Finance and tax service request processing method and device

Abstract

The embodiment of the invention discloses a finance and taxation service request processing method and device, and the method comprises the following steps: receiving a finance and taxation service request, carrying out the semantic analysis through a BERT model trained based on a finance and taxation field corpus, and extracting a user intention and a key service entity; based on a finite-state machine model, managing a dialogue flow of a business scene, triggering state transition according to a current dialogue state, a user intention and a key business entity, and orderly promoting an invoicing, checking or diagnosing business process; scheduling a corresponding AI capability module to execute a task according to the user intention and a service scene defined by the key service entity; an execution result of the AI capability module is fed back to a front end in real time through a server pushing technology and is displayed in linkage with information input by a user; based on key business entities and operation combinations recorded in a user operation log, the key business entities and the operation combinations are automatically abstracted and packaged into independent reusable functional components. According to the embodiment of the invention, finance and tax business requirements can be adapted through multi-step and scenarized dialogue response.

Description

Technical Field

[0001] This application belongs to the field of computer technology, and specifically relates to a method and apparatus for processing financial and tax business requests. Background Technology

[0002] With the rapid development of information technology, financial and tax technology is also constantly advancing, becoming an indispensable force in the financial and tax field. From e-taxation to intelligent financial and taxation, and then to the application of cloud computing and big data technologies, the landscape of financial and tax technology is constantly evolving. These technologies not only improve the efficiency of financial and tax management but also help enterprises conduct more accurate data analysis and decision-making.

[0003] However, traditional financial and tax software relies on menu-driven operations, with fragmented functional modules requiring users to frequently switch pages, resulting in lengthy operation processes. New users also need to spend a significant amount of time learning and find it difficult to master. In scenarios such as invoice processing, the identification information displayed on the front end and the compliance verification results are independent of each other, lacking a linkage mechanism. Users need to manually associate and match data, which is prone to errors. Furthermore, the interaction method of existing technologies is fixed to a single-turn command response, making it difficult to adapt to the multi-step and scenario-based processing needs of financial and tax businesses, resulting in a severe lack of flexibility.

[0004] Application content

[0005] The purpose of this application is to provide a method and apparatus for processing financial and tax business requests, so as to solve the defects of poor interactive adaptability and high operating costs of the prior art.

[0006] To solve the above-mentioned technical problems, this application is implemented as follows:

[0007] Firstly, a method for processing financial and tax business requests is provided, including the following steps:

[0008] It receives financial and tax business requests input by users via text, voice, or images, and performs semantic parsing using a BERT model trained on financial and tax corpus to extract user intent and key business entities.

[0009] The dialogue flow of the business scenario is managed based on the finite state machine model. According to the current dialogue state, the user intent and the key business entity, the state transition is triggered to orderly advance the invoicing, verification or diagnosis business process.

[0010] Based on the user intent and the business scenario defined by the key business entity, the corresponding AI capability module is scheduled to execute the task. The AI ​​capability module uses the key business entity as input parameter for processing. The AI ​​capability module includes an OCR model for invoice recognition, a decision tree model for recommending commodity codes and tax rates, and a risk diagnosis model for detecting financial and tax anomalies.

[0011] The execution results of the AI ​​capability module are fed back to the front end in real time through server-side push technology and displayed in conjunction with user input information; based on the key business entities and operation combinations recorded in the user operation log, they are automatically abstracted and encapsulated into independent reusable functional components.

[0012] Secondly, a processing device for financial and tax business requests is provided, comprising:

[0013] The receiving module is used to receive financial and tax business requests input by users through text, voice or images, and to perform semantic parsing using a BERT model trained on financial and tax domain corpus to extract user intent and key business entities.

[0014] The management module is used to manage the dialogue flow of business scenarios based on a finite state machine model. It triggers state transitions based on the current dialogue state, the user intent, and the key business entities, and orderly advances the invoicing, verification, or diagnosis business processes.

[0015] The scheduling module is used to schedule the corresponding AI capability modules to perform tasks according to the user intent and the business scenario defined by the key business entity. The AI ​​capability modules use the key business entity as input parameters for processing. The AI ​​capability modules include an OCR model for invoice recognition, a decision tree model for recommending commodity codes and tax rates, and a risk diagnosis model for detecting financial and tax anomalies.

[0016] The feedback module is used to push the execution results of the AI ​​capability module to the front end in real time via server-side push technology, and display them in conjunction with user input information. Based on the key business entities and operation combinations recorded in the user operation log, they are automatically abstracted and encapsulated into independent reusable functional components.

