Financial scene-oriented structured parameter output method and device based on large model

Abstract

The invention provides a financial scene-oriented structured parameter output method and device based on a large model, and relates to the technical field of computers. According to the method, accurate preconditions are provided for subsequent processing through document type identification; through a strategy planning agent, a parameter extraction planning problem is optimally modeled, an extraction path is adaptively planned according to the adaptation degree of parameters and an extraction method, and mixed arrangement of multiple technical means is realized; the intelligent agent is extracted through multiple paths, the accuracy of a regular expression, the knowledge retrieval capability of a retrieval enhancement generation technology and the logical reasoning capability of a large model are comprehensively utilized, and deep understanding of financial business parameters is achieved; and carrying out consistency verification on the preliminary structured parameter data, and carrying out conflict resolution processing to obtain a final structured parameter result, and outputting the final structured parameter result. According to the scheme, the accuracy and robustness of an extraction result are improved while the calculation cost is reduced, and the problems of semantic understanding and logical reasoning under a complex format are solved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and apparatus for structured parameter output based on a large model for financial scenarios. Background Technology

[0002] With the continuous improvement of the digitalization and intelligence of financial services, enhancing the service level of financial knowledge platforms has become an objective necessity for development, and these platforms often rely on structured data. Knowledge and business data for financial scenarios typically exist in the form of natural language text and semi-structured documents, and their content is complex, highly specialized, and expressed in diverse ways, posing significant challenges to traditional rule-based or template-based parameter extraction methods.

[0003] Existing methods for extracting structured financial knowledge data have the following drawbacks: First, these methods rely on fixed rules and templates, such as hard-coded regular expressions or predefined coordinate templates. They are only suitable for highly standardized documents and have poor generalization ability when dealing with complex financial documents with flexible layouts and high degrees of unstructuredness. Second, they lack the ability to integrate multimodal features and long documents. Most research focuses on text flow parsing, neglecting visual layout semantics, and easily disrupts global semantic coherence when processing long documents. Third, end-to-end large-model extraction methods have poor confidence and suffer from the "illusion" problem. They are prone to generating incorrect values ​​for precise data such as amounts and dates, and using large-model inference across all data leads to enormous computational costs. Here, "large model," also known as a Large Language Model (LLM), has three key characteristics: first, large scale, with network parameters reaching tens or hundreds of billions or even more; second, generality, meaning it is not limited to specific problems or domains; and third, emergence, meaning it generates unexpected new capabilities.

[0004] In recent years, large-scale models have made breakthroughs in natural language processing and other fields, providing new technical approaches for semantic parsing and information extraction from complex financial texts. However, existing large-scale models lack fine-grained control over financial business parameters, constraints, and output formats, making it difficult to accurately output structured parameter results. Therefore, how to fully leverage the semantic understanding capabilities of large-scale models while ensuring the accuracy and consistency of financial data, and to achieve structured data output for financial scenarios, has become a key technical problem that financial institutions urgently need to solve. Summary of the Invention

[0005] In view of the above problems, this application is proposed to provide a structured parameter output method and apparatus based on a large model for financial scenarios that overcomes or at least partially solves the above problems. The technical solution is as follows: Firstly, a structured parameter output method based on a large model for financial scenarios is provided, the method comprising: Acquire multi-source heterogeneous financial documents, identify document types of the financial documents, and obtain the financial document types; Determine the set of parameter fields to be extracted based on the financial document type; By using a strategy planning agent, a target extraction method is planned for each parameter field in the parameter field set, thus obtaining the extraction strategy; By using a multi-path extraction agent, extraction operations are performed by calling the target extraction method according to the extraction strategy to obtain the field value corresponding to each parameter field. Based on all parameter fields and the field value corresponding to each parameter field, preliminary structured parameter data is generated. The preliminary structured parameter data includes the target extraction method and source context information used for the field value corresponding to each parameter field. Based on a pre-built financial business logic constraint graph corresponding to financial document types, the consistency of the preliminary structured parameter data is verified. If the consistency check passes, the preliminary structured parameter data will be used as the final structured parameter result and output. If the consistency check fails, the conflict fields that failed the check are resolved to obtain the final structured parameter result and output it.

[0006] In one possible implementation, consistency checks are performed on the initial structured parameter data based on a pre-built financial business logic constraint graph corresponding to the financial document type, including: Obtain a pre-built financial business logic constraint graph corresponding to the financial document type. The financial business logic constraint graph includes multiple nodes and multiple edges. Each node represents a parameter field, and each edge represents the business logic relationship between two parameter fields. Traverse each edge in the financial business logic constraint graph and perform validation operations corresponding to the business logic relationship on the field values ​​of the parameter fields associated with the edges in the preliminary structured parameter data; If all verification operations pass, then the consistency check is considered successful. If any verification operation fails, the consistency verification is determined to have failed, and the conflicting fields and conflict types of the failed verifications are recorded.

[0007] In one possible implementation, if the consistency check fails, the conflicting fields that failed the check are resolved to obtain the final structured parameter result, which is then output, including: Based on the source context information, locate the original text fragment in the financial document corresponding to the conflicting field; For conflicting fields, switch the extraction strategy to perform partial re-extraction on the original text fragment to obtain the re-extracted field values; The re-extracted field values ​​are substituted into the initial structured parameter data for updating, and the updated initial structured parameter data is checked for consistency again. If the verification passes again, the updated preliminary structured parameter data will be determined as the final structured parameter result. If the validation fails again, a conflict warning label and confidence score are added to the conflict field, and the preliminary structured parameter data after adding the label is determined as the final structured parameter result.

[0008] In one possible implementation, the extraction strategy includes any one of the following: regular expression extraction strategy, retrieval enhancement generation extraction strategy, and large model inference extraction strategy. For conflicting fields, the extraction strategy is switched to perform partial re-extraction on the original text fragment, resulting in the re-extracted field values, including: If the extraction strategy for conflicting fields is the large model inference extraction strategy, then switch the extraction strategy for conflicting fields to the retrieval enhancement generation extraction strategy or the regular expression extraction strategy, and perform local re-extraction on the original text fragment to obtain the re-extracted field values. If the extraction strategy for conflict fields is a retrieval enhancement generation extraction strategy, then the retrieval context window is expanded, or the constraints corresponding to the conflict type are input into the large model as part of the prompt words, and local re-extraction is performed on the original text fragment to obtain the re-extracted field values.

