

[0774] Specific operation: The server removes duplicate data using Pandas' drop_duplicates() method, fills missing values ​​with the mean using the fillna() method, converts numeric data to a unified currency (e.g., USD), and standardizes the date format to "YYYY-MM-DD".

[0775] Step 3:

[0776] The server uses a machine learning model building tool to build a machine learning model using the preprocessed data. It splits the data into a training set, a test set, and a validation set, and trains the model using an anomaly detection algorithm, random forest, or a deep learning framework (e.g., TensorFlow).

[0777] Input: Preprocessed clean dataset.

[0778] Output: A trained machine learning model.

[0779] Specific operation: The server uses scikit-learn's train_test_split function to split the data into a 70% training set, a 20% test set, and a 10% validation set. It trains the model using the Isolation Forest algorithm or Random Forest, and builds the model using the fit() method. It also evaluates the test data using the predict() method to check precision and recall.

[0780] Step 4:

[0781] A user (accountant) inputs the company's latest financial data into a terminal, which then sends the data to a server. The server uses a generative AI model (e.g., GPT-4) and a trained model to analyze the data, detect anomalies, and predict possible window dressing.

[0782] Input: The most recent financial data entered by the user.

[0783] Output: Anomaly detection results and predicted window dressing probability.

[0784] Specific operation: The user enters the company's financial data into a form on the device and presses the "Submit" button. The device sends this data in JSON format to the server as a POST request. The server inputs the received data into the GPT-4 generation AI, which generates and analyzes the prompt: "Please analyze the sudden increase in sales and fluctuations in operating profit margins for Company B."

[0785] Step 5:

[0786] The server generates a notification containing suspicious points based on the prediction results and sends it to the user terminal in real time. The user can receive this notification and perform a detailed audit based on the displayed abnormalities.

[0787] Input: Anomaly detection results and predicted probability of window dressing.

[0788] Output: Informational message.

[0789] Specific operation: Based on the prediction results, the server generates a notification such as, "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions," and sends it to the terminal in JSON format. The terminal receives this notification and displays it to the user as a pop-up or email notification.

[0790] (Application example 1)

[0791] Next, a description will be given of Application Example 1. In the following description, the data processing device 12 will be referred to as a "server" and the headset type terminal 314 will be referred to as a "terminal."

[0792] Currently, detecting corporate accounting fraud requires a great deal of effort and time, making it difficult to detect in real time. Furthermore, it is difficult to ensure the quality and consistency of collected data, which increases the risk of false positives and oversights. For this reason, there is a need to quickly and accurately detect corporate financial fraud and improve the efficiency of accounting audits.

[0793] The specific processing by the specific processing unit 290 of the data processing device 12 in the application example 1 is realized by the following means.

[0794] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using the generative AI, a notification means for sending a notification to a specific device, a means for collecting and analyzing data from multiple information sources of the company in real time, a means for inputting new standardized data into the generative AI model, a means for immediately notifying a user based on the anomaly detection result, and a means for detecting financial fraud at the company using the prediction result of the generative AI. This makes it possible to accurately detect financial fraud at the company in real time and significantly improve the efficiency of accounting audits.

[0795] "Data collection means" refers to means for collecting corporate financial data, audit reports, news articles, and past accounting fraud cases from online and offline sources.

[0796] "Data preprocessing means" refers to means for cleaning collected data, imputing missing data, normalizing data, and standardizing data for input into generative AI models.

[0797] A "machine learning model construction means" is a means for constructing a machine learning model using preprocessed data and using that model to perform anomaly detection and prediction.

[0798] A "predictive method using generative AI" is a method of predicting corporate financial fraud by inputting new standardized data into a generative AI model.

[0799] "Notification means" refers to a means for immediately notifying users of normal or emergency conditions based on the results of anomaly detection.

[0800] "Means for collecting and analyzing data from multiple sources of information for a company in real time" refers to means for continuously collecting and analyzing a company's financial data, audit reports, news articles, and data on past accounting fraud cases in real time.

[0801] "Notification means for sending a notification to a specific device" refers to a means for sending a notification to a specific device such as a smartphone or desktop when financial fraud at a company is detected.

[0802] "Means for inputting new standardized data into the generative AI model" refers to means for inputting newly collected data into the generative AI model and analyzing it in real time.

[0803] "Means for immediately notifying the user based on the anomaly detection results" refers to a means for immediately sending a notification to the user based on an anomaly detected by the generative AI model.

[0804] The system of the present invention aims to efficiently detect corporate window dressing and reduce the burden of accounting audits. This system includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, a notification means for sending notifications to specific devices, a means for collecting and analyzing data from multiple corporate information sources in real time, a means for inputting new standardized data into the generative AI model, a means for immediately notifying users based on the results of anomaly detection, and a means for detecting corporate financial fraud using the prediction results of the generative AI.

[0805] The server collects company financial data, audit reports, news articles, and data on past accounting fraud cases via the Internet and internal systems. This allows data to be collected from companies' published financial statements, reports from third-party auditors, and even public reports. A specific example of data collection would be obtaining Company A's financial data for the past 10 years via an API.

[0806] The server also preprocesses the collected data, removing duplicates and errors, filling in missing data from other data, and standardizing the data format. This results in a clean dataset with consistent numerical data and dates. For example, missing sales data can be filled in using statistical methods from other relevant data.

[0807] The server then uses the preprocessed data to build a machine learning model. The data is divided into a training set, a test set, and a validation set, and the model is trained using techniques such as anomaly detection algorithms, random forests, and deep learning. For example, a random forest model is trained using the training set, and the performance of the constructed model is verified using the test set.

[0808] As a prediction method using generative AI, the user inputs the latest data of the audited company into a terminal and sends the data from the terminal to the server. The server analyzes the latest data using the generative AI model and the trained model and compares it with past data. At this point, if sales have increased significantly compared to the same period last year but the operating profit margin has not changed, it predicts that there is a high possibility of window dressing.

[0809] Finally, the server uses the notification method to notify the user in real time based on the prediction results. The notification details any abnormal patterns and specific suspicious points. For example, the notification might say, "Second-quarter sales increased by 50% compared to the same month last year, but there was no significant change in the operating profit margin, so there is a high possibility of abnormal transactions."

[0810] The hardware used includes a server for collecting and processing data, and a terminal where users can input data and receive notifications. The software includes an API used for data collection and preprocessing, generative AI models and machine learning algorithms used for data analysis, and a communication protocol for notifications. Specifically, Python's pandas and scikit-learn, and deep learning libraries (such as TensorFlow and PyTorch) are used.

[0811] As an example, here is a prompt that uses a generative AI model to analyze new financial data:

[0812] Prompt statement

[0813] Analyze new financial data. This data includes information such as sales, operating profit, net profit, and debt ratio. Also, if sales increase by 50% year-on-year, but the operating profit margin remains unchanged, this may indicate window dressing. Detect and report such abnormal patterns.

[0814] As described above, by implementing the system of the present invention, it is possible to accurately detect corporate window dressing in real time, thereby significantly improving the efficiency of accounting audits.

[0815] The flow of the specific processing in the application example 1 will be described with reference to FIG.

[0816] Step 1:

[0817] The server collects corporate financial data, audit reports, news articles, and data on past accounting fraud cases via the internet and internal systems. Specifically, it obtains this information using APIs. For example, it uses financial data APIs and audit report APIs. The API endpoint and authentication information are required as input, and the collected raw data is obtained as output.

[0818] Step 2:

[0819] The server preprocesses the collected data, removing duplicates and erroneous data and imputing missing data from other data. It also standardizes and normalizes numerical data units and makes date formats consistent. Specifically, it uses the Python pandas library for data cleaning and imputation. The input is the collected raw data, and the output is a preprocessed, clean dataset.

[0820] Step 3:

[0821] The server uses the preprocessed data to build a machine learning model. Specifically, it divides the data into a training set, a test set, and a validation set, and trains it using anomaly detection algorithms, random forests, and deep learning techniques. Libraries used include scikit-learn and TensorFlow. The input is the preprocessed dataset, and the output is a trained machine learning model.

[0822] Step 4:

[0823] The user inputs the latest financial statement data of the company being audited into the terminal. The terminal sends this data to the server in real time. For example, the user inputs financial statements and related data into an input form. The input is the user's latest financial statement data, and the output is the data sent from the terminal to the server.

[0824] Step 5:

[0825] The server analyzes the latest input data using the generative AI model and the trained model. This allows it to compare the latest data with past data and detect abnormal patterns. The prompt text used is something like, "Please analyze the new financial data. This data includes information such as sales, operating profit, net profit, and debt ratio. Furthermore, if sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, this could be window dressing. Please detect and report such abnormal patterns." The input is the latest financial data, and the output is the results of anomaly detection.

[0826] Step 6:

[0827] The server generates a notification containing specific suspicious points based on the prediction results. For example, it generates a notification containing specific content such as, "Sales for the second quarter have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." The input is the anomaly detection result, and the output is the generated notification message.

[0828] Step 7:

[0829] The device displays notifications received from the server to the user in real time. Notifications are displayed as push notifications on smartphones or desktop notifications. Specifically, notifications are received and displayed on the device using push notification APIs or desktop notification APIs. The input is the notification message sent from the server, and the output is the notification that the user can receive and check.

[0830] Furthermore, an emotion engine that estimates the user's emotion may be further combined. That is, the identification processing unit 290 may estimate the user's emotion using the emotion identification model 59, and perform identification processing using the user's emotion.

[0831] System Overview

[0832] The system of the present invention comprises a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, a notification means, and an emotion engine that recognizes user emotions. This system is designed not only to efficiently detect corporate window dressing and reduce the burden of accounting audits, but also to respond according to the user's emotional state.

[0833] Data collection

[0834] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[0835] Data Preprocessing

[0836] The server cleans the collected data. Specifically, it removes duplicates and errors from the data and fills in missing data from other sources. It also standardizes the format of numeric data and dates and normalizes all data. This prepares a clean dataset suitable for analysis.

[0837] Building a learning model

[0838] The server uses the preprocessed data to build a machine learning model. First, it divides the data into a training set, a test set, and a validation set. Next, it uses the training set to train the model by applying techniques such as anomaly detection algorithms, random forests, and deep learning. Finally, it evaluates the model's performance using the test set, checking precision and recall.

[0839] Predictions using generative AI

[0840] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal then sends this data to the server in real time. The server then uses generative AI and a trained model to analyze the latest data and compare it with past data. This allows it to predict whether there is a possibility of window dressing.

[0841] Emotion Engine Operation

[0842] The device's built-in emotion engine analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. For example, it analyzes the user's stress level and sense of security from changes in voice tone and facial expressions. This information is reflected when the prediction results are notified.

[0843] Real-time notifications

[0844] The server generates a notification containing specific suspicious points based on the prediction results. The notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structure. At the same time, the emotion engine determines the user's emotional state and adjusts the tone of the notification and the way detailed information is presented. For example, if the user is feeling stressed, the notification content will be concise and clear.

