Financial question answer generation method and system, electronic device and storage medium

Abstract

The application provides a financial question answering generation method and system, an electronic device and a storage medium. The method comprises: acquiring a financial question and answer request input by a user, and performing question analysis on the financial question and answer request to generate a financial question portrait; based on the financial question portrait, recalling multiple modal evidence units in parallel as candidate evidence units; taking the candidate evidence units as seed nodes, expanding along the relationship edges in the pre-constructed evidence unit relationship graph for one or more hops to form multiple candidate evidence chains; fusing and sorting the multiple candidate evidence chains to determine a target evidence chain; and generating a candidate answer corresponding to the financial question and answer request based on the target evidence chain. The application can improve the factual consistency, credibility, explainability and auditability of the financial question and answer result.

Description

Technical Field

[0001] This application belongs to the field of financial intelligent question answering technology, and relates to a method and system for generating answers to financial questions, electronic devices, and storage media. Background Technology

[0002] Existing financial question-answering systems are typically based on Large Language Models (LLMs) or Retrieval-Augmented Generation (RAG) frameworks. These approaches usually begin by retrieving relevant content from a knowledge base based on the user's question, and then generate an answer based on the search results. While this technical approach is well-suited for general knowledge-based question answering, it still faces challenges in financial business scenarios such as financial report analysis, interpretation of listed company announcements, fund product Q&A, and regulatory policy interpretation. These challenges include incomplete evidence retrieval, inconsistent statistical methods, insufficient factual basis for answers, and unauditable results.

[0003] On the one hand, financial information does not exist solely in the form of continuous text. Key factual information in listed companies' annual reports, quarterly reports, fund periodic reports, regulatory documents, and industry research reports is distributed across various formats, including financial statements, expense tables, holding tables, bar charts, line charts, pie charts, chart captions, table notes, footnotes, and scanned pages. This information not only includes textual content but also various structured semantic information such as table structural relationships, chart trend relationships, and page layout relationships. Existing technologies typically focus on the main text paragraphs as the primary retrieval target, making it difficult to fully utilize the row and column correspondences in tables, trend change information in charts, and page adjacency relationships. This results in some key evidence not being effectively recalled, leading to problems of missing evidence and incomplete evidence retrieval.

[0004] On the other hand, while some existing multimodal question-answering solutions can handle documents with mixed text and images, they typically employ Optical Character Recognition (OCR) technology to convert tables, charts, and images into text before performing retrieval and generation. Although this approach is relatively simple to implement, it is prone to losing crucial structured information during the conversion process, such as multi-level header structures, statistical definitions, units of measurement, time granularity, coordinate axis semantics, legend correspondences, and trend change characteristics. This results in the incomplete preservation of numerical, temporal, and logical relationships within financial data, consequently affecting the accuracy of financial question-answering tasks such as numerical verification, trend analysis, causal induction, and risk assessment.

[0005] Furthermore, financial question-answering scenarios place high demands on the factual consistency and traceability of answers. Compared to general question-answering tasks, financial question-answering not only requires generating semantically correct and fluent natural language answers, but also ensures that key information such as numerical values, subjects, time periods, statistical standards, and policy application conditions in the answers are consistent with the original evidence. However, most existing financial document consistency processing solutions focus on data consistency verification between the main text and tables within the document. Their main goal is to discover and correct data inconsistencies within the document itself, rather than providing effective solutions for verifying the consistency between answers and evidence in generative financial question-answering scenarios. Limited by factors such as retrieval granularity, evidence organization methods, and the lack of fact-checking mechanisms in the generation process, the generative model is prone to problems during the answering process, such as numerical citation errors, confusion between reporting periods, confusion between net profit attributable to shareholders of the parent company and net profit concepts, confusion between cumulative values ​​and single-quarter values, errors in judging chart trends, and causal attribution inferences lacking supporting evidence. While these problems may not necessarily affect the fluency of the generated text, they significantly reduce the factual consistency, credibility, and auditability of the answers, making it difficult to meet the requirements of financial business scenarios for accuracy, interpretability, and compliance. Summary of the Invention

[0006] This application provides a method and system for generating answers to financial questions, an electronic device, and a storage medium to address the technical problems of existing financial question-and-answer systems, such as incomplete evidence recall, insufficient consistency of facts in answers, low credibility of generated results, and lack of auditability.