[0017] This application embodiment constructs an interaction system centered on scenario-based dialogue flow, which can complete complex financial and tax business processing without frequent page switching, avoids errors caused by users manually associating data, and improves the flexibility of the interaction method, adapting to financial and tax business needs with multi-step, scenario-based dialogue responses. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a method for processing financial and tax business requests provided in an embodiment of this application;

[0019] Figure 2 This is a schematic diagram of the structure of a processing device for financial and tax business requests provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Existing financial and tax processing technologies have several significant shortcomings, as detailed below:

[0022] 1. High operating costs: Traditional financial and tax software relies on menu-based operation, with scattered functional modules, requiring users to frequently switch pages, resulting in lengthy operation processes; new users need to spend a lot of time learning and are difficult to master.

[0023] 2. Data linkage gap: In scenarios such as invoice processing, the identification information displayed on the front end and the compliance verification results are independent of each other and lack a linkage mechanism. Users need to manually link and match the data, which is prone to errors.

[0024] 3. Limitations of AI Functions: The natural language interaction of existing financial and tax AI tools is mostly limited to single question-and-answer scenarios, which cannot cover the entire process of complex business such as invoicing and risk diagnosis, resulting in a disconnect between AI functions and actual business processes.

[0025] 4. Poor interaction adaptability: The interaction method is fixed to a single-turn command response, which is difficult to adapt to the multi-step and scenario-based processing needs of financial and tax business, and the flexibility is seriously insufficient.

[0026] 5. Weak functional scalability: AI capabilities are encapsulated into fixed modules, making it impossible to extract and abstract derivative functions such as invoice collection from dialogue interactions, which makes it difficult to meet the personalized financial and tax processing needs of enterprises.

[0027] 6. Imbalance in data value: Enterprise financial and tax data are highly sensitive. Existing systems struggle to balance ensuring data security with using data to train AI models. The value of user behavior data is not fully explored, directly impacting the effectiveness of intelligent recommendations and function optimization.

[0028] To address the aforementioned issues, this application provides an intelligent financial and tax processing system and its derivative functionalities based on a scenario-based chatflow, with the specific objectives as follows:

[0029] 1. Construct an interactive system centered on scenario-based dialogue flow to simplify operation processes, reduce learning costs, and enable complex financial and tax business processing without frequent page switching.

[0030] 2. Enable the linked presentation of front-end recognition information and compliance verification results to avoid errors caused by users manually associating data.

[0031] 3. Break away from the limitation of AI functions to single question and answer, and deeply integrate it into the entire business process such as invoicing and risk diagnosis.

[0032] 4. Enhance the flexibility of interaction methods, and adapt to the needs of financial and tax business with multi-step and scenario-based dialogue responses.

[0033] 5. The system is given the ability to extract and abstract derivative functions such as invoice collection from dialogue interactions to meet the personalized financial and tax processing needs of enterprises.

[0034] 6. While ensuring the security of corporate financial and tax data, fully explore the value of user behavior data, optimize intelligent recommendations and functional effects, and comprehensively improve the intelligence level of financial and tax processing and user experience.

[0035] Specifically, the core interaction system in this embodiment includes:

[0036] (1) Front-end interaction design

[0037] ① Dialogue Interface Components: Develop multimodal input components that support text input (integrating pinyin association and intelligent completion of financial and tax terms), voice input (using speech-to-text technology, supporting dialect recognition and interjection filtering), and image upload (supporting batch upload and drag-and-drop sorting). The input area displays historical dialogue records and current step guidance.

[0038] ②Scenario-based template display: The front end has a list of smart agents for scenarios such as invoicing and input verification. When the user clicks, they can switch to the corresponding smart agent. For example, if the user selects the "Output Invoicing" smart agent, the interface will immediately display invoicing-related information and interact with preset questions.

[0039] ③ Progress visualization: The top progress bar displays the progress of the current business in the conversation flow, such as the completion status of each node in the invoicing process: "Fill in information -> Verify -> Submit -> Complete". Uncompleted nodes are marked with the required operations, making it easy for users to understand the overall process.

[0040] (2) Backend engine support

[0041] ① Dialogue Flow Management Engine: Designed based on a finite state machine model, each business scenario corresponds to an independent state transition diagram. For example, the invoicing scenario includes states such as "waiting for buyer information", "waiting for product details", "verification in progress" and "completed". State switching is triggered by the user input and the system's automatic verification result, ensuring that the dialogue process proceeds in an orderly manner.

[0042] ② Intent recognition and parsing module: Accepts user input data from the front end, performs semantic parsing through the BERT pre-trained model, extracts key information (such as "invoice type = VAT invoice" and "amount = specific value" from "issue a VAT invoice with a specific amount", and maps it to the corresponding business interface).

[0043] In addition, the front-end and back-end data linkage mechanism in this embodiment includes:

[0044] (1) Front-end data processing and display

[0045] ① Data visualization component: Develop an invoice information visualization component, which displays a thumbnail of the scanned invoice on the left and displays the invoice information recognized by OCR in different areas on the right (user-customizable configuration).