[0009] In one possible implementation, document type identification is performed on financial documents to obtain financial document types, including: Financial documents are processed using optical character recognition (OCR) to obtain a set of text information. Based on the preset document type system and text information set, a preliminary classification process is performed through rule matching to obtain the preliminary classification result and the confidence level corresponding to the preliminary classification result; If the confidence level is greater than or equal to the preset threshold, the primary classification result will be determined as a financial document type. If the confidence level is less than the preset threshold, the text information set is input into the large model for secondary classification processing to obtain the classification result of the large model, and the classification result of the large model is determined as a financial document type.

[0010] In one possible implementation, the text information set includes the document title and a text sequence; Based on a pre-defined document type system and text information set, a preliminary classification process is performed using rule matching to obtain the preliminary classification results and their corresponding confidence levels, including: For each candidate document type in the preset document type system, the document title is matched with the standard title of the candidate document type to obtain the title matching degree; The keyword coverage rate is determined based on the number of preset business feature keywords corresponding to the candidate document types contained in the text sequence; The title matching degree and keyword coverage are weighted and fused to obtain the confidence degree corresponding to the candidate document type; The candidate document type with the highest confidence level is determined as the primary classification result, and the confidence level corresponding to the candidate document type with the highest confidence level is determined as the confidence level corresponding to the primary classification result.

[0011] In one possible implementation, a policy planning agent plans a target extraction method for each parameter field in the parameter field set, resulting in an extraction policy, including: By using a strategy planning agent, each parameter field in the parameter field set is matched and scored with multiple preset candidate extraction methods to obtain a comprehensive matching score for each parameter field relative to each candidate extraction method. By using a policy planning agent, based on multiple comprehensive adaptation scores corresponding to each parameter field, and with the goal of maximizing the sum of the comprehensive adaptation scores corresponding to the set of parameter fields, a candidate extraction method is assigned to each parameter field as the target extraction method, thus obtaining the extraction strategy.

[0012] In one possible implementation, the preset multiple candidate extraction methods include regular expression extraction, retrieval-enhanced generation extraction, and large model inference extraction. By using a multi-path extraction agent, the extraction operation is performed by calling the target extraction method according to the extraction strategy, and the field values ​​corresponding to each parameter field are obtained, including: If the target extraction method for the parameter field is regular expression extraction, then the multi-path extraction agent will match the text sequence of the financial document according to the regular expression corresponding to the parameter field to obtain the field value. If the target extraction method for the parameter field is retrieval-enhanced generation extraction, then a multi-path extraction agent is used to retrieve a set of relevant text fragments from the text sequence of the financial document based on the semantics of the parameter field; the parameter field and the set of relevant text fragments are then input into a large model for processing to obtain the field value. If the target extraction method for the parameter fields is the large model inference extraction method, then the agent is extracted through multiple paths, and an inference thought chain is constructed based on the parameter fields; the inference thought chain and the text sequence of the financial document are input into the large model for inference processing to obtain the field values.

[0013] In one possible implementation, the set of parameter fields to be extracted is determined based on the financial document type, including: Based on the preset association between document types and parameter field sets, determine the set of parameter fields to be extracted that are associated with financial document types.

[0014] Secondly, a structured parameter output device based on a large model for financial scenarios is provided, the device comprising: The identification unit is used to acquire multi-source heterogeneous financial documents, identify the document type of the financial documents, and obtain the financial document type. The determination unit is used to determine the set of parameter fields to be extracted based on the financial document type; The planning unit is used to plan the target extraction method for each parameter field in the parameter field set through the strategy planning agent, so as to obtain the extraction strategy. The extraction and generation unit is used to extract the agent through multiple paths. According to the extraction strategy, it calls the target extraction method to perform the extraction operation and obtain the field value corresponding to each parameter field. Based on all parameter fields and the field value corresponding to each parameter field, it generates preliminary structured parameter data. The preliminary structured parameter data includes the target extraction method and source context information used for the field value corresponding to each parameter field. The verification output unit is used to perform consistency verification on the preliminary structured parameter data based on the pre-built financial business logic constraint graph corresponding to the financial document type. If the consistency verification passes, the preliminary structured parameter data is output as the final structured parameter result. If the consistency verification fails, the conflict fields that failed the verification are resolved to obtain the final structured parameter result and output it.

[0015] By employing the aforementioned technical solutions, this application provides a structured parameter output method and apparatus based on a large model for financial scenarios. This method provides accurate preconditions for subsequent processing through document type recognition; it plans target extraction methods for each parameter field in the parameter field set through a strategy planning agent, adaptively planning extraction paths based on the fit between parameters and extraction methods to obtain extraction strategies, achieving a hybrid arrangement of multiple technical means; through a multi-path extraction agent, it comprehensively utilizes the accuracy of regular expressions, the knowledge retrieval capabilities of retrieval enhancement generation technology, and the logical reasoning capabilities of the large model to achieve a deep understanding of financial business parameters and output preliminary structured parameter data; it performs consistency verification of the preliminary structured parameter data through a financial business logic constraint graph and performs conflict resolution processing to obtain and output the final structured parameter results. This solution reduces computational costs while improving the accuracy and robustness of the extraction results, solving the challenges of semantic understanding and logical reasoning under complex formats. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below.

[0017] Figure 1 This document illustrates a flowchart of a structured parameter output method based on a large model for financial scenarios, provided in an embodiment of this application. Figure 2 This paper illustrates the technical architecture of a large-model-based structured parameter outputter for financial scenarios provided in an embodiment of this application. Figure 3 This application provides a hierarchical diagram of financial document classification. Figure 4 This paper shows a structural diagram of a large-model-based structured parameter output device for financial scenarios provided in an embodiment of this application. Figure 5 A structural diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0018] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0019] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the term "comprising" and its variations should be interpreted as open-ended terms meaning "including but not limited to."

[0020] To address the aforementioned technical problems, embodiments of this application provide a structured parameter output method based on a large model for financial scenarios, such as... Figure 1 As shown, the method may include the following steps S101 to S105: Step S101: Obtain multi-source heterogeneous financial documents, identify the document type of the financial documents, and obtain the financial document type; the financial documents here may include original electronic documents, scanned copies, non-standardized documents with stamps or handwritten annotations, etc., and this embodiment does not limit them; Step S102: Determine the set of parameter fields to be extracted based on the financial document type; Step S103: Through the strategy planning agent, a target extraction method is planned for each parameter field in the parameter field set to obtain the extraction strategy; Step S104: The agent extracts data through a multi-path approach. Based on the extraction strategy, the target extraction method is called to perform the extraction operation, obtaining the field value corresponding to each parameter field. Preliminary structured parameter data is generated based on all parameter fields and their corresponding field values. This preliminary structured parameter data includes the target extraction method and source context information used for the field value corresponding to each parameter field. The extraction strategy can include any one of the following: regular expression extraction strategy, retrieval enhancement generation extraction strategy, or large model inference extraction strategy. Step S105: Based on the pre-constructed financial business logic constraint graph corresponding to the financial document type, perform consistency verification on the preliminary structured parameter data; if the consistency verification passes, output the preliminary structured parameter data as the final structured parameter result; if the consistency verification fails, resolve the conflict fields that failed the verification to obtain the final structured parameter result and output it.