[0845] Logging

[0846] The server logs the user's emotional state as determined by the emotion engine and stores it in a database for later analysis and improvement. This data helps understand user emotional trends over time and further optimize the system.

[0847] Specific examples

[0848] Examples of data collection

[0849] The server collects Company A's financial statement data, audit reports, and related publicly available news articles for the past 10 years.

[0850] Specific examples of data preprocessing

[0851] The server removes duplicate entries from the collected data, uses statistical methods to fill in missing sales data from other relevant financial data, standardizes all numerical data to a single currency (e.g., USD), and converts dates to "YYYY-MM-DD."

[0852] Specific examples of building learning models

[0853] The server divides the preprocessed data into a training set (70%), a test set (20%), and a validation set (10%). An anomaly detection algorithm is used to extract features, such as sales spikes and cost structure fluctuations. The model is then trained using a random forest algorithm, and its accuracy is verified on the test set.

[0854] Examples of predictions using generative AI

[0855] The user (accountant) inputs the latest financial data for Company B, and the device sends the data to the server. The server analyzes the data using generative AI and detects that sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, predicting a high possibility of window dressing.

[0856] Examples of emotion engines

[0857] The device's emotion engine analyzes the user's (the accountant's) voice and determines that their stress level is high. Based on this information, the server generates short, concise notifications to reduce the user's stress.

[0858] Real-time notification examples

[0859] Based on the forecast results, the server generates a notification with specific suspicious points, such as "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." Because the emotion engine determined the stress level to be high, only the main points are included in the detailed information. The server sends this result to the terminal, which receives it and displays it to the user (accountant). Based on this, the user can immediately start a detailed audit.

[0860] In this way, by implementing the system of the present invention, it is possible to efficiently detect signs of window dressing and take the most appropriate action depending on the user's emotional state.

[0861] The processing flow will be explained below.

[0862] Step 1:

[0863] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[0864] Step 2:

[0865] The server cleans the collected data. Specifically, it removes duplicates and errors from the data. It also supplements missing data from other data and standardizes the format of numerical data and dates. For example, it standardizes data in different currency units and converts them to the "YYYY-MM-DD" format.

[0866] Step 3:

[0867] The server uses the preprocessed data to build a machine learning model by first splitting the data into a training set, a test set, and a validation set, and then using the training set to apply techniques such as anomaly detection algorithms, random forests, or deep learning to train the model.

[0868] Step 4:

[0869] The server evaluates the trained model on the test set to check precision and recall, and retrains the model as needed to optimize performance.

[0870] Step 5:

[0871] The user (accountant) inputs the latest financial statement data of the company being audited into the terminal, such as data on sales, operating profit, and costs.

[0872] Step 6:

[0873] The terminal transmits the input data to the server in real time.

[0874] Step 7:

[0875] The server uses generative AI and trained models to analyze the latest financial data it receives, comparing the input data with past data to detect any abnormal patterns, such as sudden increases in sales or cost inconsistencies.

[0876] Step 8:

[0877] Based on the analysis results, the server predicts the possibility of window dressing, including specific suspicious points such as sudden increases in sales or fluctuations in cost structure.

[0878] Step 9:

[0879] The device's built-in emotion engine analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. For example, it can determine their stress level based on changes in their voice tone and facial expressions.

[0880] Step 10:

[0881] Based on the emotion engine's findings, the server adjusts the content of the notification. For example, if the user is feeling stressed, the notification will be brief and clear.

[0882] Step 11:

[0883] Based on the predictions, the server generates specific suspicious alerts, including details such as unusual sales spikes or fluctuations in cost structures.

[0884] Step 12:

[0885] The server sends this notification to the terminal in real time.

[0886] Step 13:

[0887] The terminal receives the notification and displays it to the user (accountant), who can then immediately begin a detailed audit, for example by gathering additional information on specific items or preparing questions for the company.

[0888] Step 14:

[0889] The server logs the user's emotional state as determined by the emotion engine and stores it in a database for later analysis and improvement, for example, to understand the user's emotional trends over a long period of time and to help further optimize the system.

[0890] Example 2

[0891] Next, a description will be given of Example 2. In the following description, the data processing device 12 will be referred to as a "server" and the headset type terminal 314 will be referred to as a "terminal."

[0892] In audit work, it is important to quickly and accurately detect corporate accounting fraud, but conventional methods require a great deal of time and effort from data collection to analysis. Furthermore, notification methods that do not take into account the stress and emotional state of accountants may reduce the efficiency of audit work. To solve these problems, efficient data processing and prediction methods, as well as appropriate responses according to the user's emotional state, are required.

[0893] The specific processing by the specific processing unit 290 of the data processing device 12 in the second embodiment is realized by the following means.

[0894] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, a notification means, and a log recording means, which enable fast and accurate detection of corporate window dressing and provide an optimal notification method according to the user's emotional state.

[0895] "Data collection methods" are methods for collecting corporate financial information, audit reports, news articles, and information on past accounting fraud cases from online and offline sources.

[0896] "Data preprocessing means" refers to means for cleaning collected information, filling in missing data, and normalizing information.

[0897] A "machine learning model building means" is a means for training and building a machine learning model using preprocessed data.

[0898] "Predictive methods using generative AI" are methods that analyze a company's financial information based on machine learning models and predict the possibility of accounting fraud.

[0899] The "emotion recognition means" is a means for analyzing the user's voice or facial expression and determining the user's emotional state.

[0900] The "notification means" is a means for adjusting and providing notification content according to the prediction result and the emotional state determined by the emotion recognition means.

[0901] The "log recording means" is a means for recording the emotional state of the user determined by the emotion recognition means and storing it in a database.

[0902] The system of the present invention includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, a notification means, and a log recording means. This system is designed to efficiently detect corporate window dressing, reduce the burden of accounting audits, and respond according to the user's emotional state.

[0903] First, the server uses data collection tools to collect company financial information, audit reports, news articles, and information on past accounting fraud cases from the Internet and internal systems. Specifically, it efficiently collects the required information using Python libraries (such as BeautifulSoup and Selenium) and APIs (such as FinAPI). For example, it can collect Company A's financial statement data for the past 10 years, audit reports, and publicly available related news articles.

[0904] Next, the server uses data preprocessing tools to clean the collected data, impute missing data, and normalize the data. Specifically, it removes duplicate data using the drop_duplicates() function in the pandas library, detects missing values ​​using the isnull() function, and imputes them using statistical methods. It also converts numeric data to a unified currency unit (e.g., USD) and standardizes date formats to "YYYY-MM-DD" using the datetime module.

[0905] The server then uses the preprocessed data to build a machine learning model using a machine learning model building method. The data set is split into a training set, a test set, and a validation set, and the train_test_split function in the sklearn library is used. The server also uses Isolation Forest or Random Forest Classifier for feature extraction to train and build the model. For example, an anomaly detection algorithm is used to extract features using indicators such as sudden sales increases or fluctuations in cost structure.

[0906] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal sends the data to the server in real time. The server analyzes the input data using a predictive method that uses generative AI, and compares it with past data to predict the possibility of window dressing. The generative AI uses a model such as OpenAI's GPT-4. For example, a prompt statement could be entered as follows: "Enter the latest financial data of Company B, have the server analyze it, and detect that sales have increased by 50% compared to the same period last year, but the operating profit margin has remained unchanged, and predict that there is a high possibility of window dressing."

[0907] Furthermore, the device's built-in emotion recognition means analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. It uses the librosa library for voice analysis and the OpenCV and dlib libraries for facial expression recognition. For example, it detects changes in voice tone or abnormal facial expressions to determine whether the user is in a high-stress state. This information is reflected in the notification content sent by the prediction result notification means.

[0908] The server uses a real-time notification mechanism to generate a notification containing specific suspicious points based on the prediction results. This notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structure. The notification content is adjusted to be concise and clear, taking into account the user's emotional information determined by the emotion recognition mechanism. For example, a notification may be generated that states, "Sales in the second quarter have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." This notification is sent to the terminal and displayed to the user (accountant).

[0909] Finally, the server uses the logging means to log the user's emotional state as determined by the emotion recognition means and stores the log in a database. This data is used to understand the user's emotional trends over a long period of time and to further optimize the system. For example, if a user is in a state of high stress under certain conditions, this data can be used as feedback to revise the system's interface and notification methods.

[0910] The above configuration and processing make it possible to quickly and accurately detect corporate window dressing and respond optimally according to the user's emotional state.

[0911] The flow of the identification process in the second embodiment will be described with reference to FIG.

[0912] Processing flow

[0913] Step 1: Collect data

[0914] The server collects corporate financial information, audit reports, news articles, and information on past accounting fraud cases via the internet and internal systems. It efficiently collects the necessary data using libraries such as Python's BeautifulSoup and Selenium, and APIs such as FinAPI. Specifically, it retrieves data from the API using the requests library, and parses HTML with BeautifulSoup to extract financial information.

[0915] Input: URL or API endpoint to collect data

[0916] Output: Dataset of company financial information, audit reports, press articles, and information on past accounting fraud cases

[0917] Step 2: Preprocessing the data

[0918] The server cleans the collected data, imputes missing data, and normalizes the data. Specifically, it uses the pandas library to remove duplicate data, the isnull() function to detect missing values, and statistical methods to impute them. It also standardizes date formats using the datetime module and converts numeric data to a uniform currency unit.

[0919] Input: Dataset of company financial information, audit reports, news articles, and information on past accounting fraud cases

[0920] Output: A clean and normalized dataset

[0921] Step 3: Building a learning model

[0922] The server builds a machine learning model using the preprocessed data. It splits the dataset into training, test, and validation sets using the train_test_split function in the sklearn library, and uses IsolationForest and RandomForestClassifier for feature extraction. After training the model, it evaluates its accuracy on the test set.

[0923] Input: Clean and normalized dataset

[0924] Output: A trained machine learning model

[0925] Step 4: Predict using generative AI

[0926] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal sends the data in real time to the server. The server uses generative AI and a trained model to analyze the input data and predict the possibility of window dressing. The generative AI uses a model such as GPT-4.

[0927] Input: Latest financial data

[0928] Output: Prediction result regarding the possibility of window dressing

[0929] Step 5: Emotion Engine in Action

[0930] The emotion engine installed on the device analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. Changes in voice tone are analyzed using librosa, and facial expressions are evaluated using OpenCV and dlib. This determines stress levels and emotional states.

[0931] Input: User's voice data, facial expression data

[0932] Output: User's emotional state

[0933] Step 6: Real-time notifications

[0934] The server generates a notification based on the prediction results, including specific suspicious points, such as sudden sales increases or fluctuations in cost structure, and adjusts the content and tone of the notification based on the emotional information determined by the emotion recognition means.

[0935] Input: Prediction result, user's emotional state

[0936] Output: Adjusted notification content

[0937] Step 7: Logging

[0938] The server records the user's emotional state as determined by the emotion engine as a log and stores it in a database. The log data is used to improve the system and optimize user responses.

