Financial reporting analysis method, apparatus, device, medium, and program product

Abstract

The application provides a financial report analysis method, device, equipment, medium and program product, which can be applied to the fields of artificial intelligence and financial technology, and specifically relates to application of a large model in a financial report analysis scene. The method comprises the following steps: preprocessing financial report data to obtain a to-be-processed object and routing information; performing image-text information recognition on the to-be-processed object through a processing channel corresponding to the to-be-processed object to obtain financial report content data based on document type information corresponding to the to-be-processed object; performing structured analysis on the financial report content data through a pre-trained recognition model to obtain structured data; performing standardized mapping on the structured data to generate a financial indicator table; performing balancing verification and abnormal positioning on the financial indicator table based on a formula configuration rule library to obtain a financial report verification result; and performing analysis on the financial indicator table through a pre-trained large model based on the financial report verification result and context information to obtain a financial report analysis result.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and fintech, specifically to the application of large models in financial statement analysis scenarios, and more specifically to a financial report analysis method, apparatus, equipment, medium, and program product. Background Technology

[0002] With the digitalization of corporate financial reporting, various financial, auditing, and consulting institutions, as well as corporate finance departments, need to parse financial reports in batches to support credit granting, investment, risk control, and operational monitoring. However, financial reports come in various formats and formats, and there are significant differences in report structure, fields, and statistical definitions under different entities, periods, and standards. Data extraction and analysis still heavily rely on manual operations.

[0003] Existing financial statement analysis solutions mainly analyze financial statement data by extracting keywords and matching templates. However, due to the limitations of document formats and reporting standards, the accuracy of recognition for complex financial reports is low. They rely on manual verification of each item, which not only lacks automation but also makes it difficult to quickly identify data errors. Summary of the Invention

[0004] In view of the above problems, embodiments of this application provide a financial reporting analysis method, apparatus, device, medium, and program product.

[0005] According to a first aspect of this application, a financial report analysis method is provided, comprising: preprocessing financial report data to obtain objects to be processed and routing information; wherein, the routing information includes document type information and processing channel corresponding to the objects to be processed; based on the document type information corresponding to the objects to be processed, and through the processing channel corresponding to the objects to be processed, performing graphic and textual information recognition on the objects to be processed to obtain financial report content data; performing structured parsing on the financial report content data through a pre-trained recognition model to obtain structured data; standardizing and mapping the structured data to generate a financial indicator table; performing balance verification and anomaly location on the financial indicator table based on a formula configuration rule base to obtain a financial report verification result; and analyzing the financial indicator table through a pre-trained large model based on the financial report verification result and context information to obtain a financial report analysis result.

[0006] According to embodiments of this application, the document type information includes image type information and / or text type information, the processing channel includes an optical character recognition channel and / or a document parsing channel, and the step of performing image and text information recognition on the object to be processed based on the document type information corresponding to the object to be processed through the processing channel corresponding to the object to be processed includes: performing optical character recognition on the object to be processed based on the image type information through the optical character recognition channel; and / or performing document parsing on the object to be processed based on the text type information through the document parsing channel.

[0007] According to an embodiment of this application, the step of performing optical character recognition on the object to be processed through the optical character recognition channel includes: performing text character recognition on the object to be processed to obtain text content and confidence level; performing table structure recognition on the object to be processed to determine page space information; recognizing and extracting numerical fields in the object to be processed to obtain numerical field content; and integrating the text content, the confidence level, the page space information and / or the numerical field content.

[0008] According to an embodiment of this application, the step of parsing the object to be processed through the document parsing channel includes: performing spreadsheet structure extraction on the object to be processed to obtain spreadsheet data; performing portable document parsing on the object to be processed to obtain document data; performing table identifier semantic extraction on the object to be processed to obtain table annotation information; reconstructing the table structure of the object to be processed to generate table structure data; and integrating the spreadsheet data, the document data, the table annotation information, and / or the table structure data.

[0009] According to an embodiment of this application, the step of balancing and anomaly location of the financial indicator table based on the formula configuration rule base includes: extracting the indicator rules and tolerance thresholds corresponding to the configuration formulas based on the formula configuration rule base; calling a balancing tool to perform indicator validation on the financial indicator table according to the indicator rules and the tolerance thresholds; the indicator validation includes at least one or more of the following: cross-reference validation, cross-source validation, period continuity validation, and tolerance judgment; in the case of indicator validation failure, locating abnormal fields according to the indicator validation results and field traceability information; performing candidate value iterative replacement on abnormal fields with multiple candidate values ​​to balance the configuration formulas; and integrating the indicator validation results, the abnormal fields, and / or the balancing results.

[0010] According to an embodiment of this application, the method further includes: receiving a financial report document; and performing document access processing on the financial report document to obtain the financial report data and the context information, wherein the document access processing includes one or more of basic information detection, file type identification, metadata extraction, and associated document encapsulation.

[0011] According to an embodiment of this application, the financial report data includes standardized financial statement documents and document format types. The preprocessing of the financial report data to obtain objects to be processed and routing information includes: extracting table objects, text objects, and / or page objects from the standardized financial statement documents; performing image preprocessing on the page objects and performing page layout analysis on the preprocessed page objects to obtain region objects; determining the objects to be processed based on the table objects, the text objects, the page objects, and the region objects; and performing format parsing on the objects to be processed based on the document format types to determine the routing information.

[0012] According to an embodiment of this application, the step of analyzing the financial indicator table using a pre-trained large model based on the financial statement verification results and context information includes: quantifying the financial indicator table using multi-dimensional financial features to extract quantitative financial features; performing multi-period time-series fluctuation analysis on the financial indicator table to extract time-series trend features; performing risk analysis on the financial indicator table based on the financial statement verification results to extract risk features; and performing operational status analysis based on the context information, the quantitative financial features, the time-series trend features, and the risk features.

[0013] According to embodiments of this application, the identification model includes a classification model, a first sequence labeling model, a second sequence labeling model, a third sequence labeling model, a table structure parsing model, and / or a neural network model; the step of performing structured parsing of the financial statement content data using the pre-trained identification model includes: identifying the report type of the financial statement content data using the classification model; identifying field labels of the financial statement content data using the first sequence labeling model; performing period alignment and mapping of the financial statement content data based on the context information using the second sequence labeling model; matching numerical correlation elements of the financial statement content data using the third sequence labeling model; deconstructing the table header level of the financial statement content data using the table structure parsing model; and / or performing multi-source information fusion processing of the financial statement content data using the neural network model.

[0014] According to an embodiment of this application, the method further includes: visualizing the financial indicator table and the financial statement analysis results; extracting indicator status information from the financial indicator table and visualizing the indicator status information; extracting abnormal information based on the balancing results and the financial statement verification results and visualizing the abnormal information; and performing serialization encapsulation processing on the financial statement verification results, the financial indicator table, and the financial statement analysis results based on a preset output format to generate interface output data.

[0015] According to a second aspect of this application, a financial report analysis apparatus is provided, comprising: a preprocessing module for preprocessing financial report data to obtain a subject to be processed and routing information; wherein the routing information includes document type information and processing channel corresponding to the subject to be processed; an identification module for recognizing textual and graphical information of the subject to be processed based on the document type information and processing channel corresponding to the subject to be processed, thereby obtaining financial report content data; a parsing module for performing structured parsing of the financial report content data using a pre-trained identification model to obtain structured data; a mapping module for standardizing and mapping the structured data to generate a financial indicator table; a verification module for performing balance verification and anomaly location on the financial indicator table based on a formula configuration rule base, thereby obtaining a financial report verification result; and an analysis module for analyzing the financial indicator table based on the financial report verification result and context information using a pre-trained large model, thereby obtaining a financial report analysis result.

[0016] According to a third aspect of this application, an electronic device is provided, comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0017] According to a fourth aspect of this application, a computer-readable storage medium is also provided, on which a computer program or instructions are stored, wherein the computer program or instructions, when executed by a processor, implement the steps of the above-described method.

[0018] According to a fifth aspect of this application, a computer program product is also provided, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method.

[0019] In the embodiments of this application, financial statements are automatically processed through preprocessing, channel differentiation identification, model structured parsing, standardized mapping, automatic balancing verification, and comprehensive analysis of financial statements using a large model. Different file types are paired with corresponding processing channels, which can adapt to the processing of messy financial statements with multiple carriers, multiple criteria, and multiple formats, significantly improving the accuracy of financial statement identification. Automatic standardization unifies data caliber, and the rule base is configured based on formulas to complete balancing verification and quickly locate data anomalies without the need for manual item-by-item investigation. The degree of automation is improved, and combined with intelligent analysis of contextual fusion using a large model, the reliance on manual labor is reduced, and the analysis efficiency is improved while ensuring data accuracy. Attached Figure Description

[0020] The above-mentioned contents, other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0021] Figure 1 The illustrations depict application scenarios of financial report analysis methods, apparatus, devices, media, and program products according to embodiments of this application.

[0022] Figure 2 A flowchart illustrating a financial report analysis method according to an embodiment of this application is shown schematically.

[0023] Figure 3 This illustration shows a flowchart of the text and image information recognition method for financial report analysis according to an embodiment of this application;

[0024] Figure 4 A flowchart illustrating the intelligent financial statement recognition and analysis method according to an embodiment of this application is shown in the schematic diagram.