[0046] ② Compliance Information Display: The compliance results are displayed in the official compliance result prompt window on the outside. Clicking the compliance result window will take you to the compliance details, which includes three statuses: (compliant, non-compliant, and unverified). Each status can be customized by the user to indicate whether it is compliant and the content.

[0047] ③ Real-time communication module: The model processes the results in real time through the SSE server, without requiring users to refresh to get the latest status, and has low server resource consumption.

[0048] (2) Backend data interaction and processing

[0049] ① API Gateway: As an intermediate layer for data interaction between the front-end and back-end, it is responsible for request routing (routing invoicing requests to the invoicing service and verification requests to the verification service), data format conversion (converting front-end JSON format data to the Protocol Buffers format required by the back-end service), request rate limiting (preventing high-frequency requests from causing system overload), and authentication (verifying the validity of user tokens).

[0050] ② Data caching mechanism: The information entered by the user in the dialogue flow is temporarily cached. For example, the purchase information entered by the user will be automatically filled in when the invoice is issued again within a certain period of time. The cached data is stored in encrypted form using the user ID to ensure data isolation. At the same time, a cache cleanup strategy is set to periodically delete expired data to free up storage space.

[0051] In addition, the integration of AI capabilities with business processes in this embodiment includes:

[0052] (1) Natural Language Processing Model

[0053] ① Model Architecture: The model is based on an existing open-source model with improvements. It is pre-trained on corpus of financial and tax terms, and word vector dimensions such as "tax number" and "tax rate" are added to improve the recognition accuracy of professional terms. The model supports online incremental training and new financial and tax policy terms and business language are regularly added to the training data.

[0054] ② Multi-turn dialogue processing: Supports contextual understanding. For example, if the user first enters "issued to Company A" and then enters "a specific amount", the model can recognize that "a specific amount" corresponds to Company A's invoice, without the user having to repeat the explanation. It can also handle referential relationships, such as in "its tax rate is a specific percentage", where "it" refers to the goods mentioned earlier.

[0055] (2) Machine learning model

[0056] ① Invoice verification model: Construct a rule base for matching commodity codes and tax rates, and train it with a decision tree model. When entering the commodity name when issuing an invoice, the model automatically recommends matching commodity codes and applicable tax rates, and calculates the matching score. When the score is lower than a certain standard, the front end displays the prompt "It is recommended to confirm the code and tax rate" and lists similar codes for selection.

[0057] ② Risk Diagnosis Model: An anomaly detection model is trained based on the company's historical financial and tax data (such as fluctuations in invoice amounts and changes in tax burden rates). The isolated forest algorithm is used to identify abnormal data points. For example, if a company's invoice amount increases significantly in a certain month compared to the previous month without a reasonable explanation, the model marks it as high risk. The system will automatically trigger the question "Is there any special business situation?" in the dialogue flow and display the abnormal indicators and historical comparison data in the risk report.

[0058] (3) Image recognition model

[0059] ①OCR Recognition Module: The backend adopts a deep learning model with a CNN+LSTM+CTC architecture, which is trained on invoice images. It supports automatic classification and element extraction of different types of invoices (VAT special invoices, ordinary invoices, electronic invoices, etc.). It preprocesses (deblurring, correction, and noise reduction) invoice images that are blurry, tilted, or have creases to improve recognition accuracy.

[0060] ② Result verification mechanism: After the recognition is completed, the model automatically performs a preliminary verification of the extracted key information (such as the format of the invoice code and number, and the validity of the amount). Information that fails the verification is marked as "suspicious" and is specially marked when pushed to the front end. For example, if the number of digits in the invoice code is abnormal, the front end will display "The invoice code recognition may be incorrect. It is recommended to check".

[0061] Furthermore, the derived function abstraction mechanism in this embodiment includes:

[0062] (1) Abstraction of front-end functional components

[0063] ① Invoice Recognition Component: An independent component is abstracted from the invoice processing flow of the dialog stream. It includes an image upload button, a preview window, a recognition progress bar, and a recognition result editing box. The component supports custom styles (such as button color and preview window size) and can be embedded in different pages. When calling it, you only need to pass in the scene parameters.

[0064] ② Compliance Verification Component: This component encapsulates compliance judgment logic and display style. After inputting invoice recognition information, the component automatically calls the backend compliance rule engine, returns the verification results, and displays them in a fixed format (such as the information list on the left and the compliance details on the right). It also supports customizing the display priority of verification rules through configuration files.

[0065] (2) Backend function encapsulation

[0066] ① Functional Abstraction Engine: By analyzing user operation logs in the dialogue flow, the frequency of each operation and data interaction pattern are statistically analyzed. When the weekly call frequency of a certain operation combination (such as "upload invoice -> recognition -> compliance verification") exceeds a certain number, the functional abstraction process is automatically triggered to generate independent API interfaces and functional descriptions.