[0939] Input: User's emotional state

[0940] Output: Saved log data

[0941] (Application example 2)

[0942] Next, a description will be given of Application Example 2. In the following description, the data processing device 12 will be referred to as a "server" and the headset type terminal 314 will be referred to as a "terminal."

[0943] In modern business activities, window dressing remains a serious problem, placing a heavy burden on accountants and auditors. Audit work also creates significant stress and strain. To address this issue, a system is needed that efficiently detects abnormal transactions and provides appropriate notifications based on the accountant's emotional state. The present invention aims to improve corporate reliability and reduce the burden of accounting audits by solving these problems.

[0944] The specific processing by the specific processing unit 290 of the data processing device 12 in the application example 2 is realized by the following means.

[0945] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, and a real-time notification means, which enable efficient detection of anomalies in a company's financial data and appropriate notification according to the user's emotional state.

[0946] "Data collection methods" are methods for collecting financial data, audit reports, information articles, and historical anomaly data from online or offline sources.

[0947] "Data preprocessing means" refers to methods for cleaning collected data, filling in missing data, and normalizing the data into a format that is easy to analyze.

[0948] "Machine learning model building means" refers to a method of using pre-processed data to apply anomaly detection algorithms and other machine learning techniques to build a model for detecting anomalous transactions.

[0949] "Prediction methods using generative AI" are methods that use trained machine learning models and generative AI to analyze newly obtained data and predict the possibility of abnormal transactions.

[0950] "Emotion recognition means" is a method for determining the user's emotional state by analyzing the user's voice and facial expressions, and adjusting the system's response based on that information.

[0951] "Real-time notification means" is a method of generating a notification based on the results of abnormal transaction detection, with content that corresponds to the user's emotional state, and quickly notifying the user.

[0952] The present invention is a system for monitoring and detecting anomalies in corporate financial data, and is intended to be used on end-user devices such as smartphone applications. Specific procedures for implementing the invention are described below.

[0953] Hardware and software used

[0954] Hardware: Smartphones, cloud servers

[0955] Software: Python, Scikit-learn, Keras, EmotionEngine, NotificationService

[0956] System configuration

[0957] 1. Data collection methods:

[0958] The server collects financial data, audit reports, news articles, and past anomaly data from online and offline sources. For example, financial data is collected through APIs, and audit reports and news articles are collected using a web crawling system.

[0959] 2. Data preprocessing methods:

[0960] Once collected, the data is filled in with missing values, errors corrected, and formatted in a consistent manner (for example, missing sales data is filled in from historical data, and date formats are standardized).

[0961] 3. Machine learning model building methods:

[0962] The preprocessed data is used to apply an anomaly detection algorithm and build a model. Specifically, the data is divided into a training set and a test set, and training is performed using a random forest or deep learning model.

[0963] 4. Generative AI prediction methods:

[0964] When a user inputs their latest financial data, the data is sent to the server, where generative AI and trained models are used to predict anomalies. For example, if sales have increased significantly but the operating profit margin remains unchanged, the system predicts possible anomalies.

[0965] 5. Emotion recognition means:

[0966] Analyzes the user's voice and facial expressions to determine their emotional state. Emotion Engine is used to analyze voice tone and facial expression changes to determine the user's stress level.

[0967] 6. Real-time notification methods:

[0968] Based on the anomaly detection results, notifications are generated with content that corresponds to the user's emotional state. For example, if the user is particularly stressed, the notification content will be brief and only the main points will be presented.

[0969] Specific examples of realization

[0970] 1. Examples of data collection:

[0971] Use APIs to retrieve company financial data and collect audit reports for the past 10 years.

[0972] 2. Examples of data preprocessing:

[0973] Remove duplicate entries from collected data, impute missing data using statistical methods, convert all numerical data to a single currency, and standardize date formats.

[0974] 3. Specific examples of building learning models:

[0975] The preprocessed data is used to train a model using the random forest algorithm, and the performance of the anomaly detection model is verified.

[0976] 4. Specific examples of predictions using generative AI:

[0977] Users enter their company's latest financial data, which is then analyzed on the server, and if there is a high possibility of an abnormality, an alert is issued immediately.

[0978] 5. Examples of emotion recognition:

[0979] The system analyzes the user's voice and, if it determines that the stress level is high, simplifies the notification content.

[0980] 6. Examples of real-time notifications:

[0981] "Second-quarter sales have surged 50% compared to the same month last year, but there has been no significant change in operating profit margins, so there is a high possibility that this is an abnormal transaction," the notice said.

[0982] Prompt Sentence Examples

[0983] "Evaluate a company's latest financial data and detect anomalies. Write a program that adjusts the tone of notifications based on the user's stress level."

[0984] The flow of the specific processing in the application example 2 will be described with reference to FIG.

[0985] Step 1:

[0986] Data collection by data collection method

[0987] The server collects financial data, audit reports, information articles, and historical anomaly data from online and offline sources. Specifically, it retrieves financial data using APIs and collects news articles using a web crawling system. It uses the URL or API endpoint of each source as input and obtains the collected raw data as output.

[0988] Step 2:

[0989] Data cleaning through data preprocessing methods

[0990] The server cleans the collected raw data, imputes missing data, and normalizes the data into a format that is easy to analyze. Specifically, it removes duplicate data, imputes missing data (e.g., imputation using statistical methods), and standardizes date formats. It uses the collected raw data as input and obtains a cleaned dataset as output.

[0991] Step 3:

[0992] Building models using machine learning model building methods

[0993] The server uses the preprocessed data to apply an anomaly detection algorithm and build a model. Specifically, it splits the data into a training set and a test set, and trains the model using a random forest or deep learning algorithm. The preprocessed dataset is used as input, and the trained model is obtained as output.

[0994] Step 4:

[0995] Predicting anomalies using generative AI-based predictive methods

[0996] When a user inputs their latest financial data, the device sends the data to the server. The server then uses generative AI and trained models to analyze this data and predict anomalies. Specifically, it detects sudden increases in sales or inconsistencies in cost structures. The latest financial data provided by the user is used as input, and a prediction result indicating whether or not anomalies exist is obtained as output.

[0997] Step 5:

[0998] Emotional state analysis using emotion recognition tools

[0999] The device analyzes the user's voice and facial expressions in real time to determine their emotional state. Specifically, it uses the Emotion Engine to analyze changes in voice tone and facial expressions to determine stress levels and feelings of relief. The device uses the user's voice and facial expression data as input and outputs the determined emotional state.

[1000] Step 6:

[1001] Generate and send notifications via real-time notification methods

[1002] The server generates a notification based on the anomaly detection results, with content appropriate to the user's emotional state. Specifically, if an anomaly is detected, a detailed explanation is provided, and if the user is in a stressful state, a brief notification is provided. The anomaly detection results and the user's emotional state are used as inputs, and the generated notification message is obtained as output.

[1003] The specific processing unit 290 transmits the result of the specific processing to the headset type terminal 314. In the headset type terminal 314, the control unit 46A causes the speaker 240 and the display 343 to output the result of the specific processing. The microphone 238 acquires audio indicating a user input regarding the result of the specific processing. The control unit 46A transmits audio data indicating the user input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the audio data.

[1004] The data generation model 58 is a so-called generative AI (Artificial Intelligence). An example of the data generation model 58 is ChatGPT (Internet Search<URL: https: / / openai.com / blog / chatgpt> ), Gemini (Internet search <url: https: gemini.google.com ?hl="ja">) and other generation AIs. The data generation model 58 is obtained by performing deep learning on a neural network. A prompt including an instruction is input to the data generation model 58, and inference data such as voice data indicating voice, text data indicating text, and image data indicating an image is also input. The data generation model 58 performs inference on the input inference data in accordance with the instruction indicated by the prompt, and outputs the inference result in a data format such as voice data and text data. Here, inference refers to, for example, analysis, classification, prediction, and / or summarization.

[1005] In the above embodiment, an example was given in which the specific processing is performed by the data processing device 12, but the technology of the present disclosure is not limited to this, and the specific processing may be performed by the headset type terminal 314.

[1006] [Fourth embodiment]

[1007] FIG. 7 shows an example of the configuration of a data processing system 410 according to the fourth embodiment.

[1008] 7, a data processing system 410 includes a data processing device 12 and a robot 414. An example of the data processing device 12 is a server.

[1009] The data processing device 12 includes a computer 22, a database 24, and a communication I / F 26. The computer 22 is an example of a "computer" according to the technology of the present disclosure. The computer 22 includes a processor 28, a RAM 30, and a storage 32. The processor 28, the RAM 30, and the storage 32 are connected to a bus 34. The database 24 and the communication I / F 26 are also connected to the bus 34. The communication I / F 26 is connected to a network 54. Examples of the network 54 include a WAN (Wide Area Network) and / or a LAN (Local Area Network).

[1010] The robot 414 includes a computer 36, a microphone 238, a speaker 240, a camera 42, a communication I / F 44, and a control target 443. The computer 36 includes a processor 46, a RAM 48, and a storage 50. The processor 46, the RAM 48, and the storage 50 are connected to a bus 52. The microphone 238, the speaker 240, the camera 42, and the control target 443 are also connected to the bus 52.

[1011] The microphone 238 receives instructions and the like from the user 20 by receiving voice uttered by the user 20. The microphone 238 captures the voice uttered by the user 20, converts the captured voice into audio data, and outputs it to the processor 46. The speaker 240 outputs audio in accordance with instructions from the processor 46.

[1012] Camera 42 is a small digital camera equipped with an optical system including a lens, aperture, and shutter, and an imaging element such as a CMOS (Complementary Metal-Oxide-Semiconductor) image sensor or a CCD (Charge Coupled Device) image sensor, and captures images of the surroundings of user 20 (for example, an imaging range defined by an angle of view equivalent to the field of vision of a typical healthy person).

[1013] The communication I / F 44 is connected to a network 54. The communication I / Fs 44 and 26 are responsible for the exchange of various information between the processor 46 and the processor 28 via the network 54. The exchange of various information between the processor 46 and the processor 28 using the communication I / Fs 44 and 26 is carried out in a secure state.

[1014] The control object 443 includes a display device, LEDs in the eyes, and motors for driving the arms, hands, and feet. The posture and gestures of the robot 414 are controlled by controlling the motors of the arms, hands, and feet. Some of the emotions of the robot 414 can be expressed by controlling these motors. In addition, the facial expressions of the robot 414 can also be expressed by controlling the light emission state of the LEDs in the eyes of the robot 414.

[1015] Fig. 8 shows an example of the main functions of the data processing device 12 and the robot 414. As shown in Fig. 8, in the data processing device 12, a specific process is performed by the processor 28. A specific process program 56 is stored in the storage 32.

[1016] The specific processing program 56 is an example of a "program" according to the technology of the present disclosure. The processor 28 reads the specific processing program 56 from the storage 32 and executes the read specific processing program 56 on the RAM 30. The specific processing is realized by the processor 28 operating as a specific processing unit 290 in accordance with the specific processing program 56 executed on the RAM 30.