[0025] Figure 5 This schematic diagram illustrates a structural block diagram of a financial report analysis apparatus according to an embodiment of the present application;

[0026] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a financial reporting analysis method according to an embodiment of this application. Detailed Implementation

[0027] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0028] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0030] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0031] As used in this paper, the term "model" refers to a model that learns the relationship between inputs and outputs from training data, enabling it to generate corresponding outputs for a given input after training. Model generation can be based on machine learning techniques. Deep learning is a machine learning algorithm that processes inputs and provides corresponding outputs using multiple layers of processing units. A neural network model is an example of a deep learning-based model. In this paper, "model" may also be referred to as a "machine learning model," "learning model," "machine learning network," or "learning network," and these terms are used interchangeably.

[0032] It's important to note that the term "neural network" can refer to a machine learning network based on deep learning. A neural network processes input and provides corresponding output, typically consisting of an input layer, an output layer, and one or more hidden layers between them. Neural networks used in deep learning applications often include many hidden layers, increasing the network's depth. The layers of a neural network are connected sequentially, so that the output of the previous layer serves as the input to the next layer. The input layer receives the input to the neural network, while the output layer's output becomes the final output. Each layer of a neural network includes one or more nodes (also called processing nodes or neurons), each processing the input from the layer above.

[0033] It should be understood that machine learning generally includes three phases: training, testing, and application (also known as inference). In the training phase, a given model is trained using a large amount of training data, iteratively updating parameter values ​​until the model can consistently generate inferences that meet the expected goals from the training data. Through training, the model can be considered to have learned the relationship between inputs and outputs (also known as the input-output mapping) from the training data. The parameter values ​​of the trained model are determined. In the testing phase, test inputs are applied to the trained model to test whether it can provide the correct output, thus determining the model's performance. In the application phase, the model can be used to process actual inputs based on the trained parameter values ​​to determine the corresponding output.

[0034] In one or more embodiments described herein, the term "large model" can refer to a deep learning model with a large number of model parameters, which can include hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of model parameters. Large models can also be called foundational models or basic models. They are pre-trained using large-scale unlabeled corpora to produce pre-trained models with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability, such as large language models and multimodal pre-trained models. It should be understood that in practical applications, large models only require a small number of samples to fine-tune the pre-trained model before being applied to different tasks. Large models can be widely used in natural language processing, computer vision, and other fields. Specifically, they can be applied to computer vision tasks such as visual question answering, image captioning, and image generation, as well as natural language processing tasks such as text-based sentiment classification, text summarization, and machine translation. Major application scenarios for large models can include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0035] As corporate financial reporting becomes increasingly digitized, financial institutions, investment firms, auditing firms, consulting firms, and internal financial management departments all require the rapid parsing and analysis of large volumes of financial reports to support business scenarios such as credit approval, investment judgment, risk assessment, peer benchmarking, operational monitoring, and automated report entry. In practice, the sources of financial reports are complex, including scanned documents, image-based reports, structured electronic ledgers, exported natively parsable documents, and mixed documents with various formatting options. Significant differences in report structure, field naming, unit definitions, and disclosure methods across different companies, periods, and accounting standards have led to a long-standing reliance on manual data entry, verification, and interpretation for financial statement data extraction and analysis.

[0036] Existing technologies typically employ general character recognition techniques, template-based parsing techniques, or single text extraction schemes to process financial reports. While these solutions can achieve partial automation in highly standardized scenarios, they still have significant limitations when dealing with complex layouts, mixed multi-table displays, heterogeneous field names, complex interrelationships, and the need for semantic understanding of report notes. Especially when financial data needs to be further used for operational status analysis, inaccurate recognition results or a lack of verifiable mechanisms will directly affect the reliability of subsequent analytical conclusions. Specific shortcomings are as follows:

[0037] The existing solution is not accurate enough and has poor robustness under complex financial statement formats: it is not adaptable to image-type portable documents, cross-page tables, merged cells, nested header levels, different field names, and multiple display methods of the same indicator. It is prone to field misalignment, missing amount recognition, missing unit, incorrect recognition of positive and negative signs, and incorrect period mapping, resulting in unstable structured results.

[0038] The lack of a verification mechanism oriented towards financial logic makes it difficult to detect errors in a timely manner: Traditional Optical Character Recognition (OCR) or table extraction technologies typically only focus on the recognized content and do not verify whether the recognition results satisfy financial logic relationships. When there are inconsistencies in the reconciliation relationships between the balance sheet, income statement, and cash flow statement, existing systems struggle to automatically identify the source of the problem and still require manual item-by-item verification, resulting in high verification costs.

[0039] The disconnect between financial statement recognition and financial analysis leads to insufficient reliability of analysis results: Existing solutions often directly input recognition results that have not been fully verified into the analysis model. The analysis model cannot distinguish whether the original data is reliable, nor can it combine information such as recognition confidence, source location, and verification anomalies to constrain the conclusions, thus leading to biases in the judgment of business status.

[0040] Insufficient scalability and maintainability make it difficult to adapt to scenarios with multiple industries, templates, and rules: When adding new industry reports, changing field names, adjusting indicator definitions, or adding new verification logic, solutions based on fixed templates or hard-coded rules require frequent modifications to the program logic, resulting in high expansion costs and making them unsuitable for continuously iterating financial statement recognition business scenarios.

[0041] This application provides a financial report analysis method, which preprocesses financial report data to obtain objects to be processed and routing information. The routing information includes document type information and processing channels corresponding to the objects to be processed. Based on the document type information, the method uses the processing channels to identify the text and image information of the objects to be processed, obtaining financial report content data. A pre-trained recognition model is used to perform structured parsing of the financial report content data to obtain structured data. The structured data is then standardized and mapped to generate a financial indicator table. A formula-configured rule base is used to perform balancing verification and anomaly location on the financial indicator table, obtaining financial report verification results. Finally, based on the financial report verification results and context information, a pre-trained large model is used to analyze the financial indicator table, obtaining financial report analysis results. In the embodiments of this application, financial statements are automatically processed through preprocessing, channel differentiation identification, model structured parsing, standardized mapping, automatic balancing verification, and comprehensive analysis of financial statements using a large model. Different file types are paired with corresponding processing channels, which can adapt to the processing of messy financial statements with multiple carriers, multiple criteria, and multiple formats, significantly improving the accuracy of financial statement identification. Automatic standardization unifies data caliber, and the rule base is configured based on formulas to complete balancing verification and quickly locate data anomalies without the need for manual item-by-item investigation. The degree of automation is improved, and combined with intelligent analysis of contextual fusion using a large model, the reliance on manual labor is reduced, and the analysis efficiency is improved while ensuring data accuracy.

[0042] It should be noted that the financial report analysis method and apparatus of this application can be used in the fields of artificial intelligence and fintech, specifically involving the application of large models in financial report analysis scenarios, and can also be used in any field other than artificial intelligence and fintech. The application fields of the financial report analysis method and apparatus of this application are not limited.

[0043] In the technical solution of this application, the enterprise information (including but not limited to basic enterprise information, enterprise industry information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved are all information and data authorized by the enterprise or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure and application of related data all comply with relevant laws, regulations and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding operation portals are provided for enterprises to choose to authorize or refuse.

[0044] In scenarios where enterprise information is used for automated decision-making, the methods, devices, and systems provided in this application all provide users with corresponding operation entry points, allowing users to choose to agree to or reject the automated decision results; if the user chooses to reject, the process proceeds to the expert decision-making process.

[0045] Figure 1The illustrations depict application scenarios of financial reporting analysis methods, apparatus, devices, media, and program products according to embodiments of this application.

[0046] like Figure 1 As shown, application scenario 100 according to an embodiment of this application may include a first terminal device 101, a second terminal device 102, a database 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the database 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables. For example, a user can use the first terminal device 101 and the second terminal device 102 to interact with the server 105 and / or the database 103 through the network 104 to receive or send information, etc.

[0047] The first terminal device 101 and the second terminal device 102 can be electronic devices such as smartphones, wearable devices, personal computers, intelligent voice interaction devices, smart home appliances, intelligent vehicles, in-vehicle terminals, aircraft, unmanned vending terminals, and extended reality devices. Extended reality devices can include virtual reality devices, augmented reality devices, and mixed reality devices. A client application for the target application can be installed and run on the terminal device. This target application can include, but is not limited to, financial transaction applications, payment applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, and social media platform software (these are just examples). Furthermore, this application embodiment does not limit the form of the target application, including but not limited to applications, mini-programs, etc., installed on the terminal device, and can also be in web page form.

[0048] Server 105 can be a server providing various services, such as a backend management server supporting websites browsed by users using the first terminal device 101 and the second terminal device 102 (this is just an example). The backend management server can analyze and process received user requests and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and big data. The server can be a backend server for the aforementioned target application, used to provide backend services to the clients of the target application.

[0049] Database 103 is a professional storage system for storing and managing data. It can store various types of data related to the target application, such as user account information, business transaction records, content resource data, configuration formulas, and field mapping relationships. It supports structured, semi-structured, or unstructured data storage and has management capabilities such as data creation, deletion, modification, querying, backup, and recovery. In this application scenario, database 103 can be connected to server 105 via a communication link. Server 104 can retrieve the required data from database 103 for processing based on requests from the first terminal device 101 and the second terminal device 102. It can also synchronously store new data generated by the operations of the first terminal device 101 and the second terminal device 102 into database 103, thereby achieving data persistence and efficient retrieval.

[0050] It should be noted that the financial report analysis method provided in this application embodiment can generally be executed by server 105 and / or terminal devices 101-102. Accordingly, the financial report analysis device provided in this application embodiment can generally be installed in server 105 and / or terminal devices 101-102.

[0051] It should be understood that Figure 1 The number of terminal devices, networks, databases, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, databases, and servers can be included.

[0052] Figure 2 A flowchart illustrating a financial report analysis method according to an embodiment of this application is shown schematically. Figure 2 As shown, the financial report analysis method 200 according to the embodiments of this application may include steps S210 to S260.