[0067] ②Custom trigger configuration: The backend provides a configuration interface that allows administrators to set trigger conditions for derivative functions, such as "automatically enable batch collection mode when the number of invoices collected in a month exceeds a certain amount". After the mode is enabled, the frontend automatically displays the batch upload button and Excel import option, and the backend starts a multi-threaded processing mechanism accordingly.

[0068] The following description, in conjunction with the accompanying drawings, details a method for processing financial and tax business requests provided by the embodiments of this application through specific examples and application scenarios.

[0069] like Figure 1 The diagram shown is a flowchart of a method for processing financial and tax business requests according to an embodiment of this application. The method includes the following steps:

[0070] Step 101: Receive financial and tax business requests input by users via text, voice, or images, and perform semantic parsing using a BERT model trained on financial and tax domain corpus to extract user intent and key business entities.

[0071] The image can be an invoice image. After receiving the user's tax and financial service request via image input, the image input can be recognized and preprocessed using an OCR model composed of a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, and a Connection-Time Classification (CTC) model. After receiving the user's tax and financial service request via voice input, the request can be converted by calling a speech recognition API, filtering out interjections, and optimizing the matching of professional terms using a tax and financial corpus.

[0072] Step 102: Based on the finite state machine model, manage the dialogue flow of the business scenario, trigger state transition according to the current dialogue state, the user intent, and the key business entity, and orderly advance the invoicing, verification, or diagnosis business process.

[0073] The key business entities include one or more of the following: tax ID, product name, amount, invoice date, and company name. The finite state machine model defines business states including "waiting for buyer information," "waiting for product details," "verification in progress," and "completed." State transitions are triggered by user input or automatic system verification results. Accordingly, the parsed key business entities can be filled into the business interface parameters corresponding to the current dialogue state. Once the required parameters are filled, the state machine is automatically triggered to advance to the next state.

[0074] Step 103: Based on the user intent and the business scenario defined by the key business entity, schedule the corresponding AI capability module to execute the task. The AI ​​capability module uses the key business entity as input parameters for processing.

[0075] The AI ​​capability module includes an OCR model for invoice recognition, a decision tree model for recommending commodity codes and tax rates, and a risk diagnosis model for detecting financial and tax anomalies.

[0076] Specifically, for invoicing scenarios, a decision tree model can be used to match commodity codes and tax rates; for image verification scenarios, an OCR model can be used to identify invoice elements and perform format and logic checks; and for risk query scenarios, a risk diagnosis model based on the isolated forest algorithm can be used to analyze historical data.

[0077] Step 104: The execution results of the AI ​​capability module are fed back to the front end in real time through server-side push technology and displayed in conjunction with the user input information; based on the key business entities and operation combinations recorded in the user operation log, they are automatically abstracted and encapsulated into independent reusable functional components.

[0078] Specifically, if the frequency of calls to a combination of operations on a specific type of key business entity exceeds a preset threshold, the combination of operations can be encapsulated into an independent API and a corresponding front-end UI component can be generated.

[0079] This application embodiment constructs an interaction system centered on scenario-based dialogue flow, which can complete complex financial and tax business processing without frequent page switching, avoids errors caused by users manually associating data, and improves the flexibility of the interaction method, adapting to financial and tax business needs with multi-step, scenario-based dialogue responses.

[0080] In this embodiment, the user interaction layer is responsible for providing an intuitive and efficient human-computer interaction interface, supporting multimodal input (text, voice, image) and scenario-based template selection, and displaying processing progress and result data in real time, including components for invoice information visualization and compliance verification result display. The communication layer uses SSE server push technology to achieve real-time data synchronization, ensuring low-latency communication between the front-end and back-end processing engines, and supporting long-term connection maintenance and incremental data push.

[0081] In addition, the core engine layer includes:

[0082] (1) Dialogue Flow Management Engine: Implements multi-turn dialogue state management based on finite state machine.

[0083] (2) Intent recognition module: The BERT pre-trained model is used to parse the user's semantic intent.

[0084] (3) API Gateway: Unified interface management and request routing.

[0085] (4) Data caching mechanism: Temporarily store session data and intermediate processing results.

[0086] (5) Data labeling: Suspicious data is labeled when it is sent to the front end.

[0087] The AI ​​capability layer includes:

[0088] (1) Natural Language Processing Model: Supports understanding and contextual association of professional terms in the field of finance and taxation.

[0089] (2) Machine learning model: realize rule reasoning and predictive analysis such as invoice verification and risk diagnosis.

[0090] (3) Image recognition model: Invoice OCR recognition and image preprocessing are realized based on CNN+LSTM architecture.

[0091] Furthermore, the derivative functional layer is used for operation log analysis and high-frequency operation pattern recognition, automatic generation of UI components based on user behavior, custom business rule configuration, and batch task triggering.

[0092] In addition, risk diagnosis and early warning implementation includes:

[0093] (1) Data processing: Collect financial and tax data of enterprises for a certain period of time (de-identification processing).