[0007] Firstly, this application provides a method for generating answers to financial questions, including:

[0008] The system obtains financial Q&A requests input by users, parses the financial Q&A requests, and generates a financial question profile; the financial question profile includes one or more of the following: target subject, time range, financial indicators, statistical caliber, unit, statistical period, question type, and answer constraints;

[0009] Based on the aforementioned financial problem profile, multimodal evidence units are recalled in parallel as candidate evidence units; the multimodal evidence units include text evidence units, table evidence units, chart evidence units, image region evidence units, and page layout evidence units;

[0010] Using the candidate evidence unit as a seed node, the relationship edge along the pre-constructed evidence unit relationship graph is extended in one or more hops to form multiple candidate evidence chains, wherein each candidate evidence chain includes at least a main evidence unit, a circumstantial evidence unit, and a relationship path connecting the main evidence unit and the circumstantial evidence unit.

[0011] The multiple candidate evidence chains are fused and ranked to determine the target evidence chain; wherein, in the fusion and ranking process, the ranking result of each candidate evidence chain is determined based on the question matching degree, the completeness of the evidence chain, the cross-modal mutual evidence support degree, and the conflict penalty term.

[0012] Based on the target evidence chain, candidate answers corresponding to the financial question-and-answer request are generated.

[0013] In one implementation of the first aspect, the steps for constructing the evidence unit relationship graph include: acquiring financial documents; performing structure-preserving multimodal parsing on the financial documents to obtain multimodal evidence information; the multimodal evidence information includes text evidence information, table evidence information, chart evidence information, image region evidence information, and page layout evidence information; converting the multimodal evidence information into multimodal evidence units with a unified metadata structure, and establishing evidence indexes for the multimodal evidence units to obtain an evidence unit set; and constructing the evidence unit relationship graph based on the evidence unit set; the evidence unit relationship graph is represented as follows:

[0014] G=(E, R),

[0015] Wherein G is the evidence unit relationship diagram; E is the evidence unit set; R is the relationship edge set, used to record the association relationships between each evidence unit in the evidence unit set; the association relationships include at least: same-page adjacency relationships, cross-page continuation relationships, title attribution relationships, figure caption attribution relationships, table caption attribution relationships, and footnote explanation relationships representing page structure associations; text explanation table relationships and text explanation chart relationships representing text and content explanation associations; same subject relationships, same time relationships, same indicator relationships, same statistical caliber relationships, and unit conversion relationships representing financial semantic associations; and numerical support relationships, trend support relationships, and conflict relationships representing cross-evidence support and conflict associations.

[0016] In one implementation of the first aspect, the parallel recall of multimodal evidence units as candidate evidence units based on the financial problem profile includes: generating an initial modal recall strategy according to the problem type in the financial problem profile; the initial modal recall strategy is used to determine the recall priority, recall weight, and recall quota of each modal evidence unit; converting the target subject, time range, financial indicators, statistical caliber, unit, statistical period, and answer constraints in the financial problem profile into structured filtering conditions; and performing modal parallel retrieval in the evidence unit set based on the initial modal recall strategy and the structured filtering conditions to obtain each modal candidate evidence unit.

[0017] In one implementation of the first aspect, the method further includes: splitting the candidate answer into multiple fact fragments; the fact fragments include at least numerical fragments, trend fragments, causal fragments, time fragments, and policy constraint fragments; traversing each fact fragment and determining the corresponding backtracking modality based on the type of the fact fragment; the backtracking modality is used to characterize the priority backtracking evidence modality corresponding to the type of the fact fragment and its backtracking order; based on the backtracking modality, retrieving evidence units related to the fact fragments in the target evidence chain or the candidate evidence chain, and determining the entity matching relationship, evidence support relationship, and evidence conflict relationship between the fact fragments and the retrieved evidence units; calculating the fragment consistency score of the fact fragments based on the entity matching relationship, the evidence support relationship, and the evidence conflict relationship; and calculating the overall answer consistency score of the candidate answer based on the fragment consistency scores of each fact fragment and a pre-set fragment type weight, as the fragment-level consistency backtracking result of the candidate answer.

