Financial document review method, related device and system

Abstract

The invention discloses a financial document review method, a related device and a system, and relates to the technical field of data processing, after a to-be-reviewed multi-modal financial document is processed into structured financial data, related domain knowledge and risk analysis specification information of the structured financial data are retrieved from a dynamic knowledge graph, and the structured financial data are reviewed according to the related domain knowledge and risk analysis specification information. And generating a risk analysis instruction command based on the structured financial data, the related domain knowledge and the risk analysis specification information, and inputting the risk analysis instruction command into the generative pre-training model to obtain a risk analysis result. The process does not need manual participation in auditing, the auditing efficiency is greatly improved, the indication information of risk analysis is generated according to the latest related domain knowledge retrieved from the dynamic knowledge graph and the risk analysis specification information, and a more accurate risk analysis result of the structured financial data can be obtained in real time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a financial document review method, related device and system. BACKGROUND

[0002] With the acceleration of digital transformation of the financial industry, the compliance review demand of financial webpages such as contracts, financial reports and supervision and management system documents has shown explosive growth. The traditional manual review mode has problems of low efficiency and low accuracy, especially when dealing with multi-modal data and unstructured text, the efficiency and accuracy problems are particularly prominent. SUMMARY

[0003] In view of the above problems, the present application provides a financial document review method, related device and system to achieve the purpose of high-precision and high-efficiency review of financial documents. The specific scheme is as follows:

[0004] The first aspect of the present application provides a financial document review method, comprising:

[0005] Structurally processing the multi-modal financial document to be reviewed input by the client to obtain structured financial data;

[0006] Querying associated information matching the structured financial data from a pre-established dynamic knowledge graph; the pre-established dynamic knowledge graph is a knowledge graph that stores financial entity information and the association relationship between financial entities in a graph structure; the associated information includes related field knowledge and risk analysis specification information of the structured financial data;

[0007] Generating a risk analysis instruction command according to the structured financial data, the related field knowledge and the risk analysis specification information, inputting the risk analysis instruction command into a generative pre-training model, and obtaining a risk analysis result.

[0008] In one possible implementation, querying associated information matching the structured financial data from the pre-established dynamic knowledge graph comprises:

[0009] Generating a graph retrieval enhancement generation request according to the structured financial data, and sending the graph retrieval enhancement generation request to the dynamic knowledge graph;

[0010] Retrieving a dynamic knowledge graph subgraph corresponding to the structured financial data and risk analysis specification information from the dynamic knowledge graph;

[0011] Determining related field information according to the dynamic knowledge graph subgraph;

[0012] Taking the related field information and the risk analysis specification information as the associated information.

[0013] In one possible implementation, the updating process of the dynamic knowledge graph comprises:

[0014] In response to listening to the multi-source heterogeneous financial data flow, the multi-source heterogeneous financial data is preprocessed to obtain preprocessed multi-source heterogeneous financial data.

[0015] The financial entity type information, the relationship information between the financial entities, and the financial event attribute information of the multi-source heterogeneous financial data are determined.

[0016] The current dynamic knowledge graph is dynamically maintained and processed according to the financial entity type information, the relationship information between the financial entities, and the financial event attribute information to obtain an updated dynamic knowledge graph.

[0017] In a possible implementation, after the financial entity type information, the relationship information between the financial entities, and the financial event attribute information of the multi-source heterogeneous financial data are determined, before the current dynamic knowledge graph is dynamically maintained and processed according to the financial entity type information, the relationship information between the financial entities, and the financial event attribute information to obtain an updated dynamic knowledge graph, the method further includes:

[0018] The financial entity type information is compared with the financial entity information in the current dynamic knowledge graph.

[0019] When the financial entity type information and the financial entity information have the same financial entity, the financial entity type information and the financial entity information are co-reference resolved and credibility weighted to determine that the financial entity type information or the financial entity information is the standard description information of the financial entity.

[0020] In a possible implementation, the risk analysis instruction command is input into the generative pre-training model to obtain a risk analysis result, including:

[0021] The risk analysis instruction command is input into the generative pre-training model to obtain a structured report; the structured report includes risk point positioning information, traceable evidence chain, associated case information, and risk repair suggestion information.

[0022] In a possible implementation, the client-inputted to-be-inspected multi-modal financial document is structurally processed to obtain structured financial data, including:

[0023] The financial documents of different modalities in the to-be-inspected multi-modal financial document are parsed and processed into structured financial data with unified format.