[0094] (2) Triggering mechanism: At the same time, a risk report (including historical comparison line chart) is generated and pushed to the enterprise administrator's email in PDF format.

[0095] Taking the system's invoicing process as an example, the core functional flow of the aforementioned financial and tax processing system is as follows:

[0096] 1. Input layer module

[0097] Module function: As the unified entry point of the system, it is responsible for receiving and classifying various forms of user input.

[0098] Core process:

[0099] ●User Input: The system's starting node, receiving user interaction requests.

[0100] ● Three input type splitting:

[0101] ○ Voice input: Users can interact via voice, supporting natural language dialogue.

[0102] ○Text input: Users input text, including typing, copying and pasting, etc.

[0103] ○ Image Input: Users upload image files such as invoices and receipts.

[0104] Technical features: Intelligent recognition of input type, automatic routing to the corresponding processing module, and support for multimodal fusion input.

[0105] 2. Voice processing module

[0106] Module Function: Converts user voice input into standardized text and performs professional processing in the field of finance and taxation.

[0107] Detailed process:

[0108] ●Speech Recognition API: Integrates speech recognition services, providing basic speech-to-text capabilities.

[0109] ● Dialect model configuration: Supports multiple dialect recognition, including Cantonese, Sichuanese and other local dialects, improving user experience.

[0110] ●Speech-to-text: Converting speech signals into preliminary text content.

[0111] ● Interjection filtering: Intelligently removes colloquial expressions such as "um," "ah," and "that," as well as meaningless words.

[0112] ●Financial and Tax Terminology Matching: Optimized matching for financial and tax terminology to improve the accuracy of domain-specific identification.

[0113] Technical advantages: Supports dialect recognition, professional terminology optimization, and intelligent filtering of interjections, with a speech-to-text accuracy rate of over 95%.

[0114] 3. OCR Recognition Training Module

[0115] Module Functions: Processes image recognition of various invoices and financial documents, extracts key information, and performs intelligent verification.

[0116] Detailed process:

[0117] ●Image preprocessing: Perform preliminary processing on uploaded images, including format conversion, size adjustment, etc.

[0118] ●Deblurring / Correcting / Noise Reduction:

[0119] ○ Deblurring: Using algorithms such as Wiener filtering and blind deconvolution to process blurred images.

[0120] ○ Correction: Tilting angle is detected and corrected using Hough transform.

[0121] ○ Noise Reduction: Use Gaussian filtering and median filtering to remove image noise.

[0122] ●CNN+LSTM+CTC model:

[0123] ○CNN layer: Extracting image features

[0124] ○LSTM layer: Processes sequence information (such as invoice numbers)

[0125] ○CTC layer: Solves alignment issues and outputs recognition results.

[0126] ●Invoice Type Classification: Automatically identifies different types of invoices, including VAT special invoices, general invoices, and electronic invoices.

[0127] ●Key Information Extraction: Extract key elements such as invoice code, invoice number, invoice date, amount, and tax rate.

[0128] ●Result Validation: Perform format and logical validation on the extracted information.

[0129] ○ Verification passed: Information is complete and accurate, and can be directly output to the system.

[0130] ○ Invoices marked as suspicious: Information that fails verification is marked as suspicious.

[0131] ○ When pushing notifications, please specify that the front end will display a message such as "Invoice code recognition may be incorrect, please check."

[0132] Technical features: Supports multiple invoice types, has strong processing capabilities for blurry, tilted, and creased invoices, and has a high recognition accuracy rate.

[0133] 4. Intent Recognition Model Training Module

[0134] Module Functionality: Build and optimize natural language understanding capabilities to achieve accurate intent recognition and entity extraction.

[0135] Training process:

[0136] ●Collect financial and tax business language data: Collect a large amount of financial and tax-related dialogue data and business scenario language data.

[0137] ●Identify key entities: Manually identify key entities such as tax identification number, amount, date, and company name.

[0138] ●Intent type of annotation: Annotate with intent categories such as "Issuing Invoices", "Verifying Authenticity", and "Declaring Taxes".

[0139] ●PyTorch Fine-tuning: Fine-tuning training using the PyTorch framework based on the BERT-base model.

[0140] ● Validation set accuracy evaluation: Evaluate model performance on the validation set.

[0141] ●Accuracy rate meets the standard:

[0142] ○ Not met: Adjust hyperparameters and retrain.

[0143] ○ Compliance: Production Environment Deployment

[0144] Reasoning process:

[0145] ●BERT pre-trained model: Use a pre-trained BERT model for semantic understanding.

[0146] ● Semantic analysis: Analyze the semantics of user input and extract key information.

[0147] ●Business Interface Mapping: Mapping the parsed results to specific business interfaces and operations.

[0148] Technical advantages: Based on the BERT deep learning model, optimized for the financial and tax field, the intent recognition accuracy reaches over 90%.