[0053] In step S210, the financial report data is preprocessed to obtain the objects to be processed and routing information; wherein, the routing information includes the document type information and processing channel corresponding to the objects to be processed.

[0054] Preprocessing financial report data is a pre-processing procedure that involves standardized cleaning and optimization of raw data from tables, text, and page images in standardized financial report documents, providing a high-quality foundation for graphic and textual information recognition.

[0055] The objects to be processed are various independently identifiable and parseable basic units extracted from standardized financial statement documents, including page objects, area objects, table objects, and text objects generated from layout analysis, covering different forms of financial statement carrier units such as scanned image pages, image pages, native portable document format pages, portable document format workbooks, report tables, and financial statement text fragments.

[0056] The routing information is used to distribute the scheduling identification information of the objects to be processed. It consists of two parts: first, document type information, which distinguishes between image documents and text documents and carries corresponding feature parameters such as layout, resolution, and original structure; second, the matching processing channel, which guides the system to select the corresponding recognition method to process the current object to be processed.

[0057] Document type information includes image type information and text type information. Image type information includes image source identifiers such as scanned documents, screenshots, and photos; image resolution; page layout; OCR recognition attributes; and whether the image contains tables, among other image document feature parameters. Text type information includes document carrier identifiers such as native portable document formats, spreadsheet file formats, and plain text; native document object attributes; electronic document text layout; and native table structure, among other machine-readable document feature parameters.

[0058] The processing channels are adapted to dedicated recognition processing links for different document types, divided into optical character recognition channels (for scanned documents, images, and other image-based financial reports) and document parsing channels (for electronic documents that can directly read native data, such as native portable document formats and spreadsheet file formats). The corresponding channel is automatically allocated according to the routing information to complete the image and text recognition.

[0059] In step S220, based on the document type information corresponding to the object to be processed, the image and text information of the object to be processed is recognized through the processing channel corresponding to the object to be processed to obtain the financial report content data.

[0060] The system obtains the document type information of the object to be processed, determining whether it is a scanned document, a native document, or mixed content, and then matches the corresponding processing channel accordingly. Different processing channels can be selected for different pages or regions. Scanned image objects are assigned to the OCR channel, native text / parsable tables are assigned to the document parsing channel, and mixed content is split by region and processed in parallel using dual channels.

[0061] The OCR channel performs layout analysis, character detection and recognition on the scanned area, outputting character recognition results such as text content, confidence level, and numerical field content. The document parsing channel extracts text from raw text and structured tables, parses table structure, restores row and column relationships, and generates document parsing results such as spreadsheet data and document data.

[0062] The dual-channel processing results are merged, deduplicated, and verified according to regional location and hierarchy, and integrated into unified financial statement content data, including character recognition results and / or document parsing results, to achieve accurate extraction and structured output of financial statement graphic information.

[0063] When both OCR and document parsing results are obtained for the same field, the machine-readable parsing result is used first. When the parsing result is missing or has low confidence, it automatically reverts to the OCR result. When both exist but are inconsistent, the balancing tool is called to determine consistency.

[0064] In step S230, the financial statement content data is parsed in a structured manner using a pre-trained recognition model to obtain structured data.

[0065] A recognition model refers to a lightweight machine learning model deployed at the front end or middle layer of the financial reporting analysis process. It is used to automatically identify and classify field names, indicator items, amount values, period information, currency information, unit information, and inter-table relationships in financial statements. Multiple recognition models, such as lightweight neural network models, classification models, sequence labeling models, and table structure parsing models, can be flexibly selected to collaboratively complete structured parsing.

[0066] Structured data is financial data objects that have been identified, extracted, and normalized for further processing by programs. It includes field labels, amount values, time dimensions, units, currencies, source page numbers, source coordinates, confidence levels, verification results, and modification records.

[0067] In step S240, the structured data is standardized and mapped to generate a financial indicator table.

[0068] A financial indicator table refers to a structured result set formed by standardizing, mapping, and cleaning the financial statement fields of structured data. The structured result set can be stored and displayed in the form of tables, field trees, report objects, or database records, including information such as indicator name, value, period, unit, source location, identification confidence level, and verification status.

[0069] The structured data is standardized and mapped to generate a financial indicator table. Specifically: First, based on a field mapping library (such as a synonym mapping library), the names of various differentiated fields in the initial structured financial statement data are mapped, uniformly matching fields with different textual descriptions to their corresponding standard indicator codes. Second, numerical units are normalized, converting values ​​with different units of measurement such as yuan, thousand yuan, ten thousand yuan, and million yuan to a unified standard unit. Finally, standardized tags are added to the currency information within the data, completing the unification of the currency dimension.

[0070] Then, the period dimension is uniformly represented, such as the end of 2025, 2025-12-31, the end of this reporting period, etc., and various time descriptions are uniformly mapped to standardized time identifiers; then, a standard financial indicator table is generated, and the relationship between indicators and source document page numbers, coordinates, confidence levels, candidate values, and report types is established.

[0071] It can also automatically generate derived financial indicators such as gross profit margin, debt-to-equity ratio, current ratio, net profit margin, and net cash flow from operating activities based on standardized basic financial indicators, expanding the dimensions of indicator analysis and enhancing the completeness of subsequent analysis.

[0072] It is worth noting that different field mapping libraries and formula rule libraries are set up for different industries such as manufacturing, real estate, finance, and retail to improve the recognition and analysis of reports for specific industries.

[0073] In step S250, based on the formula configuration rule base, the financial indicator table is balanced and anomaly located to obtain the financial statement verification result.

[0074] The formula configuration rule library is the core repository for centrally storing and managing all rules for configuring formulas (such as indicator codes, calculation logic, and applicable conditions). It allows users to flexibly define, modify, and version-manage these rules through a graphical interface or through scripts by technical personnel. The formula configuration rule library extracts and standardizes complex calculation logic from the code, enabling the system to dynamically and on-demand retrieve rules for automatic calculations, thus improving the flexibility, accuracy, and maintainability of financial or business indicator calculations.

[0075] Configuration formulas are stored in the formula configuration rule base. Configuration formulas refer to the rule expressions used to define the logical relationships, operational relationships, period relationships and tolerance conditions between financial indicators. Configuration formulas include, but are not limited to: Total Assets = Total Liabilities + Total Owner's Equity; Ending Cash and Cash Equivalents Balance = Beginning Cash and Cash Equivalents Balance + Net Increase in Cash and Cash Equivalents; Derivation relationships between operating revenue, operating costs, gross profit and net profit; Cross-consistency relationships of the same indicator in different reports, notes and periods.

[0076] The financial indicator table is balanced and anomaly located using configuration formulas to obtain financial statement verification results. The financial statement verification results include the verified financial indicator table, a list of abnormal fields, balance status, error sources, and correction suggestions.

[0077] In step S260, based on the financial statement verification results and context information, the financial indicator table is analyzed using a pre-trained large model to obtain the financial statement analysis results.

[0078] Large models refer to large language models (LLMs) based on large-scale pre-training. They conduct comprehensive and in-depth analysis based on multi-dimensional data such as structured financial indicators, financial statement text fragments, report notes, and verification results. This enables them to achieve in-depth understanding of financial statement semantics, assessment of enterprise operating conditions, identification of financial risks, and automatic generation of complete analysis conclusions.

[0079] Financial statement analysis results refer to the enterprise operation evaluation results output by the large model analysis engine based on structured financial statement data, including but not limited to growth, profitability, solvency, cash flow status, operational efficiency, risk warnings, and comprehensive business conclusions.

[0080] The system can also process financial statements from multiple periods simultaneously, generating cross-period financial indicator tables, and the large model analysis engine performs year-on-year, month-on-month, and trend judgments.

[0081] The system inputs the verified financial indicators, anomaly markers, identification confidence levels, summary notes, basic company information, and industry information into a pre-trained large model. Through inference and analysis by the pre-trained large model, it outputs financial statement analysis results, including business status analysis results, risk warning information, analysis summaries, and visual explanatory text.

[0082] For example, the training process for a large model is as follows:

[0083] Dataset Construction and Labeling: Collect massive amounts of standardized financial statement indicator tables, multi-period financial time-series data, enterprise industry information, financial statement notes, historical verification anomaly records, and manual financial analysis reports; clean and deduplicate the data, divide the sample set, label it with financial feature labels, trend labels, risk labels, and standard analysis conclusions, construct a training corpus specific to the financial field, and distinguish between normal indicator samples, low-confidence anomaly samples, and cross-period comparison samples.

[0084] Domain-specific fine-tuning of data preprocessing: unifying the coding, unit, and time dimension formats of financial indicators; converting structured financial indicators, verification anomaly markers, confidence levels, industry context, company context, and footnote text into input sequences that the model can recognize; distinguishing between structured numerical features and natural language text features; adding financial-specific prompt templates; and marking reconciliation anomalies and identifying risk indicators such as low confidence levels.

[0085] Model fine-tuning training: Based on a general-purpose large language model, domain-specific fine-tuning is carried out, setting a multi-task loss function, and simultaneously learning three types of tasks: multi-dimensional financial indicator quantitative calculation, multi-period time series trend mining, and financial risk identification; the input includes financial indicator tables, validation anomalies, industry context, and company context, and the supervision signals use standard financial analysis conclusions and risk warning texts written by humans, iteratively optimizing the model's ability to understand, calculate, and summarize financial data; special samples are added to strengthen training for scenarios such as low confidence and field anomalies in OCR recognition, reducing model illusions.