[0018] In one implementation of the first aspect, determining the corresponding back-check mode based on the type of the factual fragment includes: when the factual fragment is a numerical fragment, prioritizing the back-check of the tabular evidence unit and verifying the unit, currency, and statistical period; when the factual fragment is a trend fragment, prioritizing the back-check of the chart evidence unit and verifying the trend direction, interval, and inflection point; when the factual fragment is a causal fragment, prioritizing the back-check of the textual evidence unit, and using the tabular evidence unit or the chart evidence unit as auxiliary support; when the factual fragment is a time fragment, prioritizing the back-check of the title, header, table heading, reporting period field, or effective date of the regulation; when the factual fragment is a policy constraint fragment, prioritizing the back-check of the regulation text, appendices, and footnotes.

[0019] In one implementation of the first aspect, the method further includes: performing local correction on the candidate answer based on the fragment-level consistency back-check results of the candidate answer; wherein performing local correction on the candidate answer includes: determining whether the fragment consistency score of the fact fragment is lower than a first threshold; if so, determining that the fact fragment has insufficient evidence support, and performing a local re-search on the fact fragment to supplement the evidence units or evidence chains related to the fact fragment; otherwise, maintaining the fact fragment unchanged; determining whether the fragment consistency score of the fact fragment is lower than a second threshold, and whether the degree of evidence conflict corresponding to the fact fragment is higher than a preset conflict threshold; If so, the factual fragment is determined to have a conflict of evidence, and the factual fragment is partially replaced or partially deleted; otherwise, the factual fragment remains unchanged. When the factual fragment is the causal fragment, it is determined whether the causal fragment lacks numerical evidence or trend evidence as corroborating evidence; if so, an uncertainty explanation is added to the causal fragment, or attribution content lacking evidence support is deleted; otherwise, the causal fragment remains unchanged. Based on the corrected factual fragments, a corrected financial Q&A answer is generated, and the cited evidence, evidence chain path, evidence modality type, fragment consistency score, and local correction record corresponding to the financial Q&A answer are output simultaneously.

[0020] In one implementation of the first aspect, the fusion and sorting of the multiple candidate evidence chains to determine the target evidence chain includes: traversing each candidate evidence chain and calculating the basic matching score, evidence chain integrity score, cross-modal mutual verification support, and conflict penalty item corresponding to the candidate evidence chain; wherein the basic matching score includes semantic matching score, keyword matching score, subject matching score, time matching score, indicator matching score, unit, period, and statistical caliber matching score, source credibility score, parsing quality score, and page adjacency or structure attribution support score; the evidence chain integrity score is used to characterize the candidate evidence chain's support for the target subject, time range, financial indicator, statistical caliber, unit, and The statistical period, question type, and answer constraints are covered; the cross-modal mutual verification support is used to characterize the mutual support between modal evidence units in the candidate evidence chain; the conflict penalty term is used to characterize the conflict between modal evidence units in the candidate evidence chain; based on the basic matching score, the evidence chain integrity score, the cross-modal mutual verification support, and the conflict penalty term, the fusion ranking score corresponding to the candidate evidence chain is calculated; based on the fusion ranking score of each candidate evidence chain, the multiple candidate evidence chains are sorted in descending order to obtain the candidate evidence chain ranking result; one or more evidence chains are selected as the target evidence chain according to the candidate evidence chain ranking result.

[0021] Secondly, this application provides a financial question answer generation system, including: an offline graphing module for performing multimodal parsing on financial documents to construct an evidence unit set, an evidence index, and an evidence unit relationship graph; and an online question answering module connected to the offline graphing module for executing the method described in any of the above-mentioned methods.

[0022] Thirdly, this application provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program stored in the memory to cause the electronic device to perform the method described in any of the preceding claims.

[0023] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the preceding claims.