[0024] The second aspect of the present application provides a financial document inspection device, including:

[0025] The structured processing unit is configured to structurally process the client-inputted to-be-inspected multi-modal financial document to obtain structured financial data.

[0026] The query unit is configured to query associated information matched with the structured financial data from a pre-established dynamic knowledge graph; the pre-established dynamic knowledge graph is a knowledge graph in which financial entity information and association relationships between financial entities are stored in a graph structure; the associated information includes relevant field knowledge and risk analysis specification information of the structured financial data.

[0027] The risk analysis unit is configured to generate a risk analysis instruction command according to the structured financial data, the relevant field knowledge and the risk analysis specification information, input the risk analysis instruction command into the generative pre-training model, and obtain a risk analysis result.

[0028] The third aspect of the present application provides a computer program product, which includes computer readable instructions. When the computer readable instructions are run on an electronic device, the financial document review method of the first aspect or any implementation manner of the first aspect is executed.

[0029] The fourth aspect of the present application provides a computer storage medium, which carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the financial document review method of the first aspect or any implementation manner of the first aspect is executed.

[0030] The fifth aspect of the present application provides a financial document review system, which includes:

[0031] The financial document review system includes a multi-modal data analysis module, a dynamic knowledge graph module and a risk analysis module.

[0032] The multi-modal data analysis module is configured to perform structured processing on a client-inputted multi-modal financial document to be reviewed, and obtain structured financial data.

[0033] The dynamic knowledge graph module is configured to construct and update a dynamic knowledge graph.

[0034] The risk analysis module is configured to query relevant field knowledge and risk analysis specification information matched with the structured financial data from the dynamic knowledge graph, generate a risk analysis instruction command according to the structured financial data and the relevant field knowledge and the risk analysis specification information, input the risk analysis instruction command into a generative pre-training model, and obtain a risk analysis result.

[0035] By means of the technical solutions, the financial document review method, the related device and the system provided by the application process the multi-modal financial document to be reviewed into structured financial data, retrieve relevant field knowledge and risk analysis specification information of the structured financial data from a dynamic knowledge graph, generate a risk analysis instruction command based on the structured financial data, the relevant field knowledge and the risk analysis specification information, and input the risk analysis instruction command into a generative pre-training model to obtain a risk analysis result. This process does not require manual review, and the review efficiency is greatly improved. Moreover, the indication information of the risk analysis is generated based on the latest relevant field knowledge and risk analysis specification information retrieved from the dynamic knowledge graph, so that more accurate risk analysis results of the structured financial data can be obtained in real time. BRIEF DESCRIPTION OF DRAWINGS

[0036] The above and other features, advantages, and aspects of the present disclosure will become more apparent by describing in detail exemplary embodiments thereof with reference to the attached drawings in which:

[0037] Figure 1 A flowchart of a financial document review method provided by the application;

[0038] Figure 2 A structural diagram of a financial document review device provided by the application;

[0039] Figure 3 A structural composition diagram of a financial document review system provided by the application;

[0040] Figure 4 An example diagram of a review process of a financial document review system provided by the application. DETAILED DESCRIPTION

[0041] The embodiments of the present application are described below with reference to the accompanying drawings. The terms used in the embodiment part of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0042] The embodiments of the present application are described below with reference to the accompanying drawings. The embodiments of the present application are described below with reference to the accompanying drawings. Those skilled in the art can know that, with the development of technology and the appearance of new scenarios, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0043] The terms "first", "second", and the like in the description and in the claims of the present application and above drawings are used for distinguishing between similar objects and not necessarily for describing a specific sequential or chronological order. It is to be understood that the terms so used are interchangeable under appropriate circumstances and that the embodiments of the present application are capable of functioning in other sequences, except where it is inherent from the procedure. Furthermore, the terms "comprise", "comprising", "include", "including", and the like are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises, includes, or includes elements or steps that are not listed is not excluded from the scope of the embodiments of the present application.

[0044] It can be understood that, before using the technical solutions disclosed in the embodiments of the present application, the type of personal information involved in the present application, the scope of use, the use scenario, etc. should be informed to the user and the authorization of the user should be obtained according to relevant laws and regulations.

[0045] In addition, the data involved in the present technical solution (including but not limited to the data itself, the acquisition or use of the data) should comply with the requirements of relevant laws and regulations and relevant provisions.