[0086] Validation and Model Tuning: The test set is used to validate the large model's performance metrics, including accuracy in trend prediction and recall in risk identification. Hyperparameters such as learning rate, prompt word templates, and context window size are adjusted to optimize output consistency. Once all metrics meet the targets, the weights are saved, resulting in a dedicated large model for financial statement analysis.

[0087] In the embodiments of this application, financial statements are automatically processed through preprocessing, channel differentiation identification, model structured parsing, standardized mapping, automatic balancing verification, and comprehensive analysis of financial statements using a large model. Different file types are paired with corresponding processing channels, which can adapt to the processing of messy financial statements with multiple carriers, multiple criteria, and multiple formats, significantly improving the accuracy of financial statement identification. Automatic standardization unifies data caliber, and the rule base is configured based on formulas to complete balancing verification and quickly locate data anomalies without the need for manual item-by-item investigation. The degree of automation is improved, and combined with intelligent analysis of contextual fusion using a large model, the reliance on manual labor is reduced, and the analysis efficiency is improved while ensuring data accuracy.

[0088] According to an embodiment of this application, before preprocessing the financial report data, the financial report analysis method further includes: receiving a financial report document; and performing document access processing on the financial report document to obtain financial report data and context information, wherein the document access processing includes one or more of the following: basic information detection, file type identification, metadata extraction, and associated document encapsulation.

[0089] Financial reporting documents can be formatted in various ways, including portable document formats, spreadsheet formats, image report formats, and hybrid formats. Portable document formats are further divided into scanned image formats and natively parsable formats. Portable document formats, also known as portable document formats, are standardized electronic document storage formats that are independent of devices, software, and operating systems, and can fully preserve text, charts, and layout. Spreadsheet file formats are specialized structured file storage formats that use a cell-row-column structure to store numerical values, formulas, and tabular data, supporting data calculation and statistical classification. Image report formats refer to image storage formats that scan or convert complete financial statements into image carriers, retaining only the visual layout and not allowing direct extraction of native tabular data.

[0090] Financial reporting documents refer to electronic documents containing corporate financial data, financial statements and related notes, including but not limited to annual reports, semi-annual reports, quarterly reports, audit reports, management reports and other financial documents in portable document formats, spreadsheet file formats, images or combinations thereof.

[0091] Users can upload financial report documents and selectively input information such as company name, basic company information, reporting period, industry category, industry information, currency, and unit.

[0092] The system receives financial report documents submitted by users. Document access processing involves basic information checks (such as basic legality verification and document integrity checks), document type identification, and metadata extraction and registration. For multi-file scenarios, associated document encapsulation is performed, allowing multiple files to be archived, associated, and batch-identified, generating standardized documents that can be uniformly parsed for subsequent analysis and processing.

[0093] After document access processing, financial report data and user context information (such as company name, industry category, financial reporting period, etc.) are obtained. The financial report data includes standardized financial report document objects, file identifiers, document format types, page number information, upload time, batch to which they belong, etc.

[0094] In the embodiments of this application, document access processing is used to complete file detection, type identification, metadata extraction and multi-document association encapsulation, unify the original input standards of various financial reports, improve the standardization of financial original data from the source, collect basic document metadata in advance, provide context information for subsequent processing, avoid information fragmentation, reduce the workload of manual sorting, entry and archiving, and improve the efficiency of financial report preprocessing.

[0095] According to an embodiment of this application, after obtaining the financial statement analysis results, the financial statement analysis method further includes: visualizing the financial indicator table and the financial statement analysis results; extracting the indicator status information of the financial indicator table and visualizing the indicator status information; extracting abnormal information based on the balancing results and the financial statement verification results and visualizing the abnormal information; and performing serialization encapsulation processing on the financial statement verification results, the financial indicator table and the financial statement analysis results based on a preset output format to generate interface output data.

[0096] The process involves visualizing and rendering standardized financial indicator tables, rendering them in a structured format to generate a visual display view. Simultaneously, the operational status analysis text, core qualitative conclusions, risk warning labels, and other financial statement analysis results are structured and formatted in a layered manner to generate a readable analysis report and visualization view, thus completing the visual presentation of the financial indicator tables and financial statement analysis results.

[0097] The financial indicator table is processed by associating and mapping the source verification information of the indicators. The indicator status information corresponding to each financial indicator is extracted, including the source location of the indicator, the recognition confidence level, and the verification status. Then, the original recognition source coordinates, model recognition confidence level values, and indicator verification status identifiers corresponding to each financial indicator are bound and mapped to the corresponding indicator items. The source information is loaded onto the visualization page to realize the visualization display of indicator status information.

[0098] Based on the balancing results and financial statement verification results, anomaly visualization and annotation processing is performed. Combining the input balancing results, all anomaly information such as formulas that passed or failed balancing and abnormal fields are extracted. All anomaly information is highlighted and visualized in different areas to complete the visualization of anomaly information.

[0099] Based on preset output formats (such as tables, structured data, report documents, interface messages, etc.), the financial statement verification results, standardized financial indicator tables, and operating status financial statement analysis results are serialized and converted, and packaged into table files, structured data, standardized report documents, system interface response messages, and other interface output data for third-party systems. Display data for the front-end interface can be generated simultaneously.

[0100] The system can be deployed as a software-as-a-service platform, a private enterprise service, or a service interface embedded in approval / risk control / investment research systems, which can be called by upstream business systems to complete the automatic identification and analysis of financial statements.

[0101] By constructing an integrated processing workflow encompassing identification, structuring, verification, error correction, analysis, and presentation, the system reduces manual data entry and review, thereby improving financial statement processing efficiency. Through the adoption of a modular identification architecture, standardized field mapping mechanism, and formula configuration mechanism, the system can be rapidly expanded to accommodate different industries, report templates, and financial rules, reducing maintenance costs.

[0102] In the embodiments of this application, indicator tables, analysis results, indicator status, and anomaly information are presented visually, intuitively displaying data issues. Simultaneously, financial report results are uniformly serialized and encapsulated for external output. This facilitates rapid tracing and verification of abnormal data, reduces manual review costs, and the standardized interface allows seamless integration with third-party systems such as risk control, auditing, and business management, adapting to automated integration across multiple business scenarios.

[0103] According to the embodiments of this application, after obtaining the financial statement analysis results, the financial statement analysis method further includes: receiving error correction feedback data, historical recognition results, verification failure logs and manual revision records, and performing continuous optimization processing based on the abnormal analysis data. Specifically, this includes updating the field synonym mapping library, updating the industry adaptation rules and formula rule library, generating training samples to optimize the lightweight recognition model, and statistically analyzing the recognition performance under different templates, different enterprises, and different report types, providing continuous optimization data support for the lightweight recognition model and balance verification.

[0104] According to embodiments of this application, financial reporting data includes standardized financial statement documents and document format types. Preprocessing the financial reporting data to obtain objects to be processed and routing information includes: extracting table objects, text objects, and / or page objects from the standardized financial statement documents; performing image preprocessing on the page objects and performing page layout analysis on the preprocessed page objects to obtain region objects; determining the objects to be processed based on the table objects, text objects, page objects, and region objects; and performing format parsing on the objects to be processed based on the document format type to determine routing information.

[0105] Standardized financial statement documents are standardized financial report documents that have undergone document access processing. They have a unified packaging standard structure and include supporting information such as file identifiers, batches, format types, and user contexts. They can be directly transferred to preprocessing, dual-channel image and text recognition, and other stages.

[0106] Document format types are used to classify documents based on their underlying storage and content delivery methods. Document format types include image format types and text format types. Image format types include scanned areas, image areas, and image-based portable document format pages, which are represented by pixel images and lack natively editable text layers. Text format types include natively portable document format pages, spreadsheet workbooks, tables, and text objects, with built-in structured and parsable text and table data, and native text layers.

[0107] Preprocessing operations are performed based on the document format type and page content characteristics, as follows:

[0108] From the financial report data, we extract the basic carrier elements such as table objects, text objects, and page objects that are inherent in the standardized financial statement documents.

[0109] Image preprocessing operations such as page sharpness detection, rotation correction, noise reduction, and binarization are performed on all page objects. Then, layout analysis is performed on the processed page objects to distinguish between text areas, table areas, image areas, and annotation areas. Table objects, text objects, page objects, and area objects are summarized to determine the objects to be processed, covering scanned areas, image areas, image-based portable document format pages, native portable document format pages, spreadsheet file format workbooks, tables, and text.

[0110] Documents are categorized based on their format type, determining whether they are native portable document formats, scanned portable document formats, spreadsheets, or mixed documents. Documents are classified as image or text, and document type information including image and text layouts is generated. Format parsing is performed using various document feature parameters to extract the document type information corresponding to the object to be processed, matching the corresponding OCR channel or document parsing channel, and finally generating routing information containing document type information and the corresponding processing channel.

[0111] In the embodiments of this application, multiple types of objects, such as tables, text, and pages, are split in the preprocessing stage. The layout of the image page is analyzed to generate region objects. The object to be processed and the corresponding routing channel are determined in combination with the file format. This enables fine-grained segmentation and distribution of financial report content, matching and adapting processing links for different regions, avoiding invalid calculations caused by uniform processing of the entire page, accurately distinguishing image and text blocks, providing accurate input objects for dual-channel recognition, improving computational efficiency, and adapting to the processing of complex financial reports with messy structures and mixed image and text layouts.

[0112] Figure 3 A flowchart illustrating the graphic information recognition process of a financial report analysis method according to an embodiment of this application is shown. Figure 3 As shown in the embodiments of this application, the document type information includes image type information and / or text type information, and the processing channel includes an optical character recognition channel and / or a document parsing channel. Based on the document type information corresponding to the object to be processed, the image and text information of the object to be processed is recognized through the processing channel corresponding to the object to be processed, which may include steps S310 to S320.