[0024] As described above, the financial question answer generation method and system, electronic device, and storage medium described in this application have the following beneficial effects:

[0025] (1) By using modal parallel recall, text semantic information, table structure information, chart trend information, image region information and page layout information can be used simultaneously to improve the recall completeness of evidence related to financial issues;

[0026] (2) By using the segment-level consistency back-checking mechanism to verify the evidence and score the consistency of each fact segment in the candidate answer, it is possible to locate fact segments with insufficient evidence support, conflicting evidence or inconsistent statistical standards, thereby improving the factual consistency of the financial question answering results;

[0027] (3) By adopting a fragment-level local correction mechanism, the semantic drift, fact coverage and new illusion content that may occur during the overall regeneration process are avoided. While maintaining the overall semantic stability of the answer, the credibility and auditability of the financial question answering results are improved. Attached Figure Description

[0028] Figure 1 The diagram shown is a structural schematic of a financial question answer generation system according to an embodiment of this application.

[0029] Figure 2 The diagram shown is a structural schematic of a financial question answer generation system according to another embodiment of this application.

[0030] Figure 3 The flowchart shown is a method for generating answers to financial questions according to an embodiment of this application.

[0031] Figure 4 The diagram shown is a flowchart of the recall process for candidate evidence units in one embodiment of this application.

[0032] Figure 5 The diagram shown is a flowchart illustrating the construction of an evidence unit relationship diagram in one embodiment of this application.

[0033] Figure 6 The flowchart shown is a process for merging and sorting the multiple candidate evidence chains in one embodiment of this application.

[0034] Figure 7 The flowchart shown is a financial question answer generation method according to another embodiment of this application.

[0035] Figure 8 The flowchart shown is a process for partially modifying candidate answers in one embodiment of this application.

[0036] Figure 9 The flowchart shown is a method for generating answers to financial questions in yet another embodiment of this application.

[0037] Figure 10 The diagram shown is a structural schematic of an electronic device according to an embodiment of this application. Detailed Implementation

[0038] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0039] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the components related to this application and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0040] The following embodiments of this application provide a method and system for generating answers to financial questions, an electronic device, and a storage medium. This solution constructs a multimodal evidence unit relationship diagram of financial documents to achieve the association and organization of information across text, tables, charts, images, and page layout. Based on the financial question profile, it performs multimodal evidence retrieval, evidence chain construction, and fusion sorting to form a target evidence chain corresponding to the user's question. Furthermore, it generates candidate answers based on the target evidence chain and uses a fragment-level consistency review mechanism to verify the factual fragments in the candidate answers, perform consistency scoring, and make local corrections, thereby improving the factual consistency, credibility, interpretability, and auditability of the financial question-and-answer results.

[0041] This application targets financial business scenarios such as financial report analysis, interpretation of listed company announcements, Q&A on fund products, and interpretation of regulatory policies. It is applicable to financial intelligent question-answering tasks involving multimodal financial document understanding, evidence retrieval, answer generation, and consistency verification.

[0042] Please see Figure 1 The diagram shown is a schematic representation of the structure of a financial question answer generation system in one embodiment of this application.

[0043] like Figure 1 As shown in the figure, this application provides a financial question answer generation system, including an offline mapping module and an online question answering module.

[0044] The offline mapping module is used to perform multimodal parsing of financial documents to construct a set of evidence units, an evidence index, and a relationship diagram of evidence units.

[0045] The online question-and-answer module is connected to the offline mapping module and is used to execute the financial question answer generation method described in this application.

[0046] Please see Figure 2 The diagram shown is a structural schematic of a financial question answer generation system in another embodiment of this application.

[0047] like Figure 2 As shown, the financial question answer generation system also includes a self-supervised training module, used to construct training samples based on the natural correspondences within financial documents, and to train an evidence recall model, an evidence chain ranking model, and a fragment consistency matching model based on the training samples. The evidence recall model, the evidence chain ranking model, and the fragment consistency matching model are respectively used in the evidence recall, evidence chain ranking, and answer consistency backtracking processes within the online question answering module.

[0048] It should be noted that the specific workflows of the offline mapping module and the online question-and-answer module will be described in detail in conjunction with the relevant accompanying drawings in the embodiments of the financial question answer generation method of this application.

[0049] The following will describe in detail the principles and implementation methods of a financial question answer generation method and system, electronic device and storage medium of this embodiment, so that those skilled in the art can understand the financial question answer generation method and system, electronic device and storage medium of this embodiment without creative effort.