[0046] Under the background of rapid development of financial technology, compliance and risk prevention of financial business have become the core proposition for stable operation of institutions. As the core carrier of transaction constraints, financial disclosure, and system specification, financial documents cover multiple scenarios such as contract review, financial report verification, and supervision and management compliance report generation, which directly affect the legal risk, financial risk, and supervision and management compliance level of financial institutions. With the continuous expansion of the scale of financial business and the accelerated iteration of product innovation, the number of financial documents has increased exponentially, and the content complexity has significantly improved, including a large amount of unstructured text (such as loan contracts with inconsistent formats, financial derivative agreements with cross-clauses), and multi-modal data (such as scanned system files, financial statements containing tables and charts, and electronic signature attachments), which puts higher requirements on the comprehensiveness, timeliness, and accuracy of review.

[0047] Currently, financial document review mainly relies on traditional manual review combined with primary NLP (Natural Language Processing) technology, but this approach is inefficient and has low accuracy, which is manifested as follows:

[0048] Artificial review is inefficient and difficult to meet the processing needs of massive documents. Specifically, artificial review requires employees with legal and financial expertise to read documents page by page, manually extract key clauses (such as breach of contract, scope of guarantee), check data logic (such as financial statement reconciliation), and identify potential risks (such as transactions beyond authorized scope). For example, a cross-border financing contract containing 500 pages often takes 3-5 working days for several experts to complete, and during peak periods (such as the end of the quarter when financial reports are concentrated and contracts are signed in batches), the backlog of review tasks may cause compliance timeliness to lag seriously, and even lead to supervision and management penalties.

[0049] Artificial review is not accurate enough, and risk omission is a major problem. Specifically, artificial review is limited by the experience level and attention fluctuation of the reviewer, and is prone to misjudgment or omission of key information: for example, in complex financial derivative contracts, the implicit "cross default clause" may be scattered in different chapters, and artificial review may miss its amplification effect on overall risk due to fatigue or professional blindness; in financial report review, abnormal fluctuations in related party transaction data (such as a 300% increase in the purchase amount of a supplier) may not be identified due to the reviewer's lack of understanding of the industry background, ultimately leading to the failure to timely detect financial fraud risks. According to industry research data, the contract risk omission rate of artificial review is as high as 15%-20%, and the accuracy rate of identifying abnormal financial data is less than 70%.

[0050] Traditional NLP technology has limited processing capabilities for multi-modal and unstructured data, making it difficult to meet dynamic compliance needs.

[0051] Existing systems attempt to introduce rule engines or basic NLP models (such as regular expression matching, keyword extraction), but can only handle documents with fixed formats and simple semantics (such as standardized loan agreements). For complex documents with flexible clause expressions and multi-modal information (such as handwritten annotations in scanned documents, financial statements with mixed text and tables), the identification accuracy is often less than 50%; in addition, rule engines rely on manually preset "risk keyword libraries" and are difficult to adapt to dynamic adjustments of supervision and management regulations, resulting in the need for frequent manual updates of rule libraries, high maintenance costs, and the risk of rule lag.

[0052] In summary, traditional artificial review and basic NLP technology have been unable to meet the urgent needs of financial document review in terms of efficiency and accuracy.

[0053] Specifically, the current mainstream review methods include rule engine-based review strategies, statistical analysis and risk scoring model-based review strategies, traditional machine learning model-based review strategies, and fine-tuning applications of pre-trained language models. Next, we will introduce the review strategies involved in the above review methods from the perspectives of technical principles and limitations:

[0054] For rule engine-based review strategies, the technical principle is to rely on manually defined regular expressions, keyword lists, and logical rules for text matching. Its limitations are: unable to handle unpredefined text patterns; rules need to be manually updated to adapt to changing management requirements; only suitable for pure text, unable to parse tables and charts for implicit risks; lack of semantic understanding;

[0055] For statistical analysis and risk scoring model-based review strategies, the technical principle is to build a risk scoring card based on statistical indicators (such as word frequency, co-occurrence network) to assist manual review. Its limitations are that the scoring card threshold relies on expert experience, which can easily lead to false positives / false negatives; static scoring models are difficult to capture market sentiment mutations;

[0056] For traditional machine learning model review strategies, the technical principle is to use feature engineering + classification models (such as SVM, random forest) for risk classification. For example, extract TF-IDF features from contract text and train a classification model to distinguish "high-risk clauses". Its limitations are that it relies on manual feature design and is difficult to capture deep semantics; poor domain transferability, difficult to apply across scenarios;

[0057] For fine-tuning applications of pre-trained language models, the technical principle is to fine-tune general pre-trained models (such as BERT, GPT) for domain-specific tasks such as entity recognition and text classification. Its limitations are: poor model interpretability, difficult to generate review evidence chain that meets management requirements; high cost of model retraining, unable to quickly adapt to sudden changes in management requirements.