[0113] In step S310, based on image type information, optical character recognition is performed on the object to be processed through the optical character recognition channel.

[0114] According to an embodiment of this application, optical character recognition is performed on the object to be processed through an optical character recognition channel, including: performing text character recognition on the object to be processed to obtain text content and confidence level; performing table structure recognition on the object to be processed to determine page space information; recognizing and extracting numerical fields in the object to be processed to obtain numerical field content; and integrating text content, confidence level, page space information and / or numerical field content.

[0115] Optical character recognition channels are used to convert character images in scanned documents and image-based portable document formats into computer-processable text information using OCR technology.

[0116] The optical character recognition (OCR) channel performs OCR on the object to be processed. First, text character recognition is performed, detecting and recognizing the text at the character or line level, and outputting the text content and its corresponding confidence score. The confidence score is a rating of the degree of trustworthiness given for the recognition result of a specific field, which can be used for subsequent priority ranking, anomaly filtering, and triggering manual review.

[0117] The system identifies table structures, performs frame detection or frameless table structure recovery on tables in images, and determines page layout information, including coordinates, table structure, row and column relationships, and page position. It also identifies numerical fields such as amount, percentage, date, unit, and currency, and extracts the corresponding values ​​to obtain the content of these numerical fields.

[0118] The text content, confidence score, page layout information, and / or numerical field content are normalized and aggregated, and the page number, coordinates, row and column relationship, and confidence score of the text block are recorded. The output includes character recognition results containing text content, numerical field content, coordinate information, table structure, confidence score, and page layout information.

[0119] The OCR channel can also output three-level structured results: text blocks, cell blocks, and table blocks, which facilitates subsequent field classification and indicator positioning.

[0120] In the embodiments of this application, the OCR channel performs multi-dimensional recognition of text, tables, and numerical values ​​in a layered manner and integrates multi-dimensional information. It not only extracts text and recognition confidence, but also accurately captures the spatial position of the page and the content of numerical fields. Based on the page information, different table blocks are distinguished, reducing the probability of field misalignment and incorrect amount recognition. The confidence data can mark low-reliability recognition content, providing a traceability basis for the verification process, and significantly improving the accuracy of structured extraction of scanned financial statements.

[0121] In step S320, based on the text type information, the document to be processed is parsed through the document parsing channel.

[0122] The document parsing channel is used to perform layout analysis, table recognition, cell extraction, field location, and content structuring on machine-readable spreadsheet files, native portable documents, or hybrid documents to extract document data.

[0123] In the embodiments of this application, image and text document types are distinguished, and differentiated image and text recognition is achieved by adapting dual independent processing channels of OCR and document parsing. An OCR channel is enabled for scanned financial reports and other image-based financial statements, while a document parsing channel is enabled for documents such as spreadsheet files and natively portable document formats. This ensures uniform compatibility with multiple mixed-format financial reports, solves the problem of poor adaptability of a single recognition link, improves the recognition accuracy of complex reports, avoids information loss caused by unified processing, and enhances the robustness of field, amount, and period recognition from the source, reducing the workload of manual pre-classification.

[0124] According to embodiments of this application, document parsing of the object to be processed is performed through a document parsing channel, including: performing spreadsheet structure extraction on the object to be processed to obtain spreadsheet data; performing portable document parsing on the object to be processed to obtain document data; extracting table identifier semantics from the object to be processed to obtain table annotation information; reconstructing the table structure of the object to be processed to generate table structure data; and integrating spreadsheet data, document data, table annotation information and / or table structure data.

[0125] The document parsing channel parses the document to be processed. First, the document to be processed is extracted in a spreadsheet structure, extracting spreadsheet data such as document location index, worksheet name, cell coordinates, merged cell relationships, table content, formula results and style features, while obtaining cell formulas and display values.

[0126] The process performs portable document parsing on the object to be processed, extracting document data such as text objects, table objects, document position indexes, paragraph objects and their relative layout relationships in the native portable document format.

[0127] The table identifier semantics of the objects to be processed are extracted to identify table annotation information such as table header level, column headings, row headings, annotation information, unit declarations, unit information and period labels.

[0128] The table structure of the object to be processed is reconstructed, the row and column structure of the table is reconstructed, a computable table representation is formed, and table structure data is generated.

[0129] By integrating the above processing results, the spreadsheet data, document data, table annotation information, and / or table structure data are normalized and aggregated to obtain the document parsing results.

[0130] In the embodiments of this application, the document parsing channel specifically analyzes mainstream electronic financial reports such as spreadsheet file formats and portable document formats, simultaneously extracts semantic annotations from the tables and reconstructs the complete table structure, integrating multi-dimensional original data from electronic documents. It can directly read native spreadsheet values, avoiding OCR translation distortion issues, reconstructing and restoring multi-level header row and column relationships, adapting to different enterprise-customized report templates, unifying the electronic document data output format, significantly improving structured extraction efficiency, and adapting to batch processing of multi-format electronic financial reports.

[0131] According to embodiments of this application, the identification model includes a classification model, a first sequence labeling model, a second sequence labeling model, a third sequence labeling model, a table structure parsing model, and / or a neural network model. The pre-trained identification model performs structured parsing of financial statement content data, including: identifying report types using the classification model; identifying field labels using the first sequence labeling model; performing period alignment and mapping on the financial statement content data based on contextual information using the second sequence labeling model; matching numerical correlation elements using the third sequence labeling model; deconstructing the table header hierarchy using the table structure parsing model; and / or performing multi-source information fusion processing on the financial statement content data using the neural network model.

[0132] The report type is identified using a classification model, determining whether the current table is a balance sheet, income statement, cash flow statement, statement of changes in equity, or notes to the financial statements.

[0133] The first sequence labeling model is used to identify field labels and uniformly identify semantically identical but different named fields such as operating revenue, total operating revenue, and main business revenue.

[0134] Based on contextual information such as company name, industry category, and financial reporting period, the second sequence labeling model is used to perform period alignment and mapping, identify period labels such as current period, previous period, year-to-date, and same period of previous year, and establish a binding relationship between numerical values ​​and corresponding periods.

[0135] The third sequence labeling model is used to match numerical association elements, complete the determination of the amount attribution, and identify the fields, units, currencies and positive and negative attributes corresponding to each amount.

[0136] By using a table structure parsing model to deconstruct the table header hierarchy, structural interference caused by multi-level headers, cross-column headings, and cross-page continuation of tables can be eliminated.

[0137] By using a neural network model to perform multi-source information fusion processing, when OCR results and document parsing results coexist, the optimal result or candidate result set is selected based on confidence level, source type, field context, and rule matching degree. The final output is initial structured financial report data containing a list of candidate values ​​for each field, the primary value, confidence level, and source location information.

[0138] It's worth noting that the classification model refers to a dedicated classification model for report types. It takes financial statement tables and related images as input and outputs report category labels. Its core function is to automatically distinguish between balance sheets, income statements, cash flow statements, statements of changes in equity, and detailed notes, thus limiting the report scenarios for field recognition. The training process of the classification model is as follows: First, a labeled dataset is constructed by collecting massive amounts of screenshots and raw financial statement tables, labeling each sample with the corresponding report type label; the table layout size is standardized, text and structural features are extracted, and the dataset is divided into training, validation, and test sets; a text-image classification network is built, taking table image and text features as input and outputting report classification labels; a cross-entropy loss function is set, and iterative training is performed with parameter tuning on the validation set; the classification accuracy is verified on the test set, and the trained classification model is saved after reaching a preset threshold.

[0139] The first sequence labeling model represents a semantic labeling model for financial fields. It performs sequence labeling on synonymous but different-named indicators in financial statements, identifies semantically equivalent synonymous fields, normalizes financial indicators, and eliminates ambiguity caused by multiple textual expressions of the same indicator. The training process of the first sequence labeling model is as follows: construct a corpus of financial statement field texts, perform sequence labeling on synonymous but different-named texts such as operating revenue and total operating revenue, and label them with unified standard financial indicator tags; clean the text, segment it into words, and convert it into vectors that can be input to the model; use a sequence labeling base network and iteratively optimize it with a labeling loss function; continuously verify the recall rate of synonymous field recognition, and save the model after convergence.

[0140] The second sequence labeling model represents a period labeling model. Based on the contextual information of financial statements, it identifies various time-dimensional labels, binding the numerical values ​​in the table to their corresponding accounting periods to distinguish data by period dimension. The training process of the second sequence labeling model is as follows: Collect financial statement text samples with period labels, labeling them with time labels such as current period, same period last year, and year-to-date; integrate contextual features such as financial statement period, industry, and company name as additional inputs to the model; train the sequence labeling network to learn the mapping relationship between text and period labels; after optimization, verify the accuracy of period alignment and matching to complete model training.

[0141] The third sequence labeling model represents a numerical association element matching and labeling model. It labels the numerical values ​​in a table with associated elements, matching the financial field, unit of measurement, currency, and positive / negative / increase / decrease attributes to which the amount belongs, thus clarifying the complete attribute attribution of each numerical value. The training process of the third sequence labeling model is as follows: All amounts, fields, units, currencies, and positive / negative values ​​in the financial report are labeled element-by-element to construct a numerical element labeling dataset; the table text, coordinates, and structural features are used as input to train the model to automatically match numerical values ​​with various associated elements; the element matching error is optimized through a loss function, and training stops once the numerical attribution accuracy reaches a certain level.