[1017] The storage 32 stores a data generation model 58 and an emotion identification model 59. The data generation model 58 and the emotion identification model 59 are used by the identification processing unit 290.

[1018] In the robot 414, the processor 46 performs the reception output process. A reception output program 60 is stored in the storage 50. The processor 46 reads the reception output program 60 from the storage 50 and executes the read reception output program 60 on the RAM 48. The reception output process is realized by the processor 46 operating as the control unit 46A in accordance with the reception output program 60 executed on the RAM 48.

[1019] Next, a description will be given of the specific processing performed by the specific processing unit 290 of the data processing device 12. In the following description, the data processing device 12 will be referred to as a "server" and the robot 414 will be referred to as a "terminal."

[1020] System Overview

[1021] The system of the present invention includes data collection, data preprocessing, machine learning model construction, prediction using generative AI, and notification methods. This system is designed to efficiently detect corporate window dressing and reduce the burden of accounting audits.

[1022] Data collection

[1023] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[1024] Data Preprocessing

[1025] The server preprocesses the collected data. Specifically, it removes duplicates and errors from the data and fills in missing data from other sources. It also standardizes the format of numeric data and dates and normalizes all data. This prepares a clean dataset suitable for analysis.

[1026] Building a learning model

[1027] The server uses the preprocessed data to build a machine learning model. First, it divides the data into a training set, a test set, and a validation set. Next, it uses the training set to train the model by applying techniques such as anomaly detection algorithms, random forests, and deep learning. Finally, it evaluates the model's performance using the test set, checking precision and recall.

[1028] Predictions using generative AI

[1029] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal then sends this data to the server in real time. The server then uses generative AI and a trained model to analyze the latest data and compare it with past data. This allows it to predict whether there is a possibility of window dressing.

[1030] Real-time notifications

[1031] Based on the prediction results, the server generates a notification containing specific suspicious points. This notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structures. The server sends this notification in real time to the terminal, which receives the notification and displays it to the user (accountant). This allows the user to immediately identify suspicious transactions or unusual financial data and begin detailed audit procedures.

[1032] Specific examples

[1033] Examples of data collection

[1034] The server collects Company A's financial statement data, audit reports, and related publicly available news articles for the past 10 years.

[1035] Specific examples of data preprocessing

[1036] The server removes duplicate entries from the collected data, uses statistical methods to fill in missing sales data from other relevant financial data, standardizes all numerical data to a single currency (e.g., USD), and converts dates to "YYYY-MM-DD."

[1037] Specific examples of building learning models

[1038] The server divides the preprocessed data into a training set (70%), a test set (20%), and a validation set (10%). An anomaly detection algorithm is used to extract features, such as sales spikes and cost structure fluctuations. The model is then trained using a random forest algorithm, and its accuracy is verified on the test set.

[1039] Examples of predictions using generative AI

[1040] The user (accountant) inputs the latest financial data for Company B, and the device sends the data to the server. The server analyzes the data using generative AI and detects that sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, predicting a high possibility of window dressing.

[1041] Real-time notification examples

[1042] Based on the forecast results, the server generates a notification stating specific suspicious points, such as "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." The server sends this result to the terminal, which receives it and displays it to the user (accountant). Based on this, the user can immediately start a detailed audit.

[1043] In this way, by implementing the system of the present invention, it is possible to efficiently detect signs of window dressing and respond quickly.

[1044] The processing flow will be explained below.

[1045] Step 1:

[1046] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from the internet and internal systems. This data comes from a variety of sources, including publicly available company financial statements, third-party auditor reports, and even public reports.

[1047] Step 2:

[1048] The server cleans the collected data, removing duplicates and errors, filling in missing data from other data, and standardizing the format of numeric data and dates.

[1049] Step 3:

[1050] The server uses the preprocessed data to build machine learning models by splitting the data into training, test, and validation sets and applying techniques such as anomaly detection algorithms, random forests, and deep learning to train the models.

[1051] Step 4:

[1052] The server evaluates the trained model on the test set to check precision and recall, and retrains the model as needed to optimize performance.

[1053] Step 5:

[1054] The user (accountant) inputs the latest financial statement data of the company being audited into the terminal.

[1055] Step 6:

[1056] The terminal transmits the input data to the server in real time.

[1057] Step 7:

[1058] The server uses generative AI and trained models to analyze the latest financial data it receives, comparing it with historical data to determine whether any unusual patterns exist.

[1059] Step 8:

[1060] As a result of the analysis, the server predicts the possibility of accounting fraud, including specific suspicious points such as sudden increases in sales or fluctuations in cost structure.

[1061] Step 9:

[1062] The server generates a notification containing specific suspicious points based on the prediction results.

[1063] Step 10:

[1064] The server sends this notification to the terminal in real time.

[1065] Step 11:

[1066] The terminal receives the notification and displays it to the user (accountant), who can then immediately initiate a detailed audit.

[1067] Example 1

[1068] Next, a description will be given of Example 1. In the following description, the data processing device 12 will be referred to as a "server" and the robot 414 will be referred to as a "terminal."

[1069] The challenge is to efficiently and quickly detect accounting fraud in corporate financial data and reduce the burden of accounting audits. Conventional methods require manual analysis of large amounts of data, which is prone to errors and is time-consuming and costly. In addition, advanced algorithms to improve the accuracy of anomaly detection have not been fully utilized.

[1070] The specific processing by the specific processing unit 290 of the data processing device 12 in the first embodiment is realized by the following means.

[1071] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using a generative AI, and a real-time notification means. This enables comprehensive collection and preprocessing of corporate financial data, audit reports, news articles, data on past window dressing cases, etc., and construction of a machine learning model for highly accurate anomaly detection. Furthermore, the use of generative AI enables real-time analysis and prediction based on the latest data entered by the user, and immediate notification can be sent if an anomaly is detected, enabling a rapid response.

[1072] "Data collection means" refers to the ability to collect corporate financial data, audit reports, news articles, and past accounting fraud cases from online and offline sources.

[1073] "Data preprocessing means" refers to functions for cleansing collected data, filling in missing data, and normalizing data.

[1074] "Machine learning model construction means" refers to the function of using preprocessed data to divide the data into a training set, a test set, and a validation set, constructing a model using an anomaly detection algorithm, random forest, or deep learning, and evaluating the accuracy and recall.

[1075] "Prediction method using generative AI" refers to a function that receives the latest financial results data of a company entered by the user, analyzes the data using a learning model and generative AI, detects anomalies, and predicts the possibility of window dressing.

[1076] The "real-time notification means" refers to a function that generates a notification containing suspicious points based on the prediction results, and transmits and displays the notification to the user terminal.

[1077] The system of the present invention is composed of a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, and a real-time notification means. This system is designed to efficiently detect corporate window dressing and reduce the burden of accounting audits.

[1078] Data collection

[1079] The server collects company financial data, audit reports, news articles, and data on past accounting fraud cases via the internet and internal systems. To collect data, it uses Python's BeautifulSoup library and Selenium to perform web scraping. It also uses an ERP system to extract data from the company's internal systems.

[1080] Example: A server sends an HTTP request to retrieve Company A's financial statements from a financial information website, extracts the necessary data from the HTML using BeautifulSoup, and retrieves financial data from the company's ERP system via an API.

[1081] Data Preprocessing

[1082] The server uses data processing libraries such as Pandas and NumPy to preprocess the collected data, specifically deleting duplicate data, correcting erroneous data, filling in missing data, standardizing the currency units of numeric data, and standardizing date formats.

[1083] Example: The server removes duplicate entries from the collected dataset using Pandas' drop_duplicates() method, fills missing data with the average value using the fillna() method, converts all currency units to USD, and standardizes the date format to "YYYY-MM-DD".

[1084] Building a learning model

[1085] The server builds a machine learning model using the preprocessed data. It divides the data into a training set, a test set, and a validation set, and trains the model using an anomaly detection algorithm (e.g., Isolation Forest), a random forest, or a deep learning framework (e.g., TensorFlow). It evaluates the precision and recall of the models and selects the best model.

[1086] Example: The server splits the preprocessed data using scikit-learn's train_test_split function, trains the model using the Isolation Forest algorithm, runs the training using the fit() method, and evaluates the test data using the predict() method.

[1087] Predictions using generative AI

[1088] A user (accountant) inputs the company's latest financial data into a terminal, which then sends the data to a server. The server uses a generative AI model (e.g., GPT-4) and a trained model to analyze the data, detect anomalies, and predict possible window dressing.

[1089] Example: A user enters financial data for Company B into a form on a terminal and presses the "Submit" button. The terminal sends the data in JSON format to the server as a POST request. The server inputs the received data into the GPT-4 generation AI, which generates a prompt such as "Please analyze the sudden increase in sales and fluctuations in operating profit margins for Company B."

[1090] Real-time notifications

[1091] The server generates a notification containing suspicious points based on the prediction results and sends it to the user's terminal. The user can receive this notification and perform a detailed audit based on the displayed abnormal points.

[1092] Example: Based on the prediction results, the server generates a notification such as, "Company B's second-quarter sales have increased by 50% compared to the same month last year, but there is no significant change in its operating profit margin, so there is a high possibility that this is an abnormal transaction," and sends it to the terminal. The terminal receives this notification and displays it to the user as a pop-up or email notification.

[1093] In this way, this system can efficiently detect corporate window dressing and enable rapid response.

[1094] The flow of the identification process in the first embodiment will be described with reference to FIG.

[1095] Step 1:

[1096] The server uses data collection tools to collect data from the internet and internal company systems. Specifically, it uses web scraping tools such as BeautifulSoup and Selenium to obtain company financial data, audit reports, news articles, and data on past accounting fraud cases. It also collects internal data from the company's ERP system via APIs.

[1097] Input: A list of URLs or API endpoints for data collection.

[1098] Output: HTML page and / or JSON formatted data.

[1099] What it does: The server crawls financial information website URLs and sends HTTP requests. It then parses the HTML content with BeautifulSoup to extract financial data. It also uses the API of the company's ERP system to obtain an authentication token and collect the necessary financial data in JSON format.

[1100] Step 2:

[1101] The server preprocesses the collected data using data preprocessing tools, such as data cleansing, removing duplicates, filling in missing data, standardizing numeric data, and standardizing date formats. This is done using data processing libraries such as Pandas and NumPy.

[1102] Input: Raw data such as collected financial data and audit reports.

[1103] Output: A preprocessed clean dataset.

[1104] Specific operation: The server removes duplicate data using Pandas' drop_duplicates() method, fills missing values ​​with the mean using the fillna() method, converts numeric data to a unified currency (e.g., USD), and standardizes the date format to "YYYY-MM-DD".

[1105] Step 3:

[1106] The server uses a machine learning model building tool to build a machine learning model using the preprocessed data. It splits the data into a training set, a test set, and a validation set, and trains the model using an anomaly detection algorithm, random forest, or a deep learning framework (e.g., TensorFlow).

[1107] Input: Preprocessed clean dataset.

[1108] Output: A trained machine learning model.