[0021] Here, the final structured parameter result may include the field identifier, field name, field value, identifier for identifying the target extraction method, and source context information for indicating the source of the field value, etc. This embodiment does not limit this.

[0022] This embodiment uses document type recognition to provide accurate preconditions for subsequent processing; it employs a strategy planning agent to plan target extraction methods for each parameter field in the parameter field set, adaptively planning extraction paths based on the compatibility between parameters and extraction methods to obtain extraction strategies, achieving a hybrid arrangement of multiple technical means; through a multi-path extraction agent, it comprehensively utilizes the accuracy of regular expressions, the knowledge retrieval capabilities of retrieval enhancement generation technology, and the logical reasoning capabilities of large models to achieve a deep understanding of financial business parameters and output preliminary structured parameter data; it uses a financial business logic constraint graph to perform consistency verification on the preliminary structured parameter data and handle conflict resolution to obtain and output the final structured parameter results. This solution reduces computational costs while improving the accuracy and robustness of extraction results, solving the challenges of semantic understanding and logical reasoning under complex layouts.

[0023] This application embodiment provides a possible implementation method. In step S101, document type identification is performed on the financial document to obtain the financial document type, which may specifically include the following steps A1 to A4: Step A1: Perform Optical Character Recognition (OCR) processing on the financial document to obtain a set of text information; Step A2: Based on the preset document type system and text information set, perform primary classification processing through rule matching to obtain the primary classification result and the confidence level corresponding to the primary classification result; Step A3: If the confidence level is greater than or equal to the preset threshold, the primary classification result is determined to be a financial document type. Step A4: If the confidence level is less than the preset threshold, the text information set is input into the large model for secondary classification processing to obtain the classification result of the large model, and the classification result of the large model is determined to be a financial document type.

[0024] This embodiment employs a two-level classification method: "rule matching first, large model as a fallback." For documents with clearly defined rules, lightweight rule matching is used to quickly complete the classification, saving computational resources. For complex or marginal documents that rules cannot accurately determine, a large model is invoked to leverage its powerful semantic understanding capabilities for accurate classification. This mechanism significantly reduces the frequency of calling the large model while ensuring classification accuracy, thereby improving overall processing efficiency.

[0025] This application provides a possible implementation method, where the text information set includes document titles and text sequences; step A2 performs primary classification processing based on a preset document type system and the text information set using rule matching to obtain the primary classification result and the corresponding confidence level, which may specifically include the following steps A2-1 to A2-4: Step A2-1: For each candidate document type in the preset document type system, match the document title with the standard title of the candidate document type to obtain the title matching degree; Step A2-2: Determine the keyword coverage rate based on the number of preset business feature keywords corresponding to the candidate document types contained in the text sequence; Step A2-3: Perform weighted fusion processing on title matching degree and keyword coverage to obtain the confidence degree corresponding to the candidate document type; Step A2-4: Determine the candidate document type with the highest confidence as the primary classification result, and determine the confidence level corresponding to the candidate document type with the highest confidence as the confidence level corresponding to the primary classification result.

[0026] This embodiment calculates classification confidence by fusing features from two dimensions: title matching degree and keyword coverage rate, making the initial classification results more robust and reliable. Title matching utilizes document naming information, while keyword coverage rate utilizes the document's body content. These two methods complement each other, providing a more comprehensive reflection of the document's type characteristics and improving the accuracy and robustness of rule matching.

[0027] This application embodiment provides a possible implementation method. Step S102 determines the set of parameter fields to be extracted based on the financial document type, which may specifically include the following step B1: Step B1: Based on the preset association between document types and parameter field sets, determine the set of parameter fields to be extracted that are associated with the financial document type.

[0028] This embodiment defines the association between document types and the set of parameter fields to be extracted, making the parameter extraction task objective clear and avoiding invalid searches and extractions on all irrelevant fields, thus improving the targeting and efficiency of the extraction.

[0029] This application embodiment provides a possible implementation method. Step S103 uses a policy planning agent to plan a target extraction method for each parameter field in the parameter field set to obtain an extraction strategy. Specifically, it may include the following steps C1 and C2: Step C1: Through the strategy planning agent, each parameter field in the parameter field set is matched and scored with a variety of preset candidate extraction methods to obtain the comprehensive matching score of each parameter field relative to each candidate extraction method. Step C2: Through the policy planning agent, based on multiple comprehensive adaptation scores corresponding to each parameter field, with the optimization objective of maximizing the sum of comprehensive adaptation scores corresponding to the set of parameter fields, a candidate extraction method is assigned to each parameter field as the target extraction method, thus obtaining the extraction strategy.

[0030] This embodiment models the parameter extraction planning problem as an optimization problem. The strategy planning agent calculates the fit score of each parameter field with different extraction methods and plans with the goal of maximizing the global fit. This ensures that the most suitable extraction method is assigned to each field, thereby achieving the best balance between extraction accuracy and efficiency as a whole.

[0031] This application embodiment provides a possible implementation method, with preset multiple candidate extraction methods including regular expression extraction, retrieval enhancement generation extraction, and large model inference extraction; in step S104, the agent is extracted through multi-path extraction, and the target extraction method is called to perform extraction operations according to the extraction strategy to obtain the field value corresponding to each parameter field, which may specifically include any one of the following steps D1 to D3: Step D1: If the target extraction method for the parameter field is regular expression extraction, then the multi-path extraction agent will match the text sequence of the financial document according to the regular expression corresponding to the parameter field to obtain the field value.

[0032] Step D2: If the target extraction method for the parameter field is retrieval-enhanced generation extraction, then a multi-path extraction agent is used to retrieve a set of relevant text fragments from the text sequence of the financial document based on the semantics of the parameter field; the parameter field and the set of relevant text fragments are input into the large model for processing to obtain the field value.

[0033] Step D3: If the target extraction method for the parameter fields is the large model inference extraction method, then extract the agent through multiple paths, construct the inference thought chain based on the parameter fields, and input the inference thought chain and the text sequence of the financial document into the large model for inference processing to obtain the field values.

[0034] This embodiment employs a multi-path extraction agent to uniformly schedule three different extraction strategies. Regular expression extraction ensures accurate matching of structurally stable fields, enhanced retrieval generation solves the problem of integrating scattered information in long documents, and large-scale model inference overcomes the challenge of extracting implicit parameters that require complex logical derivation. Through this "each performing its own function and working together" architecture, high-quality extraction of various complex parameters from financial documents is achieved.