[0050] Please see Figure 3 The above is a flowchart of a financial question answer generation method in one embodiment of this application.

[0051] like Figure 3As shown, this embodiment provides a method for generating answers to financial questions, including the following steps S100 to S500.

[0052] In step S100, the financial question and answer request input by the user is obtained, and the financial question and answer request is parsed to generate a financial question profile.

[0053] The financial problem profile includes one or more of the following: target subject, time range, financial indicators, statistical scope, unit, statistical period, problem type, and answer constraints.

[0054] Specifically, the target entity is used to represent listed companies, fund products, industry entities, or regulatory targets that users are interested in; the time range is used to represent the financial period, statistical period, or policy application period corresponding to the question; the financial indicators are used to represent target analysis indicators such as operating revenue, net profit, asset size, expense ratio, and holding ratio; the statistical caliber, unit, and statistical period are used to constrain data consistency in the subsequent evidence retrieval and answer generation process; the question type is used to represent that the financial Q&A request belongs to at least one of the following: numerical query question, trend analysis question, cause explanation question, comparative analysis question, risk identification question, or policy interpretation question; and the answer constraints are used to represent the output requirements that need to be met during the answer generation process.

[0055] The financial problem profile may also include other features, such as comparative relationships. To save space, they will not be listed here.

[0056] In step S200, based on the financial problem profile, multimodal evidence units are recalled in parallel as candidate evidence units.

[0057] In this embodiment, an evidence unit is the smallest traceable processing object for the system to perform evidence retrieval, evidence chain construction, fusion and sorting, answer generation, consistency backchecking, and local correction. Each evidence unit includes at least the following: evidence number, document number, document version, page number, page coordinates, modality type, chapter path, title level, disclosing entity, business entity, time field, financial indicator, unit, currency, statistical period, statistical caliber, original content, structured content, parsing quality score, source credibility, and traceable location information.

[0058] The multimodal evidence units include text evidence units, table evidence units, chart evidence units, image area evidence units, and page layout evidence units.

[0059] Key information in financial documents is typically scattered across text paragraphs, financial statements, statistical tables, trend charts, image areas, and page layout. Different modalities of evidence often complement and corroborate each other. This embodiment utilizes an evidence retrieval model trained offline to collaboratively retrieve multimodal evidence through a modal parallel retrieval mechanism, thereby improving the completeness of key evidence retrieval.

[0060] In some embodiments, the evidence recall model may employ one or more of the following: dual-tower retrieval model, cross-encoder model, dense vector retrieval model, sparse retrieval model, dense and sparse fusion retrieval model, multimodal contrastive learning model, graph neural network retrieval model, or retrieval model based on a large language model.

[0061] Please see Figure 4 The above is a flowchart illustrating the recall process of candidate evidence units in one embodiment of this application.

[0062] like Figure 4 As shown, based on the financial problem profile, the parallel recall of multimodal evidence units as candidate evidence units includes the following steps S201 to S203.

[0063] In step S201, an initial modal recall strategy is generated based on the problem type in the financial problem profile.

[0064] The initial modal recall strategy is used to determine the recall priority, recall weight, and recall quota for each modal evidence unit.

[0065] For example, for numerical query problems, the recall priority and recall weight of tabular evidence units are increased; for trend analysis problems, the recall priority and recall weight of chart evidence units are increased; for causal analysis problems, the recall priority of textual evidence units is increased, and the recall weight of tabular or chart evidence units associated with the textual evidence is increased simultaneously; for policy interpretation problems, the recall priority of evidence units corresponding to the main text of the regulations, appendices, and footnotes is increased; for composite problems that simultaneously contain multiple problem type features, corresponding sub-recall strategies are generated for each problem type, and multiple sub-recall strategies are merged to obtain the initial modal recall strategy.

[0066] In some embodiments, the recall quota is the upper limit on the number of each modal evidence unit allowed to enter the candidate set. The system can adaptively adjust the recall quota based on the recall status of candidate evidence.

[0067] In step S202, the target subject, time range, financial indicators, statistical caliber, unit, statistical period and answer constraints in the financial problem profile are transformed into structured filtering conditions.