[1109] Specific operation: The server uses scikit-learn's train_test_split function to split the data into a 70% training set, a 20% test set, and a 10% validation set. It trains the model using the Isolation Forest algorithm or Random Forest, and builds the model using the fit() method. It also evaluates the test data using the predict() method to check precision and recall.

[1110] Step 4:

[1111] A user (accountant) inputs the company's latest financial data into a terminal, which then sends the data to a server. The server uses a generative AI model (e.g., GPT-4) and a trained model to analyze the data, detect anomalies, and predict possible window dressing.

[1112] Input: The most recent financial data entered by the user.

[1113] Output: Anomaly detection results and predicted window dressing probability.

[1114] Specific operation: The user enters the company's financial data into a form on the device and presses the "Submit" button. The device sends this data in JSON format to the server as a POST request. The server inputs the received data into the GPT-4 generation AI, which generates and analyzes the prompt: "Please analyze the sudden increase in sales and fluctuations in operating profit margins for Company B."

[1115] Step 5:

[1116] The server generates a notification containing suspicious points based on the prediction results and sends it to the user terminal in real time. The user can receive this notification and perform a detailed audit based on the displayed abnormalities.

[1117] Input: Anomaly detection results and predicted probability of window dressing.

[1118] Output: Informational message.

[1119] Specific operation: Based on the prediction results, the server generates a notification such as, "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions," and sends it to the terminal in JSON format. The terminal receives this notification and displays it to the user as a pop-up or email notification.

[1120] (Application example 1)

[1121] Next, a description will be given of Application Example 1. In the following description, the data processing device 12 will be referred to as a "server" and the robot 414 will be referred to as a "terminal."

[1122] Currently, detecting corporate accounting fraud requires a great deal of effort and time, making it difficult to detect in real time. Furthermore, it is difficult to ensure the quality and consistency of collected data, which increases the risk of false positives and oversights. For this reason, there is a need to quickly and accurately detect corporate financial fraud and improve the efficiency of accounting audits.

[1123] The specific processing by the specific processing unit 290 of the data processing device 12 in the application example 1 is realized by the following means.

[1124] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using the generative AI, a notification means for sending a notification to a specific device, a means for collecting and analyzing data from multiple information sources of the company in real time, a means for inputting new standardized data into the generative AI model, a means for immediately notifying a user based on the anomaly detection result, and a means for detecting financial fraud at the company using the prediction result of the generative AI. This makes it possible to accurately detect financial fraud at the company in real time and significantly improve the efficiency of accounting audits.

[1125] "Data collection means" refers to means for collecting corporate financial data, audit reports, news articles, and past accounting fraud cases from online and offline sources.

[1126] "Data preprocessing means" refers to means for cleaning collected data, imputing missing data, normalizing data, and standardizing data for input into generative AI models.

[1127] A "machine learning model construction means" is a means for constructing a machine learning model using preprocessed data and using that model to perform anomaly detection and prediction.

[1128] A "predictive method using generative AI" is a method of predicting corporate financial fraud by inputting new standardized data into a generative AI model.

[1129] "Notification means" refers to a means for immediately notifying users of normal or emergency conditions based on the results of anomaly detection.

[1130] "Means for collecting and analyzing data from multiple sources of information for a company in real time" refers to means for continuously collecting and analyzing a company's financial data, audit reports, news articles, and data on past accounting fraud cases in real time.

[1131] "Notification means for sending a notification to a specific device" refers to a means for sending a notification to a specific device such as a smartphone or desktop when financial fraud at a company is detected.

[1132] "Means for inputting new standardized data into the generative AI model" refers to means for inputting newly collected data into the generative AI model and analyzing it in real time.

[1133] "Means for immediately notifying the user based on the anomaly detection results" refers to a means for immediately sending a notification to the user based on an anomaly detected by the generative AI model.

[1134] The system of the present invention aims to efficiently detect corporate window dressing and reduce the burden of accounting audits. This system includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, a notification means for sending notifications to specific devices, a means for collecting and analyzing data from multiple corporate information sources in real time, a means for inputting new standardized data into the generative AI model, a means for immediately notifying users based on the results of anomaly detection, and a means for detecting corporate financial fraud using the prediction results of the generative AI.

[1135] The server collects company financial data, audit reports, news articles, and data on past accounting fraud cases via the Internet and internal systems. This allows data to be collected from companies' published financial statements, reports from third-party auditors, and even public reports. A specific example of data collection would be obtaining Company A's financial data for the past 10 years via an API.

[1136] The server also preprocesses the collected data, removing duplicates and errors, filling in missing data from other data, and standardizing the data format. This results in a clean dataset with consistent numerical data and dates. For example, missing sales data can be filled in using statistical methods from other relevant data.

[1137] The server then uses the preprocessed data to build a machine learning model. The data is divided into a training set, a test set, and a validation set, and the model is trained using techniques such as anomaly detection algorithms, random forests, and deep learning. For example, a random forest model is trained using the training set, and the performance of the constructed model is verified using the test set.

[1138] As a prediction method using generative AI, the user inputs the latest data of the audited company into a terminal and sends the data from the terminal to the server. The server analyzes the latest data using the generative AI model and the trained model and compares it with past data. At this point, if sales have increased significantly compared to the same period last year but the operating profit margin has not changed, it predicts that there is a high possibility of window dressing.

[1139] Finally, the server uses the notification method to notify the user in real time based on the prediction results. The notification details any abnormal patterns and specific suspicious points. For example, the notification might say, "Second-quarter sales increased by 50% compared to the same month last year, but there was no significant change in the operating profit margin, so there is a high possibility of abnormal transactions."

[1140] The hardware used includes a server for collecting and processing data, and a terminal where users can input data and receive notifications. The software includes an API used for data collection and preprocessing, generative AI models and machine learning algorithms used for data analysis, and a communication protocol for notifications. Specifically, Python's pandas and scikit-learn, and deep learning libraries (such as TensorFlow and PyTorch) are used.

[1141] As an example, here is a prompt that uses a generative AI model to analyze new financial data:

[1142] Prompt statement

[1143] Analyze new financial data. This data includes information such as sales, operating profit, net profit, and debt ratio. Also, if sales increase by 50% year-on-year, but the operating profit margin remains unchanged, this may indicate window dressing. Detect and report such abnormal patterns.

[1144] As described above, by implementing the system of the present invention, it is possible to accurately detect corporate window dressing in real time, thereby significantly improving the efficiency of accounting audits.

[1145] The flow of the specific processing in the application example 1 will be described with reference to FIG.

[1146] Step 1:

[1147] The server collects corporate financial data, audit reports, news articles, and data on past accounting fraud cases via the internet and internal systems. Specifically, it obtains this information using APIs. For example, it uses financial data APIs and audit report APIs. The API endpoint and authentication information are required as input, and the collected raw data is obtained as output.

[1148] Step 2:

[1149] The server preprocesses the collected data, removing duplicates and erroneous data and imputing missing data from other data. It also standardizes and normalizes numerical data units and makes date formats consistent. Specifically, it uses the Python pandas library for data cleaning and imputation. The input is the collected raw data, and the output is a preprocessed, clean dataset.

[1150] Step 3:

[1151] The server uses the preprocessed data to build a machine learning model. Specifically, it divides the data into a training set, a test set, and a validation set, and trains it using anomaly detection algorithms, random forests, and deep learning techniques. Libraries used include scikit-learn and TensorFlow. The input is the preprocessed dataset, and the output is a trained machine learning model.

[1152] Step 4:

[1153] The user inputs the latest financial statement data of the company being audited into the terminal. The terminal sends this data to the server in real time. For example, the user inputs financial statements and related data into an input form. The input is the user's latest financial statement data, and the output is the data sent from the terminal to the server.

[1154] Step 5:

[1155] The server analyzes the latest input data using the generative AI model and the trained model. This allows it to compare the latest data with past data and detect abnormal patterns. The prompt text used is something like, "Please analyze the new financial data. This data includes information such as sales, operating profit, net profit, and debt ratio. Furthermore, if sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, this could be window dressing. Please detect and report such abnormal patterns." The input is the latest financial data, and the output is the results of anomaly detection.

[1156] Step 6:

[1157] The server generates a notification containing specific suspicious points based on the prediction results. For example, it generates a notification containing specific content such as, "Sales for the second quarter have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." The input is the anomaly detection result, and the output is the generated notification message.

[1158] Step 7:

[1159] The device displays notifications received from the server to the user in real time. Notifications are displayed as push notifications on smartphones or desktop notifications. Specifically, notifications are received and displayed on the device using push notification APIs or desktop notification APIs. The input is the notification message sent from the server, and the output is the notification that the user can receive and check.

[1160] Furthermore, an emotion engine that estimates the user's emotion may be further combined. That is, the identification processing unit 290 may estimate the user's emotion using the emotion identification model 59, and perform identification processing using the user's emotion.

[1161] System Overview

[1162] The system of the present invention comprises a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, a notification means, and an emotion engine that recognizes user emotions. This system is designed not only to efficiently detect corporate window dressing and reduce the burden of accounting audits, but also to respond according to the user's emotional state.

[1163] Data collection

[1164] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[1165] Data Preprocessing

[1166] The server cleans the collected data. Specifically, it removes duplicates and errors from the data and fills in missing data from other sources. It also standardizes the format of numeric data and dates and normalizes all data. This prepares a clean dataset suitable for analysis.

[1167] Building a learning model

[1168] The server uses the preprocessed data to build a machine learning model. First, it divides the data into a training set, a test set, and a validation set. Next, it uses the training set to train the model by applying techniques such as anomaly detection algorithms, random forests, and deep learning. Finally, it evaluates the model's performance using the test set, checking precision and recall.

[1169] Predictions using generative AI

[1170] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal then sends this data to the server in real time. The server then uses generative AI and a trained model to analyze the latest data and compare it with past data. This allows it to predict whether there is a possibility of window dressing.

[1171] Emotion Engine Operation

[1172] The device's built-in emotion engine analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. For example, it analyzes the user's stress level and sense of security from changes in voice tone and facial expressions. This information is reflected when the prediction results are notified.

[1173] Real-time notifications

[1174] The server generates a notification containing specific suspicious points based on the prediction results. The notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structure. At the same time, the emotion engine determines the user's emotional state and adjusts the tone of the notification and the way detailed information is presented. For example, if the user is feeling stressed, the notification content will be concise and clear.

[1175] Logging

[1176] The server logs the user's emotional state as determined by the emotion engine and stores it in a database for later analysis and improvement. This data helps understand user emotional trends over time and further optimize the system.

[1177] Specific examples

[1178] Examples of data collection

[1179] The server collects Company A's financial statement data, audit reports, and related publicly available news articles for the past 10 years.

[1180] Specific examples of data preprocessing

[1181] The server removes duplicate entries from the collected data, uses statistical methods to fill in missing sales data from other relevant financial data, standardizes all numerical data to a single currency (e.g., USD), and converts dates to "YYYY-MM-DD."