[0035] This application embodiment provides a possible implementation method. In step S105, based on a pre-constructed financial business logic constraint graph corresponding to the financial document type, a consistency check is performed on the preliminary structured parameter data. Specifically, this may include the following steps E1 and E2: Step E1: Obtain a pre-constructed financial business logic constraint graph corresponding to the financial document type. The financial business logic constraint graph includes multiple nodes and multiple edges. Each node represents a parameter field, and each edge represents the business logic relationship between two parameter fields. Step E2: Traverse each edge in the financial business logic constraint graph and perform validation operations corresponding to the business logic relationship on the field values ​​of the parameter fields associated with the edge in the preliminary structured parameter data. If all validation operations pass, the consistency validation is determined to be successful. If any validation operation fails, the consistency validation is determined to be unsuccessful, and the conflicting fields and conflict types that failed the validation are recorded.

[0036] This embodiment constructs a financial business logic constraint graph, solidifying professional financial business rules in the form of a graph structure. By traversing the edges in the graph and automatically verifying the extracted results, cross-field logical contradictions can be efficiently and accurately identified, such as numerical calculation errors or reversed date order, providing a precise basis for subsequent conflict resolution.

[0037] This application embodiment provides a possible implementation method. In step S105, if the consistency check fails, the conflict fields that failed the check are processed to resolve the conflict, and the final structured parameter result is obtained and output. Specifically, it may include the following steps F1 to F3: Step F1: Based on the source context information, locate the original text fragment in the financial document corresponding to the conflicting field; Step F2: For conflicting fields, switch the extraction strategy to perform local re-extraction on the original text fragment to obtain the re-extracted field values; Step F3 involves substituting the re-extracted field values ​​into the initial structured parameter data for updating, and then performing a consistency check on the updated initial structured parameter data again. If the check passes again, the updated initial structured parameter data is determined as the final structured parameter result. If the check fails again, conflict warning labels and confidence scores are added to the conflicting fields, and the initial structured parameter data with the added labels is determined as the final structured parameter result.

[0038] This embodiment achieves intelligent conflict resolution through a closed-loop mechanism of "source tracing - re-extraction - re-verification" upon detection. This mechanism only performs partial re-extraction on the conflicting fields, avoiding reprocessing the entire document and significantly improving correction efficiency. For conflicts that cannot be automatically resolved, warning labels and confidence scores are added to ensure that downstream businesses are aware of data quality risks.

[0039] This application provides a possible implementation method, where the extraction strategy includes any one of the following: regular expression extraction strategy, retrieval enhancement generation extraction strategy, and large model inference extraction strategy; step F2, for conflicting fields, switches the extraction strategy to perform local re-extraction on the original text fragment to obtain the re-extracted field value, which may specifically include any one of the following steps F2-1 and F2-2: Step F2-1: If the extraction strategy for conflicting fields is the large model inference extraction strategy, then switch the extraction strategy for conflicting fields to the retrieval enhancement generation extraction strategy or the regular expression extraction strategy, and perform local re-extraction on the original text fragment to obtain the re-extracted field values. Step F2-2: If the extraction strategy for the conflict field is the retrieval enhancement generation extraction strategy, then expand the retrieval context window, or input the constraint conditions corresponding to the conflict type as part of the prompt words into the large model, and perform local re-extraction on the original text fragment to obtain the re-extracted field values.

[0040] This embodiment designs differentiated re-extraction strategies for different original extraction methods. For high-risk large model inference results, it switches to more rigorous retrieval enhancement generation or regular expression extraction, effectively reducing the risk of "illusion." For retrieval enhancement generation results, by expanding the context or adding constraints, it guides the model to generate corrected values ​​that better conform to business logic. This differentiated processing approach improves the targeting and success rate of conflict resolution.

[0041] The above introduces Figure 1The embodiments shown have various implementation methods for each step. The following will further explain the structured parameter output method based on a large model for financial scenarios in this application through specific embodiments.

[0042] This specific embodiment proposes a structured parameter outputter based on a large model for financial scenarios, such as... Figure 2 As shown, it can achieve high-precision structured extraction of multi-source heterogeneous financial documents, solve the problems of semantic understanding and logical reasoning under complex layouts, and improve the efficiency and accuracy of automated processing of financial business. First, this embodiment proposes a two-level classification method for financial documents based on OCR parsing and a large model, identifying document types according to a pre-defined financial document classification system. Next, it proposes an intelligent agent parameter extraction planning mechanism based on optimization modeling, which can adaptively plan extraction paths based on the fit between parameters and extraction methods by optimizing the parameter extraction planning problem. Then, it introduces a structured parameter output architecture based on a multi-path extraction agent, comprehensively utilizing the accuracy of regular expressions, the knowledge retrieval capabilities of RAG (Retrieval Augmented Generation) technology, and the logical reasoning capabilities of a large model to achieve a deep understanding of financial business parameters and preliminary structured parameter data output. Next, it introduces a post-validation and conflict resolution module based on financial logic constraints, constructing a financial business logic constraint graph to perform consistency verification on the extraction results and intelligently resolve and re-extract conflicting results. Finally, it converts unstructured financial text into high-quality JSON (JavaScript object notation) structured data that conforms to business logic. The structured parameter outputter mentioned above reduces computational costs while improving the accuracy and robustness of the extraction results. A detailed introduction follows.

[0043] 1) A two-level classification method for financial documents based on OCR parsing and a large-scale model Because financial documents differ significantly in business type, content structure, and parameter semantics, the parameter definition system, extraction rules, and processing strategies also vary depending on the document type. Directly extracting parameters without clearly defining the document type can easily lead to confusion in field definitions and inappropriate selection of extraction methods, thus affecting the accuracy and consistency of the structured parameter extraction results. Therefore, this embodiment first parses and classifies the financial documents before parameter extraction, placing them into a pre-defined financial document type system. This ensures that the subsequent determination of the parameter field set, the planning of the extraction method, and the execution of parameter extraction are all based on a clear document type premise, providing the necessary preconditions for the subsequent process.

[0044] Pre-setting a financial document type system helps to clarify the required parameters for each document, thus providing a unified and stable basis for subsequent parameter extraction planning and execution processes. This embodiment first selects 12 typical types of financial documents, including account opening application forms, transaction orders, transaction confirmations, margin trading contracts, research reports, risk disclosure statements, fund contracts, anti-XX statements, annual audit reports, portfolio statements, prospectuses, and application reports; and constructs a hierarchical classification system for financial documents, such as... Figure 3 As shown, this system contains a set of first-level classifications. ={Client Accounts, Transaction Execution, Contracts & Agreements, Research Reports, Compliance & Regulatory, Financial Audit, Investment Banking IPO (Initial Public Offering)}, i.e., the first level; and the sub-sets of the above categories ={Account Opening Application Form, Transaction Order, Transaction Confirmation, Margin Trading Contract, Research Report, Risk Disclosure Statement, Fund Contract, Anti-XX Statement, Annual Audit Report, Position Statement, Prospectus, Application Report}, i.e., the second level; and for each document type, a predefined set of strong characteristic business keywords. .