[0058] In summary, the above-mentioned traditional financial document review methods still have problems such as high reliance on manual work, static knowledge base, and weak multi-modal processing. Fine-tuning pre-trained models can reduce reliance on manual work while enhancing semantic understanding, but there is still a conflict between black-box models and white-box management requirements.

[0059] To solve the above problems, the present application provides a financial document review method and related device.

[0060] Optionally, referring to Figure 1 , the present application provides a flowchart of a financial document review method. As Figure 1 shown, the financial document review method includes the following steps:

[0061] Step 101, structuring the client inputted multi-modal financial document to be reviewed to obtain structured financial data.

[0062] It should be noted that the to-be-inspected multi-modal financial document usually includes text, image, table, audio, video and other data formats, and often exists in the form of mixed modalities, such as the combination of text and table in a contract. The to-be-inspected multi-modal financial document includes structured and unstructured text, image and chart, table and structured data, audio and video, mixed modal web page and other specific categories and typical contents.

[0063] Specifically, the structured and unstructured text includes financial statements, contracts and agreements, supervision and management documents, news and public opinion, etc.; the image and chart includes scanned copies and scanned contracts, financial charts, flowcharts and framework diagrams, etc.; the table and structured data includes transaction records, database export files, supervision and management report forms, etc.; the audio and video includes telephone conference recordings, training videos and demonstration materials, etc.; the mixed modal web page includes prospectuses, research reports, supervision and management decision documents, etc.

[0064] For example, the to-be-inspected multi-modal financial document is structured, and the main purpose is to obtain structured financial data with uniform format. Specifically, the text, table, PDF scanned copy, voice and other multi-modal financial data can be parsed through OCR (Optical Character Recognition) and voice recognition technologies to obtain uniform structured financial data.

[0065] Optionally, the financial documents in different modalities in the to-be-inspected multi-modal financial document are parsed and processed into structured financial data with uniform format.

[0066] This process can realize input and parsing of multi-modal data, and has high processing efficiency and no manual intervention.

[0067] Step 102, querying associated information matched with the structured financial data from a pre-established dynamic knowledge graph.

[0068] The pre-established dynamic knowledge graph is a knowledge graph that stores financial entity information and the association relationship between financial entities in a graph structure. Specifically, the financial field entities such as enterprises, regulations and clauses, and risk events, and the relationships between the financial field entities such as stock ownership relationship and system reference can be stored in a graph structure. The pre-established dynamic knowledge graph provided in the application supports complex queries and reasoning.

[0069] The pre-established dynamic knowledge graph of the application can synchronize the related system changes and data information changes of the financial entities in real time, so as to ensure that the information retrieved from the dynamic knowledge graph is the latest, and further to enhance the accuracy and timeliness of the subsequent risk analysis results.

[0070] Next, the construction process of the dynamic knowledge graph is introduced:

[0071] Step one, multi-source heterogeneous data collection and preprocessing: Collect, clean and standardize various structured and unstructured data from various financial databases, institutional documents, corporate annual reports, equity relationship maps, historical audit reports, industry case libraries, etc.

[0072] Step two, financial entity and relationship extraction: Fine-tune the pre-trained language model in the financial field (such as FinBERT, FinGPT) to complete the tasks of named entity recognition (extracting financial entities and attributes), entity relationship extraction (such as "Company A-Controlling-Company B" "Event X-Influence-Industry Y") and event extraction (detecting market events, participants, time and impact range).

[0073] Step three, knowledge graph construction: Based on the identified financial entity type information, relationship type information and attribute information, construct a dynamic knowledge graph.

[0074] Step four, event-driven dynamic updating mechanism: Real-time data monitoring of financial information flow and using rule engine to filter high relevance events. Extract new knowledge content entities and relationships, and incrementally update the knowledge graph to achieve the purpose of real-time knowledge base.

[0075] It should be noted that the dynamic knowledge graph can associate multi-dimensional data such as enterprise equity relationship and historical violation records, which is beneficial to discover implicit association results and causality between information.