[1182] Specific examples of building learning models

[1183] The server divides the preprocessed data into a training set (70%), a test set (20%), and a validation set (10%). An anomaly detection algorithm is used to extract features, such as sales spikes and cost structure fluctuations. The model is then trained using a random forest algorithm, and its accuracy is verified on the test set.

[1184] Examples of predictions using generative AI

[1185] The user (accountant) inputs the latest financial data for Company B, and the device sends the data to the server. The server analyzes the data using generative AI and detects that sales have increased by 50% compared to the same period last year, but the operating profit margin has not changed, predicting a high possibility of window dressing.

[1186] Examples of emotion engines

[1187] The device's emotion engine analyzes the user's (the accountant's) voice and determines that their stress level is high. Based on this information, the server generates short, concise notifications to reduce the user's stress.

[1188] Real-time notification examples

[1189] Based on the forecast results, the server generates a notification with specific suspicious points, such as "Second-quarter sales have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." Because the emotion engine determined the stress level to be high, only the main points are included in the detailed information. The server sends this result to the terminal, which receives it and displays it to the user (accountant). Based on this, the user can immediately start a detailed audit.

[1190] In this way, by implementing the system of the present invention, it is possible to efficiently detect signs of window dressing and take the most appropriate action depending on the user's emotional state.

[1191] The processing flow will be explained below.

[1192] Step 1:

[1193] The server collects company financial data, audit reports, news articles, and historical accounting fraud data from various sources, including publicly available financial statements, third-party auditor reports, and even public reports, via the internet and internal systems.

[1194] Step 2:

[1195] The server cleans the collected data. Specifically, it removes duplicates and errors from the data. It also supplements missing data from other data and standardizes the format of numerical data and dates. For example, it standardizes data in different currency units and converts them to the "YYYY-MM-DD" format.

[1196] Step 3:

[1197] The server uses the preprocessed data to build a machine learning model by first splitting the data into a training set, a test set, and a validation set, and then using the training set to apply techniques such as anomaly detection algorithms, random forests, or deep learning to train the model.

[1198] Step 4:

[1199] The server evaluates the trained model on the test set to check precision and recall, and retrains the model as needed to optimize performance.

[1200] Step 5:

[1201] The user (accountant) inputs the latest financial statement data of the company being audited into the terminal, such as data on sales, operating profit, and costs.

[1202] Step 6:

[1203] The terminal transmits the input data to the server in real time.

[1204] Step 7:

[1205] The server uses generative AI and trained models to analyze the latest financial data it receives, comparing the input data with past data to detect any abnormal patterns, such as sudden increases in sales or cost inconsistencies.

[1206] Step 8:

[1207] Based on the analysis results, the server predicts the possibility of window dressing, including specific suspicious points such as sudden increases in sales or fluctuations in cost structure.

[1208] Step 9:

[1209] The device's built-in emotion engine analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. For example, it can determine their stress level based on changes in their voice tone and facial expressions.

[1210] Step 10:

[1211] Based on the emotion engine's findings, the server adjusts the content of the notification. For example, if the user is feeling stressed, the notification will be brief and clear.

[1212] Step 11:

[1213] Based on the predictions, the server generates specific suspicious alerts, including details such as unusual sales spikes or fluctuations in cost structures.

[1214] Step 12:

[1215] The server sends this notification to the terminal in real time.

[1216] Step 13:

[1217] The terminal receives the notification and displays it to the user (accountant), who can then immediately begin a detailed audit, for example by gathering additional information on specific items or preparing questions for the company.

[1218] Step 14:

[1219] The server logs the user's emotional state as determined by the emotion engine and stores it in a database for later analysis and improvement, for example, to understand the user's emotional trends over a long period of time and to help further optimize the system.

[1220] Example 2

[1221] Next, a description will be given of Example 2. In the following description, the data processing device 12 will be referred to as a "server" and the robot 414 will be referred to as a "terminal."

[1222] In audit work, it is important to quickly and accurately detect corporate accounting fraud, but conventional methods require a great deal of time and effort from data collection to analysis. Furthermore, notification methods that do not take into account the stress and emotional state of accountants may reduce the efficiency of audit work. To solve these problems, efficient data processing and prediction methods, as well as appropriate responses according to the user's emotional state, are required.

[1223] The specific processing by the specific processing unit 290 of the data processing device 12 in the second embodiment is realized by the following means.

[1224] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, a notification means, and a log recording means, which enable fast and accurate detection of corporate window dressing and provide an optimal notification method according to the user's emotional state.

[1225] "Data collection methods" are methods for collecting corporate financial information, audit reports, news articles, and information on past accounting fraud cases from online and offline sources.

[1226] "Data preprocessing means" refers to means for cleaning collected information, filling in missing data, and normalizing information.

[1227] A "machine learning model building means" is a means for training and building a machine learning model using preprocessed data.

[1228] "Predictive methods using generative AI" are methods that analyze a company's financial information based on machine learning models and predict the possibility of accounting fraud.

[1229] The "emotion recognition means" is a means for analyzing the user's voice or facial expression and determining the user's emotional state.

[1230] The "notification means" is a means for adjusting and providing notification content according to the prediction result and the emotional state determined by the emotion recognition means.

[1231] The "log recording means" is a means for recording the emotional state of the user determined by the emotion recognition means and storing it in a database.

[1232] The system of the present invention includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, a notification means, and a log recording means. This system is designed to efficiently detect corporate window dressing, reduce the burden of accounting audits, and respond according to the user's emotional state.

[1233] First, the server uses data collection tools to collect company financial information, audit reports, news articles, and information on past accounting fraud cases from the Internet and internal systems. Specifically, it efficiently collects the required information using Python libraries (such as BeautifulSoup and Selenium) and APIs (such as FinAPI). For example, it can collect Company A's financial statement data for the past 10 years, audit reports, and publicly available related news articles.

[1234] Next, the server uses data preprocessing tools to clean the collected data, impute missing data, and normalize the data. Specifically, it removes duplicate data using the drop_duplicates() function in the pandas library, detects missing values ​​using the isnull() function, and imputes them using statistical methods. It also converts numeric data to a unified currency unit (e.g., USD) and standardizes date formats to "YYYY-MM-DD" using the datetime module.

[1235] The server then uses the preprocessed data to build a machine learning model using a machine learning model building method. The data set is split into a training set, a test set, and a validation set, and the train_test_split function in the sklearn library is used. The server also uses Isolation Forest or Random Forest Classifier for feature extraction to train and build the model. For example, an anomaly detection algorithm is used to extract features using indicators such as sudden sales increases or fluctuations in cost structure.

[1236] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal sends the data to the server in real time. The server analyzes the input data using a predictive method that uses generative AI, and compares it with past data to predict the possibility of window dressing. The generative AI uses a model such as OpenAI's GPT-4. For example, a prompt statement could be entered as follows: "Enter the latest financial data of Company B, have the server analyze it, and detect that sales have increased by 50% compared to the same period last year, but the operating profit margin has remained unchanged, and predict that there is a high possibility of window dressing."

[1237] Furthermore, the device's built-in emotion recognition means analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. It uses the librosa library for voice analysis and the OpenCV and dlib libraries for facial expression recognition. For example, it detects changes in voice tone or abnormal facial expressions to determine whether the user is in a high-stress state. This information is reflected in the notification content sent by the prediction result notification means.

[1238] The server uses a real-time notification mechanism to generate a notification containing specific suspicious points based on the prediction results. This notification details any abnormal patterns, such as sudden increases in sales or fluctuations in cost structure. The notification content is adjusted to be concise and clear, taking into account the user's emotional information determined by the emotion recognition mechanism. For example, a notification may be generated that states, "Sales in the second quarter have increased by 50% compared to the same month last year, but there has been no significant change in the operating profit margin, so there is a high possibility of abnormal transactions." This notification is sent to the terminal and displayed to the user (accountant).

[1239] Finally, the server uses the logging means to log the user's emotional state as determined by the emotion recognition means and stores the log in a database. This data is used to understand the user's emotional trends over a long period of time and to further optimize the system. For example, if a user is in a state of high stress under certain conditions, this data can be used as feedback to revise the system's interface and notification methods.

[1240] The above configuration and processing make it possible to quickly and accurately detect corporate window dressing and respond optimally according to the user's emotional state.

[1241] The flow of the identification process in the second embodiment will be described with reference to FIG.

[1242] Processing flow

[1243] Step 1: Collect data

[1244] The server collects corporate financial information, audit reports, news articles, and information on past accounting fraud cases via the internet and internal systems. It efficiently collects the necessary data using libraries such as Python's BeautifulSoup and Selenium, and APIs such as FinAPI. Specifically, it retrieves data from the API using the requests library, and parses HTML with BeautifulSoup to extract financial information.

[1245] Input: URL or API endpoint to collect data

[1246] Output: Dataset of company financial information, audit reports, press articles, and information on past accounting fraud cases

[1247] Step 2: Preprocessing the data

[1248] The server cleans the collected data, imputes missing data, and normalizes the data. Specifically, it uses the pandas library to remove duplicate data, the isnull() function to detect missing values, and statistical methods to impute them. It also standardizes date formats using the datetime module and converts numeric data to a uniform currency unit.

[1249] Input: Dataset of company financial information, audit reports, news articles, and information on past accounting fraud cases

[1250] Output: A clean and normalized dataset

[1251] Step 3: Building a learning model

[1252] The server builds a machine learning model using the preprocessed data. It splits the dataset into training, test, and validation sets using the train_test_split function in the sklearn library, and uses IsolationForest and RandomForestClassifier for feature extraction. After training the model, it evaluates its accuracy on the test set.

[1253] Input: Clean and normalized dataset

[1254] Output: A trained machine learning model

[1255] Step 4: Predict using generative AI

[1256] The user (accountant) inputs the latest financial data of the company being audited into the terminal. The terminal sends the data in real time to the server. The server uses generative AI and a trained model to analyze the input data and predict the possibility of window dressing. The generative AI uses a model such as GPT-4.

[1257] Input: Latest financial data

[1258] Output: Prediction result regarding the possibility of window dressing

[1259] Step 5: Emotion Engine in Action

[1260] The emotion engine installed on the device analyzes the user's (accountant's) voice and facial expressions to determine their emotional state. Changes in voice tone are analyzed using librosa, and facial expressions are evaluated using OpenCV and dlib. This determines stress levels and emotional states.

[1261] Input: User's voice data, facial expression data

[1262] Output: User's emotional state

[1263] Step 6: Real-time notifications

[1264] The server generates a notification based on the prediction results, including specific suspicious points, such as sudden sales increases or fluctuations in cost structure, and adjusts the content and tone of the notification based on the emotional information determined by the emotion recognition means.

[1265] Input: Prediction result, user's emotional state

[1266] Output: Adjusted notification content

[1267] Step 7: Logging

[1268] The server records the user's emotional state as determined by the emotion engine as a log and stores it in a database. The log data is used to improve the system and optimize user responses.