[0149] 5. Multi-turn dialogue state management module

[0150] Module Function: The core control module of the system manages the entire dialogue process, maintains the dialogue status, and ensures the orderly progress of business processes.

[0151] Status management process:

[0152] ● Finite State Machine Model: Based on the finite state machine design, each business scenario corresponds to an independent state diagram.

[0153] ●State transition diagram: Defines the transition conditions and paths between various states.

[0154] ● Current State Determination: Determines the current state of the dialogue in real time.

[0155] Business status classification:

[0156] ● Awaiting buyer information: Waiting for users to provide the buyer's company information, tax ID, etc.

[0157] ● Awaiting Product Details: Waiting for the user to provide detailed information such as product name, quantity, and unit price.

[0158] ●Verification in progress: The system is verifying the information provided by the user.

[0159] ●Completion: The business process is complete and results can be output.

[0160] Status verification mechanism:

[0161] ●Automatic system validation: Automatically validates user-input information.

[0162] ●Verification result judgment:

[0163] ○ Verification passed: Status progresses, proceeding to the next business step.

[0164] ○ Verification failed: Status rollback, requiring user to re-enter or correct.

[0165] ● Smooth Dialogue Flow: Ensure the entire dialogue proceeds in an orderly manner according to the pre-set business logic.

[0166] Technical features: Supports complex multi-turn dialogues, can remember context, handle multiple user interactions, and ensure the integrity of business processes.

[0167] 6. AI Engine Support Module

[0168] Module Functionality: Provides advanced AI capabilities, including contextual understanding, intelligent recommendation, and risk control.

[0169] Intelligent understanding ability:

[0170] ●Contextual understanding: Understanding dialogue history and handling referential relationships (such as "it" in "What is its tax rate?").

[0171] ●Intent recognition for complex business scenarios: Handling complex business scenarios, such as issuing multiple invoices simultaneously.

[0172] ● Precise intent recognition: Accurately identify the user's true intent in complex conversations.

[0173] Intelligent recommendation capability:

[0174] ● Invoice verification model: Intelligently verifies the compliance of invoice information.

[0175] ● Decision Tree Recommendation: Provides intelligent recommendations based on decision tree algorithms.

[0176] ●Code Recommendation: Recommend suitable product codes

[0177] ●Recommended Tax Rate: Recommended tax rate

[0178] Effect optimization:

[0179] ● Reduced matching error rate: AI optimization significantly reduces the information matching error rate.

[0180] ●Significantly reduced compliance risks: Tax compliance risks are reduced through intelligent verification.

[0181] Technical advantages: It possesses powerful contextual understanding capabilities, enabling it to handle complex business logic and provide intelligent recommendation services.

[0182] 7. Feedback Optimization Module

[0183] Module function: Collect user feedback, continuously optimize system performance, and form a self-learning closed-loop system.

[0184] Optimize the process:

[0185] ●User Feedback: Collect user feedback on system performance, including satisfaction, bug corrections, etc.

[0186] ●Model Optimization: Develop optimization strategies based on the shortcomings of the feedback data analysis system.

[0187] ●Retraining: Retrain the model using new data and feedback.

[0188] Technical features: It forms a closed loop of continuous learning, and the system performance continues to improve as the usage time increases.

[0189] 8. Key relationships between modules

[0190] Service support relationships:

[0191] ● The intent recognition training module provides core semantic understanding capabilities for other modules.

[0192] ● The speech processing module's ability to match financial and tax-related speech data relies on expertise in intent recognition.

[0193] ● OCR recognition verification requires semantic support for intent recognition.

[0194] ● The smooth progress of the dialogue process requires intent recognition to provide accurate intent understanding.

[0195] Data feedback relationship:

[0196] ● All valid dialogue data, OCR-verified data, and domain matching data from speech processing will be fed back into the training module.

[0197] ● Forming a complete closed loop of data collection → model training → application services → data feedback

[0198] System output:

[0199] ● The processing results of all modules ultimately converge at the "System Output / Response" node.

[0200] ●Provide users with a unified, accurate, and timely service response.

[0201] This application's embodiments, by constructing a comprehensive financial and tax AI processing system centered on scenario-based dialogue flows, have significant advantages over existing technologies in multiple dimensions, as detailed below:

[0202] 1. Optimized user experience

[0203] (1) By replacing the traditional menu-style operation with a dialogue flow, users can complete complex business processes without frequently switching pages, effectively reducing complex operation steps; new users are guided through scenarios, which greatly shortens the learning cycle and reduces learning costs.

[0204] (2) The multimodal input component supports mixed interaction of text, voice, and images, adapting to different office scenarios (such as voice input on mobile devices and batch image upload on PCs), significantly improving the flexibility of interaction.