[0076] Specifically, the updating process of the event-driven dynamic knowledge graph can be as follows:

[0077] In response to the monitoring of multi-source heterogeneous financial data flow, the multi-source heterogeneous financial data is preprocessed to obtain preprocessed multi-source heterogeneous financial data; the financial entity type information, the relationship information between the financial entities, and the financial event attribute information of the multi-source heterogeneous financial data are determined; and the current dynamic knowledge graph is dynamically maintained and processed according to the financial entity type information, the relationship information between the financial entities, and the financial event attribute information to obtain an updated dynamic knowledge graph.

[0078] It should be noted that the dynamic knowledge graph can synchronize institutional information and enterprise data in real time, reduce the lag risk of compliance checks, and respond to new rules in a timely manner.

[0079] It should also be noted that the associated information is the relevant field knowledge and risk analysis specification information of the structured financial data retrieved from the dynamic knowledge graph.

[0080] Optionally, a graph retrieval augmented generation request is generated according to the structured financial data, and the graph retrieval augmented generation request is sent to the dynamic knowledge graph to retrieve a dynamic knowledge graph subgraph corresponding to the structured financial data and risk analysis specification information from the dynamic knowledge graph, and relevant field information is determined according to the dynamic knowledge graph subgraph, and the relevant field information and the risk analysis specification information are taken as associated information.

[0081] It should be noted that the above process is mainly based on GraphRAG (Graph Retrieval-Augmented Generation) to realize the combination of the dynamic knowledge graph and RAG (Retrieval-Augmented Generation) to retrieve the dynamic knowledge graph stored in the graph database, obtain associated information and realize augmented generation.

[0082] The retrieval process is mainly triggered by a graph retrieval augmented request generated according to structured financial data, and the graph retrieval augmented request is mainly used to query field information and risk specification information related to the structured financial data from the dynamic knowledge graph. When the graph retrieval augmented request is generated according to the structured financial data, the key information such as the financial entity in the structured financial data is first identified, and then the financial entity and the key information are converted into nodes in the dynamic knowledge graph, and are connected through relationship edges to form a query picture segment, which is the graph retrieval augmented request.

[0083] It should be further noted that in the dynamic knowledge graph updating process, after the financial entity type information, the relationship information between the financial entities, and the financial event attribute information of the multi-source heterogeneous financial data are determined, before the current dynamic knowledge graph is dynamically maintained according to the financial entity type information, the relationship information between the financial entities, and the financial event attribute information to obtain the updated dynamic knowledge graph, the financial entity type information is also compared with the financial entity information in the current dynamic knowledge graph. When the financial entity type information and the financial entity information have the same financial entity, the financial entity type information and the financial entity information are co-reference resolution processed and credibility weighted to determine the financial entity type information or the financial entity information as the standard description information of the financial entity.

[0084] Specifically, when the same financial entity exists in the financial entity to be updated as in the current dynamic knowledge graph, that is, the same financial entity has different expressions, a standard expression is determined for the financial entity by using a credibility weighting method.

[0085] Step 103, generating a risk analysis instruction command according to the structured financial data, the relevant field knowledge and the risk analysis specification information, inputting the risk analysis instruction command into the generative pre-training model to obtain a risk analysis result.

[0086] It should be noted that the risk analysis instruction command can be specifically generated based on the structured financial data, the related field knowledge, and the risk analysis specification information to generate the prompt word content input into the generative pre-training model.

[0087] Optionally, the prompt word generated based on the structured financial data, the related field knowledge, and the risk analysis specification information is input into the generative pre-training model for risk analysis to obtain a risk analysis result.

[0088] The risk analysis result can be specifically a structured report including risk point positioning information, traceable evidence chain, associated case information, and risk repair suggestion information.

[0089] It should be noted that the corresponding evidence chain in the output risk analysis result is specifically associated with a dynamic knowledge graph, which can improve the trust of a user in the financial document review provided by the application.

[0090] In summary, the financial document review method provided by the application processes the multi-modal financial document to be reviewed into structured financial data, retrieves the related field knowledge and risk analysis specification information of the structured financial data from the dynamic knowledge graph, generates the risk analysis instruction command based on the structured financial data, the related field knowledge, and the risk analysis specification information, and inputs the risk analysis instruction command into the generative pre-training model to obtain the risk analysis result. This process does not require human intervention for review, greatly improving the review efficiency. Moreover, the indication information for risk analysis is generated based on the latest related field knowledge and risk analysis specification information retrieved from the dynamic knowledge graph, so that a more accurate risk analysis result of the structured financial data can be obtained in real time.

[0091] The above introduces a financial document review method provided by an embodiment of the application. The following will introduce a device for executing the above financial document review method.