[1269] Input: User's emotional state

[1270] Output: Saved log data

[1271] (Application example 2)

[1272] Next, a description will be given of Application Example 2. In the following description, the data processing device 12 will be referred to as a "server" and the robot 414 will be referred to as a "terminal."

[1273] In modern business activities, window dressing remains a serious problem, placing a heavy burden on accountants and auditors. Audit work also creates significant stress and strain. To address this issue, a system is needed that efficiently detects abnormal transactions and provides appropriate notifications based on the accountant's emotional state. The present invention aims to improve corporate reliability and reduce the burden of accounting audits by solving these problems.

[1274] The specific processing by the specific processing unit 290 of the data processing device 12 in the application example 2 is realized by the following means.

[1275] In this invention, the server includes a data collection means, a data preprocessing means, a machine learning model construction means, a prediction means using generative AI, an emotion recognition means, and a real-time notification means, which enable efficient detection of anomalies in a company's financial data and appropriate notification according to the user's emotional state.

[1276] "Data collection methods" are methods for collecting financial data, audit reports, information articles, and historical anomaly data from online or offline sources.

[1277] "Data preprocessing means" refers to methods for cleaning collected data, filling in missing data, and normalizing the data into a format that is easy to analyze.

[1278] "Machine learning model building means" refers to a method of using pre-processed data to apply anomaly detection algorithms and other machine learning techniques to build a model for detecting anomalous transactions.

[1279] "Prediction methods using generative AI" are methods that use trained machine learning models and generative AI to analyze newly obtained data and predict the possibility of abnormal transactions.

[1280] "Emotion recognition means" is a method for determining the user's emotional state by analyzing the user's voice and facial expressions, and adjusting the system's response based on that information.

[1281] "Real-time notification means" is a method of generating a notification based on the results of abnormal transaction detection, with content that corresponds to the user's emotional state, and quickly notifying the user.

[1282] The present invention is a system for monitoring and detecting anomalies in corporate financial data, and is intended to be used on end-user devices such as smartphone applications. Specific procedures for implementing the invention are described below.

[1283] Hardware and software used

[1284] Hardware: Smartphones, cloud servers

[1285] Software: Python, Scikit-learn, Keras, EmotionEngine, NotificationService

[1286] System configuration

[1287] 1. Data collection methods:

[1288] The server collects financial data, audit reports, news articles, and past anomaly data from online and offline sources. For example, financial data is collected through APIs, and audit reports and news articles are collected using a web crawling system.

[1289] 2. Data preprocessing methods:

[1290] Once collected, the data is filled in with missing values, errors corrected, and formatted in a consistent manner (for example, missing sales data is filled in from historical data, and date formats are standardized).

[1291] 3. Machine learning model building methods:

[1292] The preprocessed data is used to apply an anomaly detection algorithm and build a model. Specifically, the data is divided into a training set and a test set, and training is performed using a random forest or deep learning model.

[1293] 4. Generative AI prediction methods:

[1294] When a user inputs their latest financial data, the data is sent to the server, where generative AI and trained models are used to predict anomalies. For example, if sales have increased significantly but the operating profit margin remains unchanged, the system predicts possible anomalies.

[1295] 5. Emotion recognition means:

[1296] Analyzes the user's voice and facial expressions to determine their emotional state. Emotion Engine is used to analyze voice tone and facial expression changes to determine the user's stress level.

[1297] 6. Real-time notification methods:

[1298] Based on the anomaly detection results, notifications are generated with content that corresponds to the user's emotional state. For example, if the user is particularly stressed, the notification content will be brief and only the main points will be presented.

[1299] Specific examples of realization

[1300] 1. Examples of data collection:

[1301] Use APIs to retrieve company financial data and collect audit reports for the past 10 years.

[1302] 2. Examples of data preprocessing:

[1303] Remove duplicate entries from collected data, impute missing data using statistical methods, convert all numerical data to a single currency, and standardize date formats.

[1304] 3. Specific examples of building learning models:

[1305] The preprocessed data is used to train a model using the random forest algorithm, and the performance of the anomaly detection model is verified.

[1306] 4. Specific examples of predictions using generative AI:

[1307] Users enter their company's latest financial data, which is then analyzed on the server, and if there is a high possibility of an abnormality, an alert is issued immediately.

[1308] 5. Examples of emotion recognition:

[1309] The system analyzes the user's voice and, if it determines that the stress level is high, simplifies the notification content.

[1310] 6. Examples of real-time notifications:

[1311] "Second-quarter sales have surged 50% compared to the same month last year, but there has been no significant change in operating profit margins, so there is a high possibility that this is an abnormal transaction," the notice said.

[1312] Prompt Sentence Examples

[1313] "Evaluate a company's latest financial data and detect anomalies. Write a program that adjusts the tone of notifications based on the user's stress level."

[1314] The flow of the specific processing in the application example 2 will be described with reference to FIG.

[1315] Step 1:

[1316] Data collection by data collection method

[1317] The server collects financial data, audit reports, information articles, and historical anomaly data from online and offline sources. Specifically, it retrieves financial data using APIs and collects news articles using a web crawling system. It uses the URL or API endpoint of each source as input and obtains the collected raw data as output.

[1318] Step 2:

[1319] Data cleaning through data preprocessing methods

[1320] The server cleans the collected raw data, imputes missing data, and normalizes the data into a format that is easy to analyze. Specifically, it removes duplicate data, imputes missing data (e.g., imputation using statistical methods), and standardizes date formats. It uses the collected raw data as input and obtains a cleaned dataset as output.

[1321] Step 3:

[1322] Building models using machine learning model building methods

[1323] The server uses the preprocessed data to apply an anomaly detection algorithm and build a model. Specifically, it splits the data into a training set and a test set, and trains the model using a random forest or deep learning algorithm. The preprocessed dataset is used as input, and the trained model is obtained as output.

[1324] Step 4:

[1325] Predicting anomalies using generative AI-based predictive methods

[1326] When a user inputs their latest financial data, the device sends the data to the server. The server then uses generative AI and trained models to analyze this data and predict anomalies. Specifically, it detects sudden increases in sales or inconsistencies in cost structures. The latest financial data provided by the user is used as input, and a prediction result indicating whether or not anomalies exist is obtained as output.

[1327] Step 5:

[1328] Emotional state analysis using emotion recognition tools

[1329] The device analyzes the user's voice and facial expressions in real time to determine their emotional state. Specifically, it uses the Emotion Engine to analyze changes in voice tone and facial expressions to determine stress levels and feelings of relief. The device uses the user's voice and facial expression data as input and outputs the determined emotional state.

[1330] Step 6:

[1331] Generate and send notifications via real-time notification methods

[1332] The server generates a notification based on the anomaly detection results, with content appropriate to the user's emotional state. Specifically, if an anomaly is detected, a detailed explanation is provided, and if the user is in a stressful state, a brief notification is provided. The anomaly detection results and the user's emotional state are used as inputs, and the generated notification message is obtained as output.

[1333] The specific processing unit 290 transmits the result of the specific processing to the robot 414. In the robot 414, the control unit 46A causes the speaker 240 and the control target 443 to output the result of the specific processing. The microphone 238 acquires voice indicating a user input regarding the result of the specific processing. The control unit 46A transmits voice data indicating the user input acquired by the microphone 238 to the data processing device 12. In the data processing device 12, the specific processing unit 290 acquires the voice data.

[1334] The data generation model 58 is a so-called generative AI (Artificial Intelligence). An example of the data generation model 58 is ChatGPT (Internet Search<URL: https: / / openai.com / blog / chatgpt> ), Gemini (Internet search <url: https: gemini.google.com ?hl="ja">) and other generation AIs. The data generation model 58 is obtained by performing deep learning on a neural network. A prompt including an instruction is input to the data generation model 58, and inference data such as voice data indicating voice, text data indicating text, and image data indicating an image is also input. The data generation model 58 performs inference on the input inference data in accordance with the instruction indicated by the prompt, and outputs the inference result in a data format such as voice data and text data. Here, inference refers to, for example, analysis, classification, prediction, and / or summarization.

[1335] In the above embodiment, an example was given in which the specific processing is performed by the data processing device 12, but the technology of the present disclosure is not limited to this, and the specific processing may be performed by the robot 414.

[1336] The emotion identification model 59 as an emotion engine may determine the user's emotion according to a specific mapping. Specifically, the emotion identification model 59 may determine the user's emotion according to an emotion map (see FIG. 9), which is a specific mapping. Similarly, the emotion identification model 59 may determine the robot's emotion, and the identification processing unit 290 may perform identification processing using the robot's emotion.

[1337] FIG. 9 is a diagram illustrating an emotion map 400 on which multiple emotions are mapped. In the emotion map 400, emotions are arranged in concentric circles radiating from the center. Emotions closer to the center of the concentric circles are more primitive. Emotions representing states and actions arising from a state of mind are arranged on the outer edges of the concentric circles. The concept of emotion includes both affect and mental states. Emotions generally generated from reactions occurring in the brain are arranged on the left side of the concentric circles. Emotions generally induced by situational judgment are arranged on the right side of the concentric circles. Emotions generally generated from reactions occurring in the brain and induced by situational judgment are arranged on the upper and lower sides of the concentric circles. Furthermore, the emotion of "pleasure" is arranged on the upper side of the concentric circles, and the emotion of "discomfort" is arranged on the lower side. In this way, in the emotion map 400, multiple emotions are mapped based on the structure by which emotions are generated, and emotions that tend to occur simultaneously are mapped close to each other.

[1338] These emotions are distributed in the 3 o'clock direction on emotion map 400, and typically fluctuate between relief and anxiety. In the right half of emotion map 400, situational awareness dominates over internal sensations, resulting in a sense of calm.

[1339] The inside of emotion map 400 represents what is going on in the mind, and the outside of emotion map 400 represents behavior, so the further you go outside emotion map 400, the more visible the emotions become (the more they are expressed in behavior).

[1340] Human emotions are based on various balances, such as posture and blood sugar levels. When these balances deviate from the ideal, a state of discomfort is indicated, and when they approach the ideal, a state of pleasure is indicated. Emotions can also be created for robots, automobiles, and motorcycles, based on various balances, such as posture and remaining battery life. When these balances deviate from the ideal, a state of discomfort is indicated, and when they approach the ideal, a state of pleasure is indicated. An emotion map can be generated, for example, based on Dr. Mitsuyoshi's emotion map (Research on Voice Emotion Recognition and Emotional Brain Physiological Signal Analysis Systems, Tokushima University, Doctoral Dissertation: https: / / ci.nii.ac.jp / naid / 500000375379). The left half of the emotion map lists emotions belonging to the "reaction" domain, where sensation is dominant. The right half of the emotion map lists emotions belonging to the "situation" domain, where situational awareness is dominant.

[1341] The emotion map defines two emotions that promote learning. One is a negative emotion on the situation side, around the middle of "repentance" or "reflection." In other words, this occurs when the robot experiences negative emotions such as "I never want to feel this way again" or "I don't want to be scolded again." The other is a positive emotion on the response side, around "desire." In other words, this occurs when the robot experiences positive feelings such as "I want more" or "I want to know more."