[0068] The structured filtering conditions are used to constrain the retrieval scope of candidate evidence units, thereby filtering out evidence units that do not match the target subject, target time range, or target indicators during the retrieval process.

[0069] In step S203, based on the initial modal recall strategy and the structured filtering conditions, a modal parallel search is performed in the evidence unit set to obtain candidate evidence units for each modality.

[0070] Specifically, the system performs initial recall in parallel across the text evidence index, table evidence index, chart evidence index, image region evidence index, and page layout evidence index to obtain the corresponding candidate evidence units.

[0071] In some embodiments, a modal Top-K recall mechanism can be used to retrieve the top K evidence units from each modal evidence index and construct a candidate evidence set:

[0072] C0=C_text∪C_table∪C_chart∪C_image∪C_layout;

[0073] Where C0 represents the candidate evidence set, C_text represents the text evidence unit, C_table represents the table evidence unit, C_chart represents the chart evidence unit, C_image represents the image region evidence unit, and C_layout represents the page layout evidence unit.

[0074] In this implementation, by using modal parallel recall, textual semantic information, table structure information, chart trend information, image region information, and page layout information can be used simultaneously to improve the recall completeness of evidence related to financial issues.

[0075] In step S300, the candidate evidence unit is used as a seed node, and one or more hops are extended along the relationship edges of the pre-constructed evidence unit relationship graph to form multiple candidate evidence chains.

[0076] Each candidate chain of evidence includes at least a primary evidence unit, circumstantial evidence units, and a relationship path connecting the primary evidence unit and the circumstantial evidence units. For example, for causal explanation questions, a candidate chain of evidence can be represented as: "Management discussion paragraph — Text explanation table relationship — Profit and loss statement sub-area — Same indicator trend relationship — Net profit trend chart".

[0077] Please see Figure 5 The above is a flowchart showing the construction process of the evidence unit relationship diagram in one embodiment of this application.

[0078] like Figure 5As shown, the construction steps of the evidence unit relationship diagram include the following steps S301 to S304.

[0079] In step S301, financial documents are obtained.

[0080] In this embodiment, the financial documents may be listed company annual reports, quarterly reports, fund periodic reports, regulatory documents, industry research reports, interim announcements, and other documents containing financial information.

[0081] In step S302, the financial document is subjected to structure-preserving multimodal parsing to obtain multimodal evidence information.

[0082] The multimodal evidence information includes text evidence information, table evidence information, chart evidence information, image area evidence information, and page layout evidence information.

[0083] In one embodiment of this application, performing structure-preserving multimodal parsing on the financial document includes extracting content information corresponding to each modality from the financial document and retaining the structural attributes corresponding to each modality.

[0084] In this implementation, the structure-preserving multimodal parsing method can reduce the loss of table structure, chart semantics, and page layout information caused by simple OCR text processing, thus preserving the numerical relationships, time relationships, indicator relationships, and page layout relationships in financial documents.

[0085] In step S303, the multimodal evidence information is transformed into multimodal evidence units with a unified metadata structure, and an evidence index is established for the multimodal evidence units to obtain a set of evidence units.

[0086] Specifically, corresponding evidence indexes are established for different modal evidence units. For example, for text evidence units, text semantic vector indexes, keyword inverted indexes, chapter-level indexes, and context window indexes are established; for table evidence units, table title indexes, row and column title indexes, cell value indexes, table sub-region indexes, unit / caliber indexes, and computable relationship indexes are established; for chart evidence units, chart title indexes, axis indexes, legend indexes, trend feature indexes, peak and trough inflection point indexes, and visual vector indexes are established; for image region evidence units, region visual vector indexes, OCR text indexes, and region location indexes are established; for page layout evidence units, page layout region indexes, adjacency relationship indexes, title attribution indexes, figure caption attribution indexes, table caption attribution indexes, and footnote explanation indexes are established.

[0087] In this embodiment, the evidence index is used to support the rapid location, accurate retrieval, and associated recall of different modal evidence units, enabling the system to perform parallel retrieval and collaborative recall among different modal evidence based on the type of financial problem and retrieval conditions.