[0205] 2. Improved data processing efficiency

[0206] The front-end recognition information and compliance verification results are displayed in conjunction, avoiding errors from manual association and achieving high data matching accuracy. The OCR recognition module, combined with preprocessing technology, effectively improves the accuracy of fuzzy invoices and reduces the workload of manual verification.

[0207] 3. Deep integration of AI functions

[0208] Multi-turn dialogue processing supports contextual understanding, making intent recognition more accurate in complex business scenarios and overcoming the limitations of single question-and-answer systems; the invoice verification model recommends codes and tax rates through decision trees, reducing the matching error rate and significantly lowering compliance risks.

[0209] 4. Enhanced functional expandability

[0210] (1) The derivative function abstraction mechanism can automatically extract high-frequency operation combinations into independent components, shortening the enterprise personalized function development cycle from weekly to hourly; the custom trigger configuration supports dynamic adjustment of function mode according to business volume, adapting to different scale needs from small and medium-sized enterprises to group enterprises.

[0211] (2) The front-end components support custom styles and rules, and the adaptation cost of the system to different industry financial and tax scenarios (such as high-frequency invoicing of e-commerce companies and input tax verification of manufacturing industries) is reduced by 70%.

[0212] like Figure 2 The diagram shown is a structural schematic of a processing device for financial and tax business requests provided in an embodiment of this application, comprising:

[0213] The receiving module 210 is used to receive financial and tax business requests input by users through text, voice or images, and to perform semantic parsing using a BERT model trained on financial and tax domain corpus to extract user intent and key business entities.

[0214] The management module 220 is used to manage the dialogue flow of business scenarios based on the finite state machine model. According to the current dialogue state, the user intent and the key business entities, it triggers state transitions and orderly promotes the invoicing, verification or diagnosis business processes.

[0215] The key business entities include one or more of the following: tax ID, product name, amount, invoice date, and company name; the finite state machine model defines business states including "waiting for buyer information", "waiting for product details", "in verification" and "completed", and state transitions are triggered by user input or automatic system verification results.

[0216] Correspondingly, the management module 220 is specifically used to fill the parsed key business entities into the business interface parameters corresponding to the current dialogue state. Once the required parameters are filled, the state machine is automatically triggered to advance to the next state.

[0217] The scheduling module 230 is used to schedule the corresponding AI capability modules to perform tasks according to the user intent and the business scenario defined by the key business entity. The AI ​​capability modules use the key business entity as input parameters for processing. The AI ​​capability modules include an OCR model for invoice recognition, a decision tree model for recommending commodity codes and tax rates, and a risk diagnosis model for detecting financial and tax anomalies.

[0218] Specifically, the scheduling module 230 is used to call the decision tree model to match the commodity code and tax rate for invoicing scenarios; to call the OCR model to identify invoice elements and perform format and logic verification for image verification scenarios; and to call the risk diagnosis model based on the isolated forest algorithm to analyze historical data for risk query scenarios.

[0219] Feedback module 240 is used to feed back the execution results of the AI ​​capability module to the front end in real time through server-side push technology, and display them in conjunction with user input information; based on the key business entities and operation combinations recorded in the user operation log, it is automatically abstracted and encapsulated into independent reusable functional components.

[0220] Specifically, the feedback module 240 is used to encapsulate the operation combination into an independent API and generate a corresponding front-end UI component when the frequency of the call to the operation combination of a specific type of key business entity exceeds a preset threshold.

[0221] Furthermore, the aforementioned device also includes:

[0222] The recognition module is used to recognize and preprocess the image input when the image is an invoice image by using an OCR model composed of a convolutional neural network (CNN), a long short-term memory network (LSTM), and a connection-temporal classification (CTC).

[0223] The conversion module is used to convert the financial and tax business requests received by the user via voice input by calling the speech recognition API, filtering out interjections, and optimizing the matching of professional terms using a financial and tax corpus.

[0224] This application embodiment constructs an interaction system centered on scenario-based dialogue flow, which can complete complex financial and tax business processing without frequent page switching, avoids errors caused by users manually associating data, and improves the flexibility of the interaction method, adapting to financial and tax business needs with multi-step, scenario-based dialogue responses.

[0225] This application also provides a computer-readable storage medium storing a computer program. When executed by a processor, this computer program implements the various processes of the above-described method embodiment for processing financial and tax business requests, achieving the same technical effects. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0226] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0227] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0228] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for processing financial and tax business requests, characterized in that, Includes the following steps: It receives financial and tax business requests input by users via text, voice, or images, and performs semantic parsing using a BERT model trained on financial and tax corpus to extract user intent and key business entities. The dialogue flow of the business scenario is managed based on the finite state machine model. According to the current dialogue state, the user intent and the key business entity, the state transition is triggered to orderly advance the invoicing, verification or diagnosis business process. Based on the user intent and the business scenario defined by the key business entity, the corresponding AI capability module is scheduled to execute the task. The AI ​​capability module uses the key business entity as input parameter for processing. The AI ​​capability module includes an OCR model for invoice recognition, a decision tree model for recommending commodity codes and tax rates, and a risk diagnosis model for detecting financial and tax anomalies. The execution results of the AI ​​capability module are fed back to the front end in real time through server-side push technology and displayed in conjunction with user input information; based on the key business entities and operation combinations recorded in the user operation log, they are automatically abstracted and encapsulated into independent reusable functional components.