[1342] The emotion identification model 59 inputs user input into a pre-trained neural network, obtains emotion values ​​indicating each emotion shown in the emotion map 400, and determines the user's emotion. This neural network is pre-trained based on multiple pieces of training data that are combinations of user input and emotion values ​​indicating each emotion shown in the emotion map 400. Furthermore, this neural network is trained so that emotions that are located close to each other have similar values, as in the emotion map 900 shown in FIG. 10. FIG. 10 shows an example in which multiple emotions, "relieved," "calm," and "reassuring," have similar emotion values.

[1343] The system according to the present disclosure has been described above mainly with respect to the functions of the data processing device 12, but the system according to the present disclosure is not necessarily implemented on a server. The system according to the present disclosure may be implemented as a general information processing system. The present disclosure may be implemented, for example, as a software program running on a personal computer or an application running on a smartphone, etc. The method according to the present disclosure may be provided to users in the form of SaaS (Software as a Service).

[1344] In the above embodiment, an example was given in which the specific processing is performed by one computer 22, but the technology of the present disclosure is not limited to this, and the specific processing may be distributed and performed by a plurality of computers including the computer 22. For example, the data generation model 58 may be provided in an external device of the data processing device 12, and data may be generated in the external device in accordance with input data.

[1345] In the above embodiment, an example in which the specific processing program 56 is stored in the storage 32 has been described, but the technology of the present disclosure is not limited to this. For example, the specific processing program 56 may be stored in a portable, computer-readable, non-transitory storage medium such as a USB (Universal Serial Bus) memory. The specific processing program 56 stored in the non-transitory storage medium is installed in the computer 22 of the data processing device 12. The processor 28 executes the specific processing in accordance with the specific processing program 56.

[1346] Alternatively, the specific processing program 56 may be stored in a storage device such as a server connected to the data processing device 12 via the network 54, and the specific processing program 56 may be downloaded and installed on the computer 22 in response to a request from the data processing device 12.

[1347] It is not necessary to store all of the specific processing program 56 in a storage device such as a server connected to the data processing device 12 via the network 54, or to store all of the specific processing program 56 in the storage 32; only a portion of the specific processing program 56 may be stored.

[1348] The hardware resource for executing a specific process can be any of the following processors: An example of a processor is a CPU, which is a general-purpose processor that functions as a hardware resource for executing a specific process by executing software, i.e., a program. Another example of a processor is a dedicated electrical circuit, such as an FPGA (Field-Programmable Gate Array), a PLD (Programmable Logic Device), or an ASIC (Application Specific Integrated Circuit), which is a processor with a circuit configuration designed specifically for executing a specific process. Each processor has built-in or connected memory, and each processor uses the memory to execute the specific process.

[1349] The hardware resource that executes the specific processing may be configured with one of these various processors, or may be configured with a combination of two or more processors of the same or different types (for example, a combination of multiple FPGAs, or a combination of a CPU and an FPGA). Also, the hardware resource that executes the specific processing may be a single processor.

[1350] As an example of a system configured with a single processor, first, one processor is configured by combining one or more CPUs and software, and this processor functions as a hardware resource that executes a specific process. Second, there is a system that uses a processor that realizes the functions of an entire system including multiple hardware resources that execute a specific process on a single IC chip, as typified by SoC (System-on-a-chip). In this way, a specific process is realized using one or more of the above-mentioned various processors as hardware resources.

[1351] Furthermore, the hardware structure of these various processors can be, more specifically, an electric circuit that combines circuit elements such as semiconductor devices. The specific processing described above is merely an example. Therefore, it goes without saying that unnecessary steps may be deleted, new steps may be added, or the processing order may be rearranged, without departing from the spirit of the invention.

[1352] The above-described description and illustrations are a detailed explanation of the parts related to the technology of the present disclosure and are merely an example of the technology of the present disclosure. For example, the above description of the configuration, functions, actions, and effects is an explanation of an example of the configuration, functions, actions, and effects of the parts related to the technology of the present disclosure. Therefore, it goes without saying that unnecessary parts may be deleted, new elements may be added, or replacements may be made to the above-described description and illustrations within the scope of the gist of the technology of the present disclosure. Furthermore, to avoid confusion and facilitate understanding of the parts related to the technology of the present disclosure, the above-described description and illustrations omit explanations of common technical knowledge that do not require particular explanation to enable the implementation of the technology of the present disclosure.

[1353] All publications, patent applications, and technical standards mentioned in this specification are herein incorporated by reference to the same extent as if each individual publication, patent application, or technical standard was specifically and individually indicated to be incorporated by reference.

[1354] The following is further disclosed regarding the above embodiment.

[1355] (Claim 1)

[1356] data collection means;

[1357] data preprocessing means;

[1358] A machine learning model building method;

[1359] A prediction method using generative AI,

[1360] A notification means;

[1361] A system including:

[1362] (Claim 2)

[1363] 2. The system according to claim 1, wherein the data collection means collects corporate financial data, audit reports, news articles and past accounting fraud case data from online and offline sources.

[1364] (Claim 3)

[1365] 2. The system of claim 1, wherein the data preprocessing means cleans the collected data, imputes missing data, and normalizes the data.

[1366] (Claim 4)

[1367] The system of claim 1, wherein the machine learning model building means uses the preprocessed data to divide it into a training set, a test set, and a validation set, and applies a machine learning algorithm to train the model.

[1368] (Claim 5)

[1369] The system described in 1 is characterized in that the prediction means using the generation AI inputs the latest financial statement data and predicts the possibility of window dressing by comparing it with past data.

[1370] (Claim 6)

[1371] 2. The system according to claim 1, wherein the notification means notifies the accounting auditor of specific suspicious points along with the prediction results in real time.

[1372] "Example 1"

[1373] (Claim 1)

[1374] data collection means;

[1375] data preprocessing means;

[1376] A machine learning model building method;

[1377] A prediction method using generative AI,

[1378] Real-time notification means;

[1379] A system including:

[1380] (Claim 2)

[1381] 2. The system according to claim 1, wherein the data collection means collects corporate financial data, audit reports, news articles and past accounting fraud case data from online and offline sources.

[1382] (Claim 3)

[1383] 2. The system of claim 1, wherein the data preprocessing means cleans the collected data, imputes missing data, and normalizes the data.

[1384] (Claim 4)

[1385] The system described in claim 1, characterized in that the machine learning model construction means divides the preprocessed data into a training set, a test set, and a validation set, constructs a model using an anomaly detection algorithm, random forest, and deep learning, and evaluates precision and recall.

[1386] (Claim 5)

[1387] The system described in claim 1 is characterized in that the prediction means using the generation AI receives the latest financial statement data of a company entered by the user, analyzes it using a learning model and the generation AI, detects anomalies in the data, and predicts the possibility of window dressing.

[1388] (Claim 6)

[1389] The system according to claim 1, wherein the real-time notification means generates a notification including suspicious points based on the predicted results, and transmits and displays the notification to the user terminal.

[1390] "Application Example 1"

[1391] (Claim 1)

[1392] data collection means;

[1393] data preprocessing means;

[1394] A machine learning model building method;

[1395] A prediction method using generative AI,

[1396] a notification means for notifying a specific device;

[1397] A means of collecting and analyzing data from multiple corporate sources in real time;

[1398] A means of inputting new standardized data into the generative AI model; and

[1399] A means for immediately notifying the user based on the anomaly detection result;

[1400] A means of detecting corporate financial fraud using the prediction results of generative AI;

[1401] A system including:

[1402] (Claim 2)

[1403] The system according to claim 1, characterized in that the data collection means collects corporate financial data, audit reports, news articles and past accounting fraud case data from online and offline sources, and performs the collection in real time.

[1404] (Claim 3)

[1405] The system of claim 1, wherein the data preprocessing means cleans the collected data, imputes missing data, and normalizes the data before inputting it into the generative AI model.

[1406] "Example 2: Combining Emotion Engines"

[1407] (Claim 1)

[1408] data collection means;

[1409] data preprocessing means;

[1410] A machine learning model building method;

[1411] A prediction method using generative AI,

[1412] An emotion recognition means;

[1413] A notification means;

[1414] A log recording means;

[1415] A system including:

[1416] (Claim 2)

[1417] 2. The system according to claim 1, wherein the data collection means collects corporate financial information, audit reports, news articles, and information on past accounting fraud cases from online and offline sources.

[1418] (Claim 3)

[1419] 2. The system of claim 1, wherein the data preprocessing means cleans the collected information, imputes missing data, and normalizes the information.

[1420] (Claim 4)

[1421] The system described in claim 1, characterized in that the predictive means using generative AI analyzes a company's financial information based on a machine learning model and predicts the possibility of accounting fraud.

[1422] (Claim 5)

[1423] 2. The system according to claim 1, wherein the emotion recognition means analyzes the user's voice or facial expression to determine the user's emotional state.

[1424] (Claim 6)

[1425] 2. The system according to claim 1, wherein the notification means adjusts and provides notification content according to the prediction result and the emotional state determined by the emotion recognition means.

[1426] (Claim 7)

[1427] 2. The system according to claim 1, wherein the log recording means records the emotional state of the user determined by the emotion recognition means and stores the recorded state in a database.

[1428] "Application example 2 when combining emotion engines"

[1429] (Claim 1)

[1430] data collection means;

[1431] data preprocessing means;

[1432] A machine learning model building method;

[1433] A prediction method using generative AI,

[1434] An emotion recognition means;

[1435] Real-time notification means;

[1436] A system including:

[1437] (Claim 2)

[1438] 2. The system according to claim 1, wherein the data collection means collects financial data, audit reports, information articles, and past abnormal case data from online and offline sources.

[1439] (Claim 3)

[1440] 2. The system according to claim 1, wherein the data preprocessing means cleans the collected data, completes missing data, and normalizes the data.

[1441] (Claim 4)

[1442] 2. The system according to claim 1, wherein the emotion recognition means analyzes the user's voice and facial expressions to determine the user's emotional state and adjusts the notification method based on the determined emotional state.

[1443] (Claim 5)

[1444] The system according to claim 1, wherein the real-time notification means generates a notification based on the anomaly detection result and according to the emotional state of the user. [Explanation of symbols]

[1445] 10, 210, 310, 410 Data Processing Systems 12 Data Processing Device 14 Smart Devices 214 Smart Glasses 314 Headset-type terminal 414 Robot< / url:> < / url:> < / url:> < / url:>

Claims

[Claim 1] A data collection means for collecting corporate financial data, audit reports, news articles and past accounting fraud case data from online and offline sources; ,Data preprocessing means for cleaning the collected data, ,imputing missing data, and normalizing the data; A machine learning model construction means for dividing the preprocessed data into a training set, a test set, and a validation set and applying a machine learning algorithm to train a model; A prediction method using generative AI that inputs the latest financial statement data and compares it with past data to predict the possibility of window dressing fraud. A notification method that notifies the accounting auditor of specific suspicious points along with the prediction results in real time, A system including:

Images