[0092] Please refer to Figure 2 , Figure 2 FIG. 1 is a structural schematic diagram of a financial document review device provided by the application. As shown in Figure 2 the figure, the device comprises:

[0093] a structured processing unit 10, a query unit 20, and a risk analysis unit 30; wherein:

[0094] The structured processing unit 10 is configured to perform structured processing on the multi-modal financial document to be reviewed input by a client to obtain structured financial data.

[0095] The query unit 20 is configured to query associated information matched with the structured financial data from a pre-established dynamic knowledge graph; the pre-established dynamic knowledge graph is a knowledge graph in which financial entity information and association relationships between the financial entities are stored in a graph structure; the associated information includes domain knowledge related to the structured financial data and risk analysis specification information;

[0096] The risk analysis unit 30 is configured to generate a risk analysis instruction command according to the structured financial data, the domain knowledge and the risk analysis specification information, input the risk analysis instruction command into the generative pre-training model, and obtain a risk analysis result.

[0097] In an embodiment, the query unit 20 is specifically configured to:

[0098] generate a graph retrieval enhancement generation request according to the structured financial data, and send the graph retrieval enhancement generation request to the dynamic knowledge graph;

[0099] retrieve a dynamic knowledge graph subgraph corresponding to the structured financial data and risk analysis specification information from the dynamic knowledge graph;

[0100] determine the domain information according to the dynamic knowledge graph subgraph;

[0101] use the domain information and the risk analysis specification information as the associated information.

[0102] In an embodiment, the updating process of the dynamic knowledge graph in the query unit 20 includes:

[0103] In response to listening to the inflow of the multi-source heterogeneous financial data, pre-processing the multi-source heterogeneous financial data to obtain pre-processed multi-source heterogeneous financial data;

[0104] determining financial entity type information, relationship information between financial entities and financial event attribute information of the multi-source heterogeneous financial data;

[0105] performing dynamic incremental maintenance processing on the current dynamic knowledge graph according to the financial entity type information, the relationship information between the financial entities and the financial event attribute information to obtain an updated dynamic knowledge graph.

[0106] In an embodiment, in the updating process of the dynamic knowledge graph in the query unit 20, after the financial entity type information, the relationship information between the financial entities and the financial event attribute information of the multi-source heterogeneous financial data are determined, before the dynamic incremental maintenance processing is performed on the current dynamic knowledge graph according to the financial entity type information, the relationship information between the financial entities and the financial event attribute information to obtain the updated dynamic knowledge graph, the process further includes:

[0107] comparing the financial entity type information with financial entity information in the current dynamic knowledge graph;

[0108] When the financial entity type information and the financial entity information are the same financial entity, the financial entity type information and the financial entity information are subjected to co-reference resolution processing and credibility weighting processing, and the financial entity type information or the financial entity information is determined as the standard description information of the financial entity.

[0109] In an embodiment, the structured processing unit 10 is specifically configured to:

[0110] The financial documents of different modalities in the to-be-inspected multi-modal financial document are parsed and processed into structured financial data with unified format.

[0111] In an embodiment of the present application, a computer program product is provided, which includes computer readable instructions. When the computer readable instructions are run on an electronic device, the electronic device implements any of the financial document inspection methods provided in the embodiments of the present application.

[0112] In an embodiment of the present application, a computer storage medium is provided, which carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the financial document inspection methods provided in the embodiments of the present application.

[0113] In an embodiment of the present application, a financial document inspection system is provided, which includes a multi-modal data parsing module, a dynamic knowledge graph module, and a risk analysis module.

[0114] The multi-modal data parsing module is configured to perform structured processing on the to-be-inspected multi-modal financial document input by the client to obtain structured financial data. The dynamic knowledge graph module is configured to construct and update a dynamic knowledge graph. The risk analysis module is configured to query related field knowledge and risk analysis specification information matching the structured financial data from the dynamic knowledge graph. A risk analysis instruction command is generated according to the structured financial data, the related field knowledge, and the risk analysis specification information. The risk analysis instruction command is input into a generative pre-training model to obtain a risk analysis result.

[0115] Optionally, referring to Figure 3 , the present application provides a structural composition diagram of the financial document inspection system.

[0116] As Figure 3 shown, the financial document inspection system includes a multi-modal data parsing module, a risk analysis module based on GraphRAG, a dynamic knowledge graph construction module, and a dynamic report generation module.

[0117] The multi-modal data analysis module is specifically configured to analyze and process multi-modal data such as text, images, and voice input by the user to obtain structured financial data in a unified format. The specific means of analysis and processing include image recognition and voice analysis.