[0088] In this implementation, by establishing a unified metadata structure for different modal evidence units, textual evidence, tabular evidence, chart evidence, image region evidence, and page layout evidence can be stored, retrieved, and analyzed using a unified data organization method, thereby achieving unified management and processing of cross-modal evidence.

[0089] In step S304, the evidence unit relationship diagram is constructed based on the evidence unit set.

[0090] In this embodiment, the evidence unit relationship diagram is represented as follows:

[0091] G=(E, R),

[0092] Where G is the relationship graph of the evidence units; E is the set of evidence units; and R is the set of relationship edges, used to record the association relationships between each evidence unit in the set of evidence units.

[0093] Specifically, the relationships include at least the following: relationships representing page structure such as same-page adjacency, cross-page continuation, title attribution, figure caption attribution, table caption attribution, and footnote explanation; relationships representing the relationship between the main text and content explanation such as table relationship and chart relationship; relationships representing the relationship between the same subject, the same time, the same indicator, the same statistical caliber, and unit conversion; and relationships representing the relationship between evidence support and conflict such as numerical support, trend support, and conflict.

[0094] In some embodiments, the relationship edge is used not only to characterize whether there is a relationship between two evidence units, but also to record the relationship type, relationship direction, relationship confidence level, relationship formation basis and auditable path corresponding to the relationship.

[0095] In this implementation, the evidence unit relationship diagram can be used to associate and organize evidence units that are originally scattered across different pages, chapters, and modalities, thus establishing an evidence association network that spans text, tables, charts, image areas, and page layouts.

[0096] In step S400, the multiple candidate evidence chains are fused and sorted to determine the target evidence chain.

[0097] This embodiment can utilize an evidence chain ranking model trained offline. During the fusion ranking process, the evidence chain ranking model determines the ranking result of each candidate evidence chain based on question matching degree, evidence chain completeness, cross-modal mutual evidence support, and conflict penalty term.

[0098] In some embodiments, the evidence chain ranking model can be a learning ranking model, a gradient boosting decision tree model, a neural network ranking model, a graph neural network model, a Transformer ranking model, a large language model sorter, or any combination of the above models.

[0099] Please see Figure 6 The above is a flowchart showing the fusion and sorting of the multiple candidate evidence chains in one embodiment of this application.

[0100] like Figure 6 As shown, the process of fusing and sorting the multiple candidate evidence chains to determine the target evidence chain includes the following steps S401 to S304.

[0101] In step S401, each candidate evidence chain is traversed, and the basic matching score, evidence chain integrity score, cross-modal mutual evidence support, and conflict penalty item corresponding to the candidate evidence chain are calculated.

[0102] Specifically, the basic matching score includes semantic matching score, keyword matching score, subject matching score, time matching score, indicator matching score, unit, period, and statistical caliber matching score, source credibility score, parsing quality score, and page adjacency or structural attribution support score. The evidence chain integrity score characterizes the extent to which the candidate evidence chain covers the target subject, time range, financial indicators, statistical caliber, unit, statistical period, question type, and answer constraints. The cross-modal mutual verification support score characterizes the degree of mutual support between modal evidence units in the candidate evidence chain. The conflict penalty term characterizes the degree of conflict between modal evidence units in the candidate evidence chain.

[0103] In step S402, based on the basic matching score, the evidence chain integrity score, the cross-modal mutual evidence support, and the conflict penalty term, the fusion ranking score corresponding to the candidate evidence chain is calculated.

[0104] Specifically, the fusion ranking score is represented as:

[0105] Score(g)=β1·S_sem+β2·S_entity+β3·S_time+β4·S_metric+β5·S_scope+β6·S_source+β7·S_quality+β8·S_layout+β9·S_cross+β10·S_chain-β11·P_conflict;

[0106] Where S_sem represents the semantic matching score, S_entity represents the subject matching score, S_time represents the time matching score, S_metric represents the metric matching score, S_scope represents the unit, period, and statistical scope matching score, S_source represents the source credibility score, S_quality represents the parsing quality score, S_layout represents the page adjacency or structure affiliation support score, S_cross represents the cross-modal mutual verification support, S_chain represents the evidence chain integrity score, P_conflict represents the conflict penalty term, and β1 to β11 are preset weight parameters.