[0045] See Figure 2 For multi-source heterogeneous financial documents uploaded by users or obtained from the system, OCR technology is used to recognize the text information of the financial documents, transforming the unstructured documents into the following set of text information I:

[0046] in, Indicates the document title. Represents a text sequence.

[0047] Based on this, according to the text information set I, the first-level classification is performed based on rule matching.

[0048] Will The standard title of the m-th category in the hierarchical classification system Perform string matching to obtain the title matching score. :

[0049] in, The length of the string representing the document title; The string length indicating the standard heading; It is a classic algorithm function for calculating edit distance, which is a non-negative integer. The larger the value, the greater the difference between the two texts. The closer it is to 1, the higher the title matching degree.

[0050] Simultaneously, the number of keywords contained in the OCR-recognized text sequence is counted based on the preset keyword set of the m-th category. Calculate keyword coverage score :

[0051] Finally, a multi-dimensional feature weighted fusion method is used to calculate the confidence score of whether a document belongs to the m-th category. :

[0052] in, , Each feature is assigned a weight coefficient, and the sum of the two is 1. The calculated confidence scores for each category are compared, and the category with the highest confidence score is determined as the correct classification result. Its confidence level is .

[0053] Introducing an adaptive threshold ,if The classification results will be... It is considered the final result; if Then, the text information set I is input into the large model for secondary classification processing, that is, the secondary classification of semantic understanding in the large model, to obtain the classification result of the large model. The classification results of the large model are then identified as financial document types; this determination logic can be expressed as:

[0054] By using the above two-level document classification method, the frequency of large model calls is significantly reduced while ensuring classification accuracy and result reliability, thus saving computing resources and improving overall processing efficiency.

[0055] 2) Agent parameter extraction planning based on optimization modeling Because financial transactions involve a wide variety of document types, the structured parameters to be extracted from different financial documents vary significantly in terms of expression, semantic complexity, and contextual dependencies. Therefore, by uniformly planning the appropriate parameter extraction methods for different parameter fields based on their characteristics, it is possible to improve the overall accuracy of structured parameter extraction and the efficiency of system operation while ensuring that the standardization and reliability requirements of financial scenarios are met.

[0056] After completing document parsing and classification, this embodiment models the parameter extraction process as an optimization problem using manually defined parameter extraction rules, and introduces a Strategic Planning Agent (SP-Agent) to solve the optimization model, thereby achieving unified planning for the parameter extraction process.

[0057] Suppose the user-defined set of structured parameters to be extracted is:

[0058] Three extraction strategies are defined: Regular Extraction Strategy (RES), Retrieval Enhancement Generation, and Large Model Reasoning Generation (LLMRG). The set of extraction strategies is M = {RES, RAG, LLMRG}. .

[0059] The parameter extraction rules are modeled as an optimization problem, with its decision variables... for:

[0060] And it satisfies the field unique selection constraint:

[0061] Its optimization objective is:

[0062] in, Representation field Extraction method The comprehensive adaptation score is determined by SP-Agent based on field structure stability, context dependency, and whether logical derivation is required, and is expressed as:

[0063] in, The score represents the structural fit between the field's structural features and the extraction method. A score indicating the fit between the degree of field context dependency and the ability of the extraction method; A score indicating the compatibility between the logical derivation and extraction method reasoning ability for field values; , , These are preset weighting coefficients.

[0064] By solving the optimization modeling problem, SP-Agent outputs a unique extraction method for each field:

[0065] Through the above planning mechanism, SP-Agent generates clear and executable extraction method decisions for each parameter, providing a clear strategic basis for subsequent parameter extraction and execution.

[0066] 3) Extracting structured parameters of the agent based on multi-path extraction After SP-Agent completes the parameter extraction method planning, this embodiment uses a multi-path extraction agent (MPE-Agent) to uniformly schedule and execute the structured parameter extraction task.

[0067] The three extraction strategies are executed as follows: (1) Regular expression extraction strategy For fields with high structural stability, weak context dependency, and no need for logical derivation, MPE-Agent can generate regular expressions either by using a pre-defined regular expression library corresponding to the parameters or by parsing the parameter information. Subsequently, the text sequence extracted from the OCR... By matching the parameter data, the final result is obtained:

[0068] (2) Search Enhancement Generation Strategy For parameter fields with strong context dependence or distributed across different paragraphs of a long document, MPE-Agent first obtains relevant information about the target parameter. Construct query vectors based on the semantics of the target fields. ; Text sequence The text is divided into a set B of blocks, which are then encoded into vectors using an embedding model. The K most relevant sub-text sets to the field are selected by calculating cosine similarity.

[0069]

[0070] Subsequently, MPE-Agent inputs the search results and field suggestions into the large model to generate the final parameter results:

[0071] (3) Large model inference generation strategy For implicit parameters or parameters that need to be obtained through logical deduction, MPE-Agent is a field. Building a chain of reasoning The large model executes the reasoning process step by step according to the thought chain, generating the final parameter results:

[0072] After extracting parameters, MPE-Agent encapsulates the results into a structured output. Each parameter result includes a field identifier, field value, extraction method type, and source context information, ensuring good traceability and interpretability.

[0073] 4) Post-validation and conflict resolution based on financial logic constraints After extracting preliminary structured parameter data from the intelligent agent via multi-path extraction, a post-validation and conflict resolution module based on financial logic constraints is introduced to address potential logical inconsistencies, illusory data, or cross-field contradictions that may arise during the extraction process of large models. This module is a crucial step in ensuring the high availability and reliability of the output data in core financial operations.

[0074] (1) Constructing a financial business logic constraint graph For different document types, a financial business logic constraint graph is pre-constructed. This graph includes multiple nodes and edges. Each node represents a parameter field, and each edge represents the business logic relationship between two parameter fields. This business logic relationship can include, but is not limited to: Numerical calculation constraints: For example, in a margin trading contract, the constraint is that the margin ratio = (financing amount / collateral value) × 100%; Identity and mutual exclusion constraints: For example, the purchase amount and the sale amount should not both be non-zero; the start date must be earlier than the end date; Enumeration and range constraints: For example, the value of the transaction type field must belong to the preset set {"buy", "sell", "subscribe", "redeem"}; the annualized rate of return must be between 0% and 100%.