[0118] The dynamic knowledge graph construction module is mainly used for the construction and updating of the dynamic knowledge graph. The construction process of the dynamic knowledge graph mainly includes three steps: multi-source heterogeneous data acquisition, extraction of entities and relationships between entities from multi-source heterogeneous data, and finally construction of the dynamic knowledge graph. The updating of the dynamic knowledge graph in the present application is mainly event-driven dynamic updating.

[0119] The risk analysis module based on GraphRAG is mainly used for risk analysis of the multi-modal analyzed user input combined with the domain knowledge in the knowledge graph using a generative pre-training model. Specifically, the multi-modal analyzed user input is processed, and if there is multi-modal data that needs to be aligned (for example, the indicators in the table need to be associated with the corresponding text description). Secondly, the knowledge graph is queried based on the GraphRAG strategy to obtain the domain knowledge and specifications related to the user input content. Then, the user input and the retrieved knowledge are processed to construct the prompt word content of the model input to enhance the context information. Finally, the processed prompt word is input into the model for risk analysis, and the model is required to generate a relevant evidence chain for the detected risk content.

[0120] The dynamic report generation module is mainly used for integrating the risk analysis results and automatically generating a structured report using a generative pre-training model, including risk point positioning, evidence chain, related cases, and repair suggestions, to enhance the explainability of the risk conclusion. At the same time, the user can trace the analysis process through natural language (such as "explain the source of a certain risk conclusion").

[0121] For example, referring to Figure 4 The above financial document review system provided by the present application provides an example of a review process diagram.

[0122] The multi-modal data analysis module receives the file to be reviewed from the user end. The file to be reviewed can be a multi-format financial document, which can include PDF, voice, scanned copies, and other forms.

[0123] The multi-modal data analysis module performs multi-modal analysis on the to-be-audited file, and different analysis processing is performed on different formats of document content in the to-be-audited file to obtain structured data. For example, a PDF scan of an enterprise credit contract is used to extract text and table data in the contract using OCR (Optical Character Recognition), and the signature in the scan is labeled as "contract effective condition"; the table in the contract is structured and analyzed, and if it is in voice, it needs to be converted to text.

[0124] Note that the GraphRAG-based risk analysis module initiates a joint retrieval request to the dynamic knowledge graph construction module and sends a graph retrieval enhancement generation request to the dynamic knowledge graph construction module, retrieves a subgraph from the dynamic knowledge graph, and returns the associated data to the GraphRAG-based risk analysis module.

[0125] The GraphRAG-based risk analysis module performs context enhancement analysis and multi-modal data alignment based on the associated data, generates an evidence chain using a generative pre-training model, and finally sends the generated risk analysis structure to the dynamic report generation module, which generates a traceable data chain including visual evidence chain links and associated historical similar cases. Specifically, it can be an interactive report containing a graph path evidence chain.

[0126] In addition, it should be noted that the apparatus embodiments described above are only schematic, and the units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. In addition, the connection relationship between the modules in the apparatus embodiment provided in the present application indicates that there is a communication connection between them, which can be implemented as one or more communication buses or signal lines.

[0127] Those skilled in the art can clearly understand that the application can be implemented by means of software plus necessary universal hardware, of course, also can be implemented by special hardware including special integrated circuit, special CPU, special memory, special component, etc. Generally, the functions completed by computer program can be easily implemented by corresponding hardware, and the specific hardware structure for implementing the same function can also be various, such as analog circuit, digital circuit or special circuit, etc. However, for the application, the software program implementation is the better embodiment. Based on such understanding, the technical solutions of the application can be embodied in the form of software product, which is stored in a readable storage medium, such as a floppy disk, a U disk, a mobile hard disk, a ROM, a RAM, a magnetic disk or an optical disk, etc., including a plurality of instructions for making a computer device (which can be a personal computer, a training device, or a network device, etc.) execute the methods described in various embodiments of the application.

[0128] In the above embodiments, all or part can be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part can be implemented in the form of a computer program product.

[0129] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, all or part of the processes or functions described in the embodiments of the application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, training device or data center to another website, computer, training device or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer storage medium can be any available medium that the computer can store or the data storage device such as training device, data center, etc. integrated with one or more available media sets. The available media can be magnetic media (such as floppy disk, hard disk, magnetic tape), optical media (such as DVD), or semiconductor media (such as solid state disk (SSD)), etc.