[0107] The cross-modal mutual verification support S_cross can be further expressed as:

[0108] S_cross=γ1·C_text-table+γ2·C_text-chart+γ3·C_table-chart+γ4·C_layout+γ5·C_footnote;

[0109] Wherein C_text-table represents the consistency between text description and table values, C_text-chart represents the consistency between text description and chart trends, C_table-chart represents the consistency between changes in table values ​​and the direction of chart trends, C_layout represents the degree to which titles, figure captions, table captions, and page adjacencies support the structure of the evidence chain, and C_footnote represents the degree to which footnotes or table captions support the definition of scope.

[0110] In step S403, based on the fusion ranking score of each candidate evidence chain, the multiple candidate evidence chains are sorted in descending order to obtain the candidate evidence chain ranking result.

[0111] In step S404, one or more evidence chains are selected as the target evidence chain based on the candidate evidence chain ranking results.

[0112] Specifically, candidate evidence chains with fusion ranking scores higher than a preset score threshold can be identified as target evidence chains; or the top N candidate evidence chains can be selected from the ranking results as target evidence chains, where N is a preset positive integer.

[0113] In this implementation, by conducting a multi-dimensional comprehensive evaluation of the candidate evidence chain, the quality of the target evidence chain can be improved, and the risk of factual errors caused by missing evidence, inconsistent statistical standards, mismatched time ranges, or cross-modal information conflicts can be reduced.

[0114] In step S500, candidate answers corresponding to the financial question-and-answer request are generated based on the target evidence chain.

[0115] The identified target evidence chain is one that has a high degree of matching with the financial problem profile, high evidence completeness, strong cross-modal mutual verification support capability, and low degree of conflict, which can provide a credible evidence basis for the subsequent generation of financial question and answer results.

[0116] Please see Figure 7 The above is a flowchart of a financial question answer generation method according to another embodiment of this application.

[0117] like Figure 7 As shown, the financial question answer generation method described in this application further includes the following steps S600 to S604.

[0118] In step S600, the candidate answer is split into multiple fact fragments.

[0119] The factual fragments include at least numerical fragments, trend fragments, causal fragments, time fragments, and policy constraint fragments.

[0120] In step S601, each of the fact fragments is traversed, and the corresponding lookup mode is determined according to the type of the fact fragment.

[0121] The backtracking modality is used to characterize the priority backtracking evidence modality and its backtracking order corresponding to the fact fragment type.

[0122] In one embodiment of this application, determining the corresponding lookup mode based on the type of the fact fragment includes:

[0123] (1) When the factual fragment is the numerical fragment, the table evidence unit is checked first, and the unit, currency and statistical period are verified;

[0124] (2) When the factual fragment is the trend fragment, the chart evidence unit is checked back first, and the trend direction, interval and inflection point are verified;

[0125] (3) When the factual fragment is the cause fragment, the textual evidence unit is checked back first, and the table evidence unit or the chart evidence unit is used as auxiliary support;

[0126] (4) When the fact segment is the time segment, prioritize checking the title, header, table header, reporting period field or the effective date of the regulation;

[0127] (5) When the factual fragment is the policy constraint fragment, the main text of the regulation, appendices and footnotes should be checked first.

[0128] In step S602, based on the back lookup mode, evidence units related to the fact fragment are retrieved in the target evidence chain or the candidate evidence chain, and entity matching relationship, evidence support relationship and evidence conflict relationship between the fact fragment and the retrieved evidence units are determined.

[0129] In step S603, the fragment consistency score of the fact fragment is calculated based on the entity matching relationship, the evidence support relationship, and the evidence conflict relationship.

[0130] Specifically, the segment consistency score corresponding to fact segment fk is Cons(fk), and the formula for calculating Cons(fk) is as follows:

[0131] Cons(fk)=δ1·Sup(fk)+δ2·Exact(fk)+δ3·Scope(fk)+δ4·Cross(fk)+δ5·Path(fk)+δ6·Source(fk)-δ7·Conflict(fk)-δ8·Uncertainty(fk);

[0132] Wherein, Sup(fk) represents the support of evidence, used to characterize the quantity, quality, and strength of evidence supporting