[0142] The table structure parsing model is a dedicated model for table header deconstruction. It's a deep learning model capable of understanding the visual structure of tables in images or portable document formats and accurately reconstructing them into a computer-readable format. It performs hierarchical deconstruction and structural correction for complex financial report headers with multi-level headers, cross-column merged titles, and cross-page continuation tables, restoring a clear and computable standard table hierarchy. The training process for the table structure parsing model is as follows: A large number of financial report samples containing multi-level headers, cross-column, and cross-page continuation tables are collected; the header hierarchy and row / column merging relationships are manually deconstructed and labeled; table borders, text positions, and cell layout features are extracted and input into the deep learning model; the model is trained to restore the standard table hierarchy structure, with focused reinforcement learning for cross-page and merged header samples; model training ends once the structure restoration error meets the requirements.

[0143] The neural network model represents a lightweight neural network for multi-source information fusion decision-making. It receives two types of heterogeneous intermediate data: OCR recognition results and document parsing results. It performs fusion judgment by comprehensively considering confidence, data source, context, and rule matching scores, and outputs the optimal recognition result or multiple candidate results. The training process of the lightweight neural network model is as follows: construct a multi-source fusion dataset, where each sample simultaneously carries the OCR recognition result and document parsing result of the financial report region, and labels the optimal standard result; input features include the confidence, source, and context matching scores of the two types of data; build a lightweight neural network, and train the fusion decision logic using the standard labeled results as supervision signals; after training, the model can autonomously select the optimal result or output a set of candidate results.

[0144] In the embodiments of this application, multiple pre-trained recognition models are used to perform report classification, field labeling, period mapping, numerical matching, table deconstruction, and multi-source fusion, thus deconstructing the complex semantic structure of financial statements. Classification models, sequence labeling models, table structure parsing models, and neural network models are specifically designed to handle different parsing tasks, solving the problem that a single model cannot adequately handle header levels, time dimensions, and numerical relationships. This ensures accurate alignment of report field definitions based on different criteria and periods, and efficient extraction of financial statement data.

[0145] According to embodiments of this application, a balance verification and anomaly location are performed on a financial indicator table based on a formula configuration rule base, including: extracting the indicator rules and tolerance thresholds corresponding to the configuration formulas based on the formula configuration rule base; calling a balancing tool to perform indicator verification on the financial indicator table according to the indicator rules and tolerance thresholds; the indicator verification includes at least one or more of the following: reconciliation relationship verification, cross-source verification, period continuity verification, and tolerance judgment; in the case of indicator verification failure, locating abnormal fields based on the indicator verification results and field traceability information; performing candidate value iterative replacement on abnormal fields with multiple candidate values ​​to balance the configuration formulas; and integrating the indicator verification results, abnormal fields, and / or balancing results.

[0146] Balancing tools are rule-based processing tools that perform consistency checks, cross-reference verification, anomaly location, and error correction assistance on OCR results, document parsing results, and recognition model output results based on preset or configurable formula rules.

[0147] Based on the formula configuration rule library, the indicator rules and tolerance thresholds corresponding to the configuration formulas are extracted. The formula configuration rule library contains indicator codes, operation expressions, conditional expressions, applicable industries, applicable report types, priorities and tolerance thresholds, etc. It supports graphical or script-based configuration of various configurable validation formulas.

[0148] The financial indicator table is validated by using a balancing tool. Validation includes reconciliation verification, cross-source verification, period continuity verification, and tolerance assessment. Specifically, reconciliation verification involves calculating and comparing the formulas for indicators based on the internal and inter-statement logical relationships within the balance sheet, income statement, and cash flow statement; cross-source verification compares the disclosure of the same indicator in the main table, notes, and summary, performing cross-consistency comparisons if the same indicator appears in multiple places in these tables; period continuity verification logically verifies the relationships between beginning, ending, current, previous, and cumulative values; and tolerance assessment allows data to be considered consistent within a preset tolerance threshold, effectively accommodating and eliminating minor calculation errors caused by rounding.

[0149] In the event of a failed indicator validation, based on the indicator validation results and field traceability information (such as the dependency chain of the configuration formula, field confidence and source location), mark the fields that may contain errors in order to locate the abnormal fields.

[0150] When a field has multiple candidate values, perform an iterative replacement operation on the abnormal field with multiple candidate values. It traverses multiple sets of candidate identification values ​​in the field, performs a combination search in combination with the formula configuration, and tries to match the combination of values ​​that can meet the conditions of the configured formula. That is, it automatically selects the value that can satisfy the correlation relationship from the candidate values ​​as the recommended value, so as to balance the configuration formula and achieve automatic error correction.

[0151] Integrate data such as indicator verification results, abnormal fields, and balance results to generate verified financial statement verification results, including a financial indicator table, a list of abnormal fields, balance status, suspected error sources, and correction suggestions.

[0152] Provide users with an interactive correction interface to enable manual error correction and feedback based on financial statement verification results. First, read the verified financial indicator table, abnormal field list, source location, candidate value and correction suggestions, and load them into the interactive correction interface for visualization. The interface simultaneously displays the original page number and coordinate area corresponding to the recognition result, and also displays the main selected value, candidate value, confidence level and formula verification status of each field.

[0153] It provides multiple manual interaction capabilities, supporting users to manually modify erroneous fields, adjust field mappings, and specify attribution or unit affiliation during revision. Users can also add, modify, or disable balancing formulas. After completing all revision operations, the system records all revision results and uses them as error correction feedback data. Feedback information generated after users modify the recognition results through the balancing tool, manual correction interface, or rule editing interface serves as error correction feedback. This feedback information can be used to update field mappings, rule bases, or training samples to improve the accuracy of subsequent financial statement recognition.

[0154] The revised financial indicator table was then rebalanced and anomaly located, and a review was conducted. Once the review was passed, the revised financial indicator table was input into the large model for further analysis.

[0155] In the embodiments of this application, multi-dimensional financial verification is achieved based on a configurable formula rule base and balancing tools. This covers multiple verification logics including reconciliation, cross-validation, period verification, and tolerance verification. Verification failures can be accurately traced back to abnormal fields, and candidate values ​​are automatically iterated and replaced to complete formula balancing. This addresses the shortcomings of traditional solutions, such as lack of financial logic verification and difficulty in error location. Abnormal data is intercepted in advance, eliminating the need for manual line-by-line data verification and improving the reliability of indicators.

[0156] According to embodiments of this application, based on financial statement verification results and contextual information, a pre-trained large model is used to analyze the financial indicator table, including: quantifying multi-dimensional financial characteristics of the financial indicator table to extract quantitative financial characteristics; performing multi-period time-series fluctuation analysis on the financial indicator table to extract time-series trend characteristics; performing risk analysis on the financial indicator table based on the financial statement verification results to extract risk characteristics; and performing operational status analysis based on contextual information, quantitative financial characteristics, time-series trend characteristics, and risk characteristics.

[0157] When analyzing financial indicator tables using a large-scale model, the first step is to quantify the multi-dimensional financial characteristics of the tables. The large-scale model then calculates and aggregates indicators related to enterprise growth, profitability, solvency, operational efficiency, and cash flow quality in batches to extract quantitative financial characteristics. The extraction of these quantitative financial characteristics uses only the structured and validated financial indicator tables as core input, supplemented by supplementary text summaries to reduce model illusions and ensure stable and consistent analytical conclusions.

[0158] Based on a multi-period time-series financial indicator sequence, a large model is used to mine the time-series fluctuation patterns of various indicators, realize multi-period time-series fluctuation analysis, identify enterprise operating trends by combining changes in multi-period financial indicators, and extract time-series trend features.

[0159] Based on the anomaly markers and confidence information identified in the financial statement verification results, the large model conducts risk analysis, completes confidence correction and risk feature mapping, extracts risk features, and imposes credibility constraints or risk warnings on the analysis results.

[0160] The large model combines contextual information such as basic enterprise information and industry information, and integrates quantitative financial characteristics, time-series trend characteristics and risk characteristics to conduct a comprehensive analysis of the enterprise's operating status. It summarizes and generates operating judgment conclusions such as "revenue growth but profit pressure", "cash flow improvement" and "weak short-term solvency". It simultaneously outputs structured analysis results and analysis reports in natural language form, including operating status analysis results, risk warning information, analysis summary and visual explanatory text.

[0161] In the embodiments of this application, a large model is used to analyze multidimensional financial characteristics, multi-period time-series fluctuations, and risks. Combined with contextual information, the operating conditions are comprehensively analyzed. It integrates financial indicators, time-series trends, and risk points, solving the problem that traditional analysis relies solely on raw text and the conclusions lack financial basis. This improves the stability and interpretability of the analysis results, and the output content is tailored to the actual analysis needs of auditing, investment and financing, and other businesses.

[0162] For example, based on the above-described financial report analysis method, embodiments of this application also provide a financial report intelligent recognition and analysis method. The following will combine... Figure 4 The method is described in detail.

[0163] Figure 4 A flowchart illustrating a financial statement intelligent recognition and analysis method according to an embodiment of this application is shown schematically. Figure 4 As shown, the function receives financial report documents: It receives financial report documents uploaded by users (portable document format, spreadsheet file format, image or mixed format documents, etc.) and records the file type, file batch and context information, etc.

[0164] Document preprocessing and type routing: The document is preprocessed and identified as a scanned document, native portable document format, spreadsheet file format, or hybrid document based on features such as whether the page is a scanned document, whether it contains parsable tables, and whether it has multi-level headers. Different pages or regions are routed to different recognition channels, and the corresponding OCR processing path, document parsing path, or hybrid processing path is selected.

[0165] OCR Recognition Module: Performs text and table recognition on image regions, extracts text and table content, and obtains OCR recognition results.

[0166] Document parsing module: Performs table parsing and cell extraction on native text or spreadsheet file formats to extract data such as document content, cell structure, header hierarchy, and text layout, and obtains document parsing results.