[0075] (2) Consistency verification of extraction results Traverse each edge in the financial business logic constraint graph, and perform validation operations corresponding to the business logic relationship on the field values ​​of the parameter fields associated with the edge in the preliminary structured parameter data; if all validation operations pass, the consistency validation is determined to be successful; if any validation operation fails, the consistency validation is determined to be unsuccessful, and the conflicting fields and conflict types that failed the validation are recorded.

[0076] (3) Intelligent resolution and re-extraction of conflicts When a conflict occurs, the result is not discarded directly, but the following hierarchical resolution process is executed: 1. Source tracing and location: Based on the source context information, locate the original text fragments in the financial document corresponding to the conflicting fields; 2. Targeted Re-extraction: For conflicting fields, switch the extraction strategy to perform local re-extraction on the original text fragment to obtain the re-extracted field values. Here, if the extraction strategy for conflicting fields is the large model inference extraction strategy, then switch the extraction strategy for conflicting fields to the retrieval enhancement generation extraction strategy or the regular expression extraction strategy, and perform local re-extraction on the original text fragment to obtain the re-extracted field values. If the extraction strategy for conflicting fields is the retrieval enhancement generation extraction strategy, then expand the retrieval context window, or input the constraint conditions corresponding to the conflict type as part of the prompt words into the large model, and perform local re-extraction on the original text fragment to obtain the re-extracted field values. 3. Re-verification of resolution results: Substitute the re-extracted field values ​​into the initial structured parameter data for updating, and perform consistency verification on the updated initial structured parameter data again; if the verification passes again, the updated initial structured parameter data is determined as the final structured parameter result; 4. Confidence labeling and risk warning: If the validation fails again, add a conflict warning label and confidence score to the conflict field, and determine the preliminary structured parameter data after adding the label as the final structured parameter result.

[0077] This embodiment proposes a structured parameter outputter based on a large model for financial scenarios, enabling high-precision structured extraction of multi-source heterogeneous financial documents. It addresses the challenges of semantic understanding and logical reasoning under complex formats, improving the efficiency and accuracy of automated financial processing. First, a two-level classification method for financial documents based on OCR parsing and a large model is proposed, identifying document types according to a pre-defined financial document classification system. Next, an intelligent agent parameter extraction planning mechanism based on optimization modeling is proposed, which adaptively plans the extraction path based on the fit between parameters and extraction methods by optimizing the parameter extraction planning problem. Then, a structured parameter output architecture based on a multi-path extraction agent is introduced, comprehensively utilizing the accuracy of regular expressions, the knowledge retrieval capabilities of RAG technology, and the logical reasoning capabilities of large language models to achieve a deep understanding of financial business parameters and initial structured parameter data output. Next, a post-validation and conflict resolution module based on financial logic constraints is introduced, constructing a financial business logic constraint graph to verify the consistency of extraction results and intelligently resolve and re-extract conflicting results. Finally, unstructured financial text is converted into high-quality JSON structured data that conforms to business logic. The structured parameter outputter mentioned above reduces computational costs while improving the accuracy and robustness of the extraction results.

[0078] It should be noted that the sequence numbers of the steps in the above embodiments do not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. In practical applications, all the above possible implementation methods can be arbitrarily combined in a combined manner to form possible embodiments of this application, which will not be described in detail here.

[0079] Based on the structured parameter output method for financial scenarios based on large models provided in the above embodiments, and based on the same inventive concept, this application also provides a structured parameter output device for financial scenarios based on large models.

[0080] Figure 4 This is a structural diagram of a structured parameter output device based on a large model for financial scenarios, provided in an embodiment of this application. Figure 4 As shown, the structured parameter output device based on a large model for financial scenarios may specifically include an identification unit 410, a determination unit 420, a planning unit 430, an extraction and generation unit 440, and a verification output unit 450.

[0081] The identification unit 410 is used to acquire multi-source heterogeneous financial documents, identify the document type of the financial documents, and obtain the financial document type. Unit 420 is used to determine the set of parameter fields to be extracted based on the financial document type; Planning unit 430 is used to plan the target extraction method for each parameter field in the parameter field set through a strategy planning agent, thereby obtaining the extraction strategy; The extraction and generation unit 440 is used to extract the agent through multiple paths. According to the extraction strategy, it calls the target extraction method to perform the extraction operation to obtain the field value corresponding to each parameter field. Based on all parameter fields and the field value corresponding to each parameter field, it generates preliminary structured parameter data. The preliminary structured parameter data includes the target extraction method and source context information used for the field value corresponding to each parameter field. The verification output unit 450 is used to perform consistency verification on the preliminary structured parameter data based on the pre-built financial business logic constraint graph corresponding to the financial document type. If the consistency verification passes, the preliminary structured parameter data is output as the final structured parameter result. If the consistency verification fails, the conflict fields that failed the verification are resolved to obtain the final structured parameter result and output it.

[0082] This application embodiment provides a possible implementation, wherein the verification output unit 450 is further configured to: Obtain a pre-built financial business logic constraint graph corresponding to the financial document type. The financial business logic constraint graph includes multiple nodes and multiple edges. Each node represents a parameter field, and each edge represents the business logic relationship between two parameter fields. Traverse each edge in the financial business logic constraint graph and perform validation operations corresponding to the business logic relationship on the field values ​​of the parameter fields associated with the edges in the preliminary structured parameter data; If all verification operations pass, then the consistency check is considered successful. If any verification operation fails, the consistency verification is determined to have failed, and the conflicting fields and conflict types of the failed verifications are recorded.

[0083] This application embodiment provides a possible implementation, wherein the verification output unit 450 is further configured to: Based on the source context information, locate the original text fragment in the financial document corresponding to the conflicting field; For conflicting fields, switch the extraction strategy to perform partial re-extraction on the original text fragment to obtain the re-extracted field values; The re-extracted field values ​​are substituted into the initial structured parameter data for updating, and the updated initial structured parameter data is checked for consistency again. If the verification passes again, the updated preliminary structured parameter data will be determined as the final structured parameter result. If the validation fails again, a conflict warning label and confidence score are added to the conflict field, and the preliminary structured parameter data after adding the label is determined as the final structured parameter result.

[0084] This application embodiment provides a possible implementation method, and the extraction strategy includes any one of the following: regular expression extraction strategy, retrieval enhancement generation extraction strategy, and large model inference extraction strategy; the verification output unit 450 is further used for: If the extraction strategy for conflicting fields is the large model inference extraction strategy, then switch the extraction strategy for conflicting fields to the retrieval enhancement generation extraction strategy or the regular expression extraction strategy, and perform local re-extraction on the original text fragment to obtain the re-extracted field values. If the extraction strategy for conflict fields is a retrieval enhancement generation extraction strategy, then the retrieval context window is expanded, or the constraints corresponding to the conflict type are input into the large model as part of the prompt words, and local re-extraction is performed on the original text fragment to obtain the re-extracted field values.