Claims

1. A method for reviewing financial documents, characterized in that, include: The multimodal financial documents input by the client for review are processed in a structured manner to obtain structured financial data; The system queries a pre-established dynamic knowledge graph to find related information that matches the structured financial data. The pre-established dynamic knowledge graph is a knowledge graph that stores information about financial entities and the relationships between financial entities in a graph structure. The related information includes relevant domain knowledge and risk analysis standard information of the structured financial data. Based on the structured financial data, relevant domain knowledge, and risk analysis standard information, a risk analysis instruction command is generated. The risk analysis instruction command is then input into a generative pre-trained model to obtain the risk analysis results.

2. The financial document review method according to claim 1, characterized in that, The step of querying the association information matching the structured financial data from the pre-established dynamic knowledge graph includes: A graph retrieval enhancement generation request is generated based on the structured financial data, and the graph retrieval enhancement generation request is sent to the dynamic knowledge graph; Retrieve from the dynamic knowledge graph subgraphs and risk analysis specification information corresponding to the structured financial data; Relevant domain information is determined based on the dynamic knowledge graph subgraph; The relevant field information and the risk analysis specification information are used as the associated information.

3. The financial document review method according to claim 1, characterized in that, The update process of the dynamic knowledge graph includes: In response to the detection of multi-source heterogeneous financial data inflow, the multi-source heterogeneous financial data is preprocessed to obtain preprocessed multi-source heterogeneous financial data. Determine the financial entity type information, the relationship information between financial entities, and the financial event attribute information of the multi-source heterogeneous financial data; Based on the information on financial entity types, relationships between financial entities, and attributes of financial events, the current dynamic knowledge graph is dynamically incrementally maintained to obtain an updated dynamic knowledge graph.

4. The financial document review method according to claim 3, characterized in that, After determining the financial entity type information, inter-entity relationship information, and financial event attribute information of the multi-source heterogeneous financial data, and before performing dynamic incremental maintenance processing on the current dynamic knowledge graph based on the financial entity type information, inter-entity relationship information, and financial event attribute information to obtain the updated dynamic knowledge graph, the process further includes: Compare the financial entity type information with the financial entity information in the current dynamic knowledge graph; When the financial entity type information and the financial entity information belong to the same financial entity, the financial entity type information and the financial entity information are subjected to common reference resolution processing and credibility weighting processing to determine that the financial entity type information or the financial entity information is the standard description information of the financial entity.

5. The financial document review method according to claim 1, characterized in that, The risk analysis instructions are input into the generative pre-trained model to obtain risk analysis results, including: The risk analysis instructions are input into a generative pre-trained model to obtain a structured report; the structured report includes risk point location information, traceable evidence chain, related case information, and risk remediation suggestions.

6. The financial document review method according to claim 1, characterized in that, The process of structuring the multimodal financial documents input by the client to be reviewed, resulting in structured financial data, includes: The financial documents of different modalities in the multimodal financial documents to be reviewed are parsed and processed into structured financial data with a unified format.

7. A financial document review device, characterized in that, include: The structured processing unit is used to perform structured processing on the multimodal financial documents to be reviewed input by the client, so as to obtain structured financial data. The query unit is used to query the associated information that matches the structured financial data from a pre-established dynamic knowledge graph; the pre-established dynamic knowledge graph is a knowledge graph that stores financial entity information and the relationships between financial entities in a graph structure; the associated information includes relevant domain knowledge and risk analysis standard information of the structured financial data; The risk analysis unit is used to generate risk analysis instructions based on the structured financial data, the relevant domain knowledge, and risk analysis specification information, and input the risk analysis instructions into the generative pre-trained model to obtain risk analysis results.

8. A computer program product, characterized in that, It includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to perform the financial document review method as described in any one of claims 1 to 6.

9. A computer storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the financial document review method as described in any one of claims 1 to 6.

10. A financial document review system, characterized in that, include: Multimodal data parsing module, dynamic knowledge graph module, and risk analysis module; The multimodal data parsing module is used to perform structured processing on the multimodal financial documents to be reviewed input by the client, and obtain structured financial data. The dynamic knowledge graph module is used to construct and update the dynamic knowledge graph; The risk analysis module is used to query relevant domain knowledge and risk analysis standard information that match the structured financial data from the dynamic knowledge graph; generate risk analysis instruction commands based on the structured financial data, the relevant domain knowledge, and the risk analysis standard information; input the risk analysis instruction commands into the generative pre-trained model to obtain risk analysis results.

Images