[0167] Lightweight recognition model fusion recognition: It integrates results from multiple sources, identifies report types, financial fields, amounts, periods, units, and currencies, and establishes the correspondence between fields and values ​​to form initial structured financial statement data.

[0168] Field standardization and financial indicator table construction: The field names, units, currencies, periods and numerical formats are standardized and converted into a financial indicator table with a unified format. Field source and confidence indexes are also established to form a standardized financial indicator table.

[0169] The balancing tool performs formula verification: Based on preset or user-configured formulas, the balancing tool performs cross-reference verification, cross-source verification, and period continuity verification on the financial indicator table, and outputs the pass status, anomalies, and correction suggestions.

[0170] Anomaly location and manual correction: When the verification fails, there are abnormal or low-confidence fields. The system locates the suspected error fields, marks the abnormal fields and generates correction suggestions. The system can also enter the manual correction process. After the correction is completed, the indicator table is generated again, the user's modification results are received and the verification is re-executed until a financial indicator table that meets the preset conditions is obtained.

[0171] Large-scale model operation status analysis: After the verification is passed, the system will input the verified financial indicator table, notes summary and abnormal status into the large-scale model analysis engine, output the operation status analysis report and risk warning, and generate the enterprise operation status analysis results.

[0172] Generate analysis reports and tables: The system outputs structured indicator tables, verification results, and business analysis reports to users, displaying financial indicator tables, balance status, and business status analysis results, and supports export.

[0173] From a data flow perspective, the original financial statement document is first converted into page objects, region objects, and table objects; then, the OCR recognition results and document parsing results together constitute multi-source intermediate data; in the lightweight recognition model, the multi-source intermediate data is converted into structured financial fields; during field standardization and indicator construction, the structured financial fields are further converted into standard financial indicator tables; after balancing verification and manual correction, highly reliable structured financial statement data is formed, which ultimately drives the large model analysis.

[0174] From a control flow perspective, the system prioritizes automatic identification and automatic verification. When verification fails or the identification confidence level is below the threshold, a manual error correction process is triggered. After manual correction is completed, the system re-enters the verification process. When verification passes or the preset output conditions are met, a large model analysis process is triggered. After analysis is completed, the system enters the display and export process.

[0175] Compared with the prior art, this application has at least the following advantages:

[0176] Significantly improves recognition accuracy in complex financial reporting scenarios: By integrating OCR recognition, spreadsheet file format / portable document format parsing, and lightweight recognition models, it can adapt to scanned documents, native documents, and mixed financial reports, reducing the probability of field misalignment, misidentification of amounts, omission of units, and errors in period mapping.

[0177] Enhancing the credibility and usability of structured financial data: By using balancing tools to perform formula validation, cross-reference validation, and cross-source validation on financial indicators, erroneous fields can be identified and located during the identification phase, preventing obviously erroneous data from being used directly in subsequent analysis and improving the quality of output results.

[0178] Reduce the cost of manual review and improve the efficiency of automated processing: The financial statement processing flow is transformed from "manual entry + manual verification + manual analysis" to "automatic identification + automatic verification + anomaly correction + automatic analysis". Manual intervention is triggered only when abnormal or low-confidence fields appear, thereby significantly reducing the workload of manual processing and shortening the processing time.

[0179] Improve the stability and interpretability of operational status analysis: The large model analysis engine uses validated structured financial indicators as core inputs and combines source information, confidence levels, and anomaly markers for analysis constraints. This ensures that operational status analysis is no longer based on unverified raw text, thereby improving the reliability and interpretability of the analysis results.

[0180] It has good scalability and rule configurability: Through field mapping mechanism and formula configuration mechanism, the system can quickly adapt to different enterprise templates, different industry standards and different analysis scenarios, reduce the development cost when adding new rules and templates, and enhance the long-term maintainability of the system.

[0181] Supports continuous optimization of the closed loop to improve long-term recognition performance: Manual error correction results, verification failure logs, and revision records can be fed back into the rule base and model training data to form an optimization closed loop of "recognition-verification-error correction-iteration", so that the system's recognition capability and rule coverage capability can be continuously improved with use.

[0182] Balancing user experience with engineering feasibility: The system not only outputs the final analysis report, but also financial indicator tables, source locations, verification status, and reasons for anomalies, making it easy for users to verify and trace. It is suitable for direct application in scenarios such as financial risk control, investment and financing analysis, audit assistance, and enterprise operation analysis.

[0183] Based on the above-described financial report analysis method, embodiments of this application also provide a financial report analysis apparatus. The following will be combined with... Figure 5 The device is described in detail.

[0184] Figure 5 A schematic block diagram of a financial report analysis apparatus according to an embodiment of this application is shown.

[0185] like Figure 5 As shown, the financial report analysis device 1000 of this embodiment includes a preprocessing module 1010, an identification module 1020, a parsing module 1030, a mapping module 1040, a verification module 1050, and an analysis module 1060.

[0186] The preprocessing module 1010 is used to preprocess financial report data to obtain objects to be processed and routing information; wherein, the routing information includes document type information and processing channel corresponding to the objects to be processed. In one embodiment, the preprocessing module 1010 can be used to perform step S210 described above, which will not be repeated here.

[0187] The identification module 1020 is used to identify the text and image information of the object to be processed based on the document type information corresponding to the object to be processed, through the processing channel corresponding to the object to be processed, to obtain financial statement content data. In one embodiment, the identification module 1020 can be used to execute the step S220 described above, which will not be repeated here.

[0188] The parsing module 1030 is used to perform structured parsing of the financial statement content data using a pre-trained recognition model to obtain structured data. In one embodiment, the parsing module 1030 can be used to execute step S230 described above, which will not be repeated here.

[0189] The mapping module 1040 is used to standardize and map the structured data to generate a financial indicator table. In one embodiment, the mapping module 1040 can be used to perform step S240 described above, which will not be repeated here.

[0190] The verification module 1050 is used to perform balance verification and anomaly location on the financial indicator table based on the formula configuration rule base, and obtain the financial statement verification result. In one embodiment, the verification module 1050 can be used to execute the step S250 described above, which will not be repeated here.

[0191] The analysis module 1060 is used to analyze the financial indicator table based on the financial statement verification results and context information using a pre-trained large model to obtain financial statement analysis results. In one embodiment, the analysis module 1060 can be used to execute step S260 described above, which will not be repeated here.

[0192] According to an embodiment of this application, the document type information includes image type information and / or text type information, the processing channel includes an optical character recognition channel and / or a document parsing channel, and the recognition module 1020 is further configured to perform optical character recognition on the object to be processed based on the image type information through the optical character recognition channel; and / or perform document parsing on the object to be processed based on the text type information through the document parsing channel.

[0193] According to an embodiment of this application, the recognition module 1020 is further configured to perform text character recognition on the object to be processed to obtain text content and confidence level; perform table structure recognition on the object to be processed to determine page space information; recognize and extract numerical fields in the object to be processed to obtain numerical field content; and integrate the text content, the confidence level, the page space information and / or the numerical field content.

[0194] According to an embodiment of this application, the identification module 1020 is further configured to perform spreadsheet structure extraction on the object to be processed to obtain spreadsheet data; perform portable document parsing on the object to be processed to obtain document data; perform table identifier semantic extraction on the object to be processed to obtain table annotation information; perform table structure reconstruction on the object to be processed to generate table structure data; and integrate the spreadsheet data, the document data, the table annotation information and / or the table structure data.

[0195] According to an embodiment of this application, the verification module 1050 is further configured to extract the indicator rules and tolerance thresholds corresponding to the configuration formula based on the formula configuration rule base; call the balancing tool to perform indicator verification on the financial indicator table according to the indicator rules and the tolerance thresholds; the indicator verification includes at least one or more of the following: reconciliation relationship verification, cross-source verification, period continuity verification, and tolerance judgment; in the case of indicator verification failure, locate abnormal fields according to the indicator verification results and field traceability information; perform candidate value iterative replacement on abnormal fields with multiple candidate values ​​to balance the configuration formula; and integrate the indicator verification results, the abnormal fields, and / or the balancing results.

[0196] According to an embodiment of this application, the apparatus 1000 further includes: a document access module for receiving financial report documents; and performing document access processing on the financial report documents to obtain the financial report data and the context information, wherein the document access processing includes one or more of basic information detection, file type identification, metadata extraction, and associated document encapsulation.

[0197] According to an embodiment of this application, the financial report data includes standardized financial statement documents and document format types. The preprocessing module 1010 is further configured to extract table objects, text objects, and / or page objects from the standardized financial statement documents; perform image preprocessing on the page objects, and perform page layout analysis on the preprocessed page objects to obtain region objects; determine the object to be processed based on the table objects, the text objects, the page objects, and the region objects; and perform format parsing on the object to be processed based on the document format type to determine the routing information.

[0198] According to an embodiment of this application, the analysis module 1060 is further configured to quantify the financial characteristics of the financial indicator table in multiple dimensions to extract quantitative financial characteristics; perform multi-period time-series fluctuation analysis on the financial indicator table to extract time-series trend characteristics; perform risk analysis on the financial indicator table based on the financial statement verification results to extract risk characteristics; and perform operational status analysis based on the context information, the quantitative financial characteristics, the time-series trend characteristics, and the risk characteristics.

[0199] According to embodiments of this application, the identification model includes a classification model, a first sequence labeling model, a second sequence labeling model, a third sequence labeling model, a table structure parsing model, and / or a neural network model; the parsing module 1030 is further configured to: identify the report type of the financial statement content data using the classification model; identify field labels of the financial statement content data using the first sequence labeling model; perform period alignment and mapping of the financial statement content data using the second sequence labeling model based on the context information; perform numerical correlation element matching of the financial statement content data using the third sequence labeling model; perform header level deconstruction of the financial statement content data using the table structure parsing model; and / or perform multi-source information fusion processing of the financial statement content data using the neural network model.