[0085] This application embodiment provides a possible implementation, wherein the identification unit 410 is further configured to: Financial documents are processed using optical character recognition (OCR) to obtain a set of text information. Based on the preset document type system and text information set, a preliminary classification process is performed through rule matching to obtain the preliminary classification result and the confidence level corresponding to the preliminary classification result; If the confidence level is greater than or equal to the preset threshold, the primary classification result will be determined as a financial document type. If the confidence level is less than the preset threshold, the text information set is input into the large model for secondary classification processing to obtain the classification result of the large model, and the classification result of the large model is determined as a financial document type.

[0086] This application embodiment provides a possible implementation, wherein the text information set includes a document title and a text sequence; the recognition unit 410 is further configured to: For each candidate document type in the preset document type system, the document title is matched with the standard title of the candidate document type to obtain the title matching degree; The keyword coverage rate is determined based on the number of preset business feature keywords corresponding to the candidate document types contained in the text sequence; The title matching degree and keyword coverage are weighted and fused to obtain the confidence degree corresponding to the candidate document type; The candidate document type with the highest confidence level is determined as the primary classification result, and the confidence level corresponding to the candidate document type with the highest confidence level is determined as the confidence level corresponding to the primary classification result.

[0087] This application embodiment provides a possible implementation, wherein the planning unit 430 is further used for: By using a strategy planning agent, each parameter field in the parameter field set is matched and scored with multiple preset candidate extraction methods to obtain a comprehensive matching score for each parameter field relative to each candidate extraction method. By using a policy planning agent, based on multiple comprehensive adaptation scores corresponding to each parameter field, and with the goal of maximizing the sum of the comprehensive adaptation scores corresponding to the set of parameter fields, a candidate extraction method is assigned to each parameter field as the target extraction method, thus obtaining the extraction strategy.

[0088] This application embodiment provides a possible implementation, where multiple preset candidate extraction methods include regular expression extraction, retrieval enhancement generation extraction, and large model inference extraction; the extraction generation unit 440 is further used for: If the target extraction method for the parameter field is regular expression extraction, then the multi-path extraction agent will match the text sequence of the financial document according to the regular expression corresponding to the parameter field to obtain the field value. If the target extraction method for the parameter field is retrieval-enhanced generation extraction, then a multi-path extraction agent is used to retrieve a set of relevant text fragments from the text sequence of the financial document based on the semantics of the parameter field; the parameter field and the set of relevant text fragments are then input into a large model for processing to obtain the field value. If the target extraction method for the parameter fields is the large model inference extraction method, then the agent is extracted through multiple paths, and an inference thought chain is constructed based on the parameter fields; the inference thought chain and the text sequence of the financial document are input into the large model for inference processing to obtain the field values.

[0089] This application embodiment provides a possible implementation, wherein the determining unit 420 is further configured to: Based on the preset association between document types and parameter field sets, determine the set of parameter fields to be extracted that are associated with financial document types.

[0090] Based on the same inventive concept, this application also provides an electronic device, including a processor and a memory, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the structured parameter output method based on a large model for financial scenarios in any of the above embodiments.

[0091] In an exemplary embodiment, an electronic device is provided, such as Figure 5 As shown, Figure 5 The illustrated electronic device 500 includes a processor 501 and a memory 503. The processor 501 and the memory 503 are connected, for example, via a bus 502. Optionally, the electronic device 500 may also include a transceiver 504. It should be noted that in practical applications, the transceiver 504 is not limited to one type, and the structure of this electronic device 500 does not constitute a limitation on the embodiments of this application.

[0092] Processor 501 may be a CPU (Central Processing Unit), GPU (Graphics Processing Unit), DSP (Digital Signal Processor), ASIC (Application Specific Integrated Circuit), FPGA, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 501 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0093] Bus 502 may include a pathway for transmitting information between the aforementioned components. Bus 502 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 502 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0094] The memory 503 may be a ROM (Read Only Memory) or other type of static storage device capable of storing static information and instructions, RAM (Random Access Memory) or other type of dynamic storage device capable of storing information and instructions, or an EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0095] The memory 503 stores computer program code that executes the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the computer program code stored in the memory 503 to implement the content shown in the foregoing method embodiments.

[0096] Among them, electronic devices include, but are not limited to: mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (personal digital assistants), PADs (tablet computers), PMPs (portable multimedia players), and in-vehicle terminals (such as in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 5 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0097] Those skilled in the art will clearly understand that the specific working process of the systems, devices, and modules described above can be referred to the corresponding process in the foregoing method embodiments. For the sake of brevity, it will not be repeated here.

[0098] Those skilled in the art will understand that the technical solution of this application, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several program instructions to cause an electronic device (e.g., a personal computer, server, or network device) to execute all or part of the steps of the methods described in the embodiments of this application when running the program instructions. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory, random access memory, magnetic disks, or optical disks.

[0099] Alternatively, all or part of the steps of the foregoing method embodiments can be implemented by hardware (such as electronic devices like personal computers, servers, or network devices) associated with program instructions. The program instructions can be stored in a computer-readable storage medium. When the program instructions are executed by the processor of the electronic device, the electronic device executes all or part of the steps of the methods described in the embodiments of this application.

[0100] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that within the spirit and principles of this application, modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein; and these modifications or substitutions do not cause the corresponding technical solutions to leave the protection scope of this application.

Claims

1. A structured parameter output method based on a large model for financial scenarios, characterized in that, The method includes: Acquire multi-source heterogeneous financial documents, identify document types of the financial documents, and obtain the financial document types; Determine the set of parameter fields to be extracted based on the financial document type; By using a strategy planning agent, a target extraction method is planned for each parameter field in the parameter field set, thus obtaining the extraction strategy; By using a multi-path extraction agent, extraction operations are performed by calling the target extraction method according to the extraction strategy to obtain the field value corresponding to each parameter field. Based on all parameter fields and the field value corresponding to each parameter field, preliminary structured parameter data is generated. The preliminary structured parameter data includes the target extraction method and source context information used for the field value corresponding to each parameter field. Based on a pre-built financial business logic constraint graph corresponding to financial document types, the consistency of the preliminary structured parameter data is verified. If the consistency check passes, the preliminary structured parameter data will be used as the final structured parameter result and output. If the consistency check fails, the conflict fields that failed the check are resolved to obtain the final structured parameter result and output it.