2. The method according to claim 1, characterized in that, The image is an invoice image. After receiving the user's tax and financial service request input via the image, the system also includes: The image input is identified and preprocessed using an OCR model consisting of a Convolutional Neural Network (CNN), a Long Short-Term Memory Network (LSTM), and a Connection-Time Classification (CTC). After receiving the user's financial and tax service request via voice input, it also includes: By calling the speech recognition API, the financial and tax business requests are converted, interjections are filtered, and professional terminology matching and optimization are performed using a financial and tax corpus.

3. The method according to claim 1, characterized in that, The key business entities include one or more of the following: tax identification number, product name, amount, invoice date, and company name; The finite state machine model defines business states including "waiting for buyer information", "waiting for product details", "verification in progress" and "completed". State transitions are triggered by user input or automatic system verification results. The step of triggering a state transition based on the current dialogue state, the user intent, and the key business entity specifically includes: The parsed key business entities are filled into the business interface parameters corresponding to the current dialogue state. Once the required parameters are filled, the state machine is automatically triggered to advance to the next state.

4. The method according to claim 1, characterized in that, The step of scheduling corresponding AI capability modules to execute tasks based on the user intent and the business scenario defined by the key business entity specifically includes: For invoicing scenarios, a decision tree model is used to match commodity codes and tax rates; for image verification scenarios, an OCR model is used to identify invoice elements and perform format and logic checks; for risk query scenarios, a risk diagnosis model based on the isolated forest algorithm is used to analyze historical data.

5. The method according to claim 1, characterized in that, The key business entities and operation combinations recorded in the user operation logs are automatically abstracted and encapsulated into independent, reusable functional components, specifically including: If the frequency of calls to a combination of operations on a specific type of key business entity exceeds a preset threshold, the combination of operations will be encapsulated into an independent API, and a corresponding front-end UI component will be generated.

6. A processing device for financial and tax business requests, characterized in that, include: The receiving module is used to receive financial and tax business requests input by users through text, voice or images, and to perform semantic parsing using a BERT model trained on financial and tax domain corpus to extract user intent and key business entities. The management module is used to manage the dialogue flow of business scenarios based on a finite state machine model. It triggers state transitions based on the current dialogue state, the user intent, and the key business entities, and orderly advances the invoicing, verification, or diagnosis business processes. The scheduling module is used to schedule the corresponding AI capability modules to perform tasks according to the user intent and the business scenario defined by the key business entity. The AI ​​capability modules use the key business entity as input parameters for processing. The AI ​​capability modules include an OCR model for invoice recognition, a decision tree model for recommending commodity codes and tax rates, and a risk diagnosis model for detecting financial and tax anomalies. The feedback module is used to push the execution results of the AI ​​capability module to the front end in real time through server-side push technology, and display them in conjunction with the user's input information. Based on the key business entities and operation combinations recorded in the user operation logs, they are automatically abstracted and encapsulated into independent, reusable functional components.

7. The apparatus according to claim 6, characterized in that, Also includes: The recognition module is used to recognize and preprocess the image input when the image is an invoice image by using an OCR model composed of a convolutional neural network (CNN), a long short-term memory network (LSTM), and a connection-temporal classification (CTC). The conversion module is used to convert the financial and tax business requests received by the user via voice input by calling the speech recognition API, filtering out interjections, and optimizing the matching of professional terms using a financial and tax corpus.

8. The apparatus according to claim 6, characterized in that, The key business entities include one or more of the following: tax identification number, product name, amount, invoice date, and company name; The finite state machine model defines business states including "waiting for buyer information", "waiting for product details", "verification in progress" and "completed". State transitions are triggered by user input or automatic system verification results. The management module is specifically used to fill the parsed key business entities into the business interface parameters corresponding to the current dialogue state. Once the required parameters are filled, the state machine is automatically triggered to advance to the next state.

9. The apparatus according to claim 6, characterized in that, The scheduling module is specifically used for the following scenarios: for invoicing, it calls a decision tree model to match commodity codes and tax rates; for image verification, it calls an OCR model to identify invoice elements and perform format and logic verification; and for risk query, it calls a risk diagnosis model based on the isolated forest algorithm to analyze historical data.

10. The apparatus according to claim 6, characterized in that, The feedback module is specifically used to encapsulate the operation combination into an independent API and generate a corresponding front-end UI component when the frequency of calling the operation combination of a specific type of key business entity exceeds a preset threshold.

Images