[0200] According to an embodiment of this application, the device 1000 further includes: a visualization and data encapsulation module for visualizing the financial indicator table and the financial statement analysis results; extracting indicator status information from the financial indicator table and visualizing the indicator status information; extracting abnormal information based on the balancing results and the financial statement verification results and visualizing the abnormal information; and performing serialization encapsulation processing on the financial statement verification results, the financial indicator table, and the financial statement analysis results based on a preset output format to generate interface output data.

[0201] According to embodiments of this application, any multiple modules among the preprocessing module 1010, identification module 1020, parsing module 1030, mapping module 1040, verification module 1050, analysis module 1060, document access module, and visualization and data encapsulation module can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the preprocessing module 1010, identification module 1020, parsing module 1030, mapping module 1040, verification module 1050, analysis module 1060, document access module, and visualization and data encapsulation module can be at least partially implemented as hardware circuits, such as field-programmable gate arrays, programmable logic arrays, systems-on-a-chip, systems-on-a-substrate, systems-on-package, application-specific integrated circuits, or any other reasonable means of integrating or encapsulating circuits, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the preprocessing module 1010, recognition module 1020, parsing module 1030, mapping module 1040, verification module 1050, analysis module 1060, document access module, and visualization and data encapsulation module can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0202] Figure 6 A block diagram schematically illustrates an electronic device suitable for implementing a financial reporting analysis method according to an embodiment of this application.

[0203] like Figure 6 As shown, an electronic device 1200 according to an embodiment of this application includes a processor 1201, which can perform various appropriate actions and processes according to a program stored in a read-only memory 1202 or a program loaded from a storage portion 1208 into a random access memory 1203. The processor 1201 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 1201 may also include onboard memory for caching purposes. The processor 1201 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0204] Random access memory 1203 stores various programs and data required for the operation of electronic device 1200. Processor 1201, read-only memory 1202, and random access memory 1203 are interconnected via bus 1204. Processor 1201 executes various steps of the method flow according to embodiments of this application by executing programs in read-only memory 1202 and / or random access memory 1203. It should be noted that the programs may also be stored in one or more memories other than read-only memory 1202 and random access memory 1203. Processor 1201 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0205] According to embodiments of this application, the electronic device 1200 may further include an input / output interface 1205, which is also connected to the bus 1204. The electronic device 1200 may also include one or more of the following components connected to the input / output interface 1205: an input section 1206 including a keyboard, mouse, etc.; an output section 1207 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 1208 including a hard disk, etc.; and a communication section 1209 including a network interface card, such as a local area network card, modem, etc. The communication section 1209 performs communication processing via a network such as the Internet. A drive 1210 is also connected to the input / output interface 1205 as needed. A removable medium 1211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 1210 as needed so that computer programs read from it can be installed into the storage section 1208 as needed.

[0206] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0207] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 1202, and / or random access memory 1203, and / or one or more memories other than read-only memory 1202 and random access memory 1203 described above.

[0208] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0209] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 1209, and / or installed from the removable medium 1211. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0210] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 1209, and / or installed from removable medium 1211. When the computer program is executed by processor 1201, it performs the functions defined in the system of this application embodiment. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0211] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0212] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0213] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A financial report analysis method, characterized in that, The method includes: Preprocess financial report data to obtain objects to be processed and routing information; wherein, the routing information includes document type information and processing channel corresponding to the objects to be processed; Based on the document type information corresponding to the object to be processed, the image and text information of the object to be processed is identified through the processing channel corresponding to the object to be processed to obtain financial report content data; The financial statement content data is structured and parsed using a pre-trained recognition model to obtain structured data. The structured data is standardized and mapped to generate a financial indicator table; Based on the formula configuration rule base, the financial indicator table is balanced, validated, and anomaly located to obtain the financial statement validation results; and Based on the financial statement verification results and contextual information, the financial indicator table is analyzed using a pre-trained large model to obtain financial statement analysis results.

2. The method according to claim 1, characterized in that, The document type information includes image type information and / or text type information; the processing channel includes an optical character recognition channel and / or a document parsing channel; the step of performing image and text information recognition on the object to be processed based on the document type information corresponding to the object to be processed, through the processing channel corresponding to the object to be processed, includes: Based on the image type information, optical character recognition is performed on the object to be processed through the optical character recognition channel; and / or Based on the text type information, the document parsing channel is used to parse the object to be processed.

3. The method according to claim 2, characterized in that, The step of performing optical character recognition on the object to be processed through the optical character recognition channel includes: The object to be processed is subjected to text character recognition to obtain the text content and confidence level; The table structure of the object to be processed is identified to determine the page space information; Identify and extract the numeric fields from the object to be processed to obtain the content of the numeric fields; and Integrate the text content, the confidence level, the page layout information, and / or the content of the numerical fields.

4. The method according to claim 2, characterized in that, The step of parsing the object to be processed through the document parsing channel includes: The object to be processed is subjected to spreadsheet structure extraction to obtain spreadsheet data; The object to be processed is subjected to portable document parsing to obtain document data; The table identifier semantics of the object to be processed are extracted to obtain table annotation information; The object to be processed is reconstructed into a table structure to generate table structure data; and Integrate the spreadsheet data, the document data, the table annotation information, and / or the table structure data.

5. The method according to claim 1, characterized in that, The method of balancing, verifying, and locating anomalies in the financial indicator table based on the formula-configured rule base includes: Based on the formula configuration rule base, extract the indicator rules and tolerance thresholds corresponding to the configuration formula; The balancing tool is invoked to perform indicator verification on the financial indicator table according to the indicator rules and the tolerance threshold; the indicator verification includes at least one or more of the following: reconciliation relationship verification, cross-source verification, period continuity verification, and tolerance judgment. In the event of a failed indicator verification, the abnormal field can be located based on the indicator verification results and field traceability information. For outlier fields with multiple candidate values, iterative replacement of candidate values ​​is performed to balance the configuration formula; and Integrate the indicator verification results, the abnormal fields, and / or the balancing results.

6. The method according to claim 1, characterized in that, The method further includes: Receiving financial reporting documents; and The financial report document is subjected to document access processing to obtain the financial report data and the context information. The document access processing includes one or more of the following: basic information detection, file type identification, metadata extraction, and associated document encapsulation.

7. The method according to claim 1, characterized in that, The financial reporting data includes standardized financial statement documents and document format types. The preprocessing of the financial reporting data, obtaining the objects to be processed and routing information, includes: Extract the table objects, text objects, and / or page objects from the standardized financial statement document; The page object is preprocessed with an image, and the page layout of the preprocessed page object is analyzed to obtain the region object; The object to be processed is determined based on the table object, the text object, the page object, and the region object; and Based on the document format type, the format of the object to be processed is parsed to determine the routing information.

8. The method according to claim 1, characterized in that, The analysis of the financial indicator table based on the financial statement verification results and context information, using a pre-trained large model, includes: The financial indicator table is subjected to multidimensional financial feature quantification to extract quantitative financial features; The financial indicator table is subjected to multi-period time-series fluctuation analysis to extract time-series trend characteristics; Based on the financial statement verification results, a risk analysis is performed on the financial indicator table to extract risk characteristics; and Operational status analysis is performed based on the contextual information, the quantitative financial characteristics, the time-series trend characteristics, and the risk characteristics.

9. The method according to claim 1, characterized in that, The recognition model includes a classification model, a first sequence labeling model, a second sequence labeling model, a third sequence labeling model, a table structure parsing model, and / or a neural network model; the structured parsing of the financial statement content data using the pre-trained recognition model includes: The financial statement content data is identified by the classification model. The financial statement content data is labeled using the first sequence labeling model; Based on the context information, the financial statement content data is period-aligned and mapped using the second sequence labeling model; The financial statement content data is matched with numerical correlation elements using the third sequence labeling model. The table structure parsing model is used to deconstruct the financial statement data by header level; and / or The financial report data is processed by multi-source information fusion using the neural network model.

10. The method according to claim 5, characterized in that, The method further includes: Visualize the financial indicator table and the financial statement analysis results; Extract the indicator status information from the financial indicator table and visualize the indicator status information; Based on the balancing results and the financial statement verification results, abnormal information is extracted and visualized; and Based on a preset output format, the financial statement verification results, the financial indicator table, and the financial statement analysis results are serialized and encapsulated to generate interface output data.

11. A financial report analysis device, characterized in that, The device includes: The preprocessing module is used to preprocess financial report data to obtain objects to be processed and routing information; wherein, the routing information includes document type information and processing channel corresponding to the objects to be processed; The recognition module is used to recognize the text and image information of the object to be processed based on the document type information corresponding to the object to be processed and through the processing channel corresponding to the object to be processed, so as to obtain financial report content data. The parsing module is used to perform structured parsing of the financial statement content data using a pre-trained recognition model to obtain structured data; The mapping module is used to standardize and map the structured data to generate a financial indicator table; The verification module is used to perform balance verification and anomaly location on the financial indicator table based on the formula-configured rule base, and obtain the financial statement verification results; and The analysis module is used to analyze the financial indicator table based on the financial statement verification results and context information using a pre-trained large model to obtain financial statement analysis results.

12. An electronic device, comprising: One or more processors; Memory, used to store one or more computer programs. The characteristic feature is that the one or more processors execute the one or more computer programs to implement the steps of the method according to any one of claims 1 to 10.

13. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 10.

14. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1 to 10.

Images