2. The method according to claim 1, characterized in that, Based on a pre-constructed financial business logic constraint graph corresponding to financial document types, consistency checks are performed on the initial structured parameter data, including: Obtain a pre-built financial business logic constraint graph corresponding to the financial document type. The financial business logic constraint graph includes multiple nodes and multiple edges. Each node represents a parameter field, and each edge represents the business logic relationship between two parameter fields. Traverse each edge in the financial business logic constraint graph and perform validation operations corresponding to the business logic relationship on the field values ​​of the parameter fields associated with the edges in the preliminary structured parameter data; If all verification operations pass, then the consistency check is considered successful. If any verification operation fails, the consistency verification is determined to have failed, and the conflicting fields and conflict types of the failed verifications are recorded.

3. The method according to claim 2, characterized in that, If the consistency check fails, conflict resolution is performed on the conflicting fields that failed the check, and the final structured parameter result is obtained and output, including: Based on the source context information, locate the original text fragment in the financial document corresponding to the conflicting field; For conflicting fields, switch the extraction strategy to perform partial re-extraction on the original text fragment to obtain the re-extracted field values; The re-extracted field values ​​are substituted into the initial structured parameter data for updating, and the updated initial structured parameter data is checked for consistency again. If the verification passes again, the updated preliminary structured parameter data will be determined as the final structured parameter result. If the validation fails again, a conflict warning label and confidence score are added to the conflict field, and the preliminary structured parameter data after adding the label is determined as the final structured parameter result.

4. The method according to claim 3, characterized in that, Extraction strategies include any one of the following: regular expression extraction strategy, retrieval enhancement generation extraction strategy, and large model inference extraction strategy; For conflicting fields, the extraction strategy is switched to perform partial re-extraction on the original text fragment, resulting in the re-extracted field values, including: If the extraction strategy for conflicting fields is the large model inference extraction strategy, then switch the extraction strategy for conflicting fields to the retrieval enhancement generation extraction strategy or the regular expression extraction strategy, and perform local re-extraction on the original text fragment to obtain the re-extracted field values. If the extraction strategy for conflict fields is a retrieval enhancement generation extraction strategy, then the retrieval context window is expanded, or the constraints corresponding to the conflict type are input into the large model as part of the prompt words, and local re-extraction is performed on the original text fragment to obtain the re-extracted field values.

5. The method according to claim 1, characterized in that, Document type identification is performed on financial documents to obtain their types, including: Financial documents are processed using optical character recognition (OCR) to obtain a set of text information. Based on the preset document type system and text information set, a preliminary classification process is performed through rule matching to obtain the preliminary classification result and the confidence level corresponding to the preliminary classification result; If the confidence level is greater than or equal to the preset threshold, the primary classification result will be determined as a financial document type. If the confidence level is less than the preset threshold, the text information set is input into the large model for secondary classification processing to obtain the classification result of the large model, and the classification result of the large model is determined as a financial document type.

6. The method according to claim 5, characterized in that, The text information set includes the document title and a sequence of text. Based on a pre-defined document type system and text information set, a preliminary classification process is performed using rule matching to obtain the preliminary classification results and their corresponding confidence levels, including: For each candidate document type in the preset document type system, the document title is matched with the standard title of the candidate document type to obtain the title matching degree; The keyword coverage rate is determined based on the number of preset business feature keywords corresponding to the candidate document types contained in the text sequence; The title matching degree and keyword coverage are weighted and fused to obtain the confidence degree corresponding to the candidate document type; The candidate document type with the highest confidence level is determined as the primary classification result, and the confidence level corresponding to the candidate document type with the highest confidence level is determined as the confidence level corresponding to the primary classification result.

7. The method according to claim 1, characterized in that, By using a policy planning agent, a target extraction method is planned for each parameter field in the parameter field set, resulting in an extraction strategy, including: By using a strategy planning agent, each parameter field in the parameter field set is matched and scored with multiple preset candidate extraction methods to obtain a comprehensive matching score for each parameter field relative to each candidate extraction method. By using a policy planning agent, based on multiple comprehensive adaptation scores corresponding to each parameter field, and with the goal of maximizing the sum of the comprehensive adaptation scores corresponding to the set of parameter fields, a candidate extraction method is assigned to each parameter field as the target extraction method, thus obtaining the extraction strategy.

8. The method according to claim 7, characterized in that, The preset candidate extraction methods include regular expression extraction, retrieval enhancement generation extraction, and large model inference extraction. By using a multi-path extraction agent, the extraction operation is performed by calling the target extraction method according to the extraction strategy, and the field values ​​corresponding to each parameter field are obtained, including: If the target extraction method for the parameter field is regular expression extraction, then the multi-path extraction agent will match the text sequence of the financial document according to the regular expression corresponding to the parameter field to obtain the field value. If the target extraction method for the parameter field is retrieval-enhanced generation extraction, then a multi-path extraction agent is used to retrieve a set of relevant text fragments from the text sequence of the financial document based on the semantics of the parameter field; the parameter field and the set of relevant text fragments are then input into a large model for processing to obtain the field value. If the target extraction method for the parameter fields is the large model inference extraction method, then the agent is extracted through multiple paths, and an inference thought chain is constructed based on the parameter fields; the inference thought chain and the text sequence of the financial document are input into the large model for inference processing to obtain the field values.

9. The method according to claim 1, characterized in that, Based on the financial document type, determine the set of parameter fields to be extracted, including: Based on the preset association between document types and parameter field sets, determine the set of parameter fields to be extracted that are associated with financial document types.

10. A structured parameter output device based on a large model for financial scenarios, characterized in that, The device includes: The identification unit is used to acquire multi-source heterogeneous financial documents, identify the document type of the financial documents, and obtain the financial document type. The determination unit is used to determine the set of parameter fields to be extracted based on the financial document type; The planning unit is used to plan the target extraction method for each parameter field in the parameter field set through the strategy planning agent, so as to obtain the extraction strategy. The extraction and generation unit is used to extract the agent through multiple paths. According to the extraction strategy, it calls the target extraction method to perform the extraction operation and obtain the field value corresponding to each parameter field. Based on all parameter fields and the field value corresponding to each parameter field, it generates preliminary structured parameter data. The preliminary structured parameter data includes the target extraction method and source context information used for the field value corresponding to each parameter field. The verification output unit is used to perform consistency verification on the preliminary structured parameter data based on the pre-built financial business logic constraint graph corresponding to the financial document type. If the consistency verification passes, the preliminary structured parameter data is output as the final structured parameter result. If the consistency verification fails, the conflict fields that failed the verification are resolved to obtain the final structured parameter result and output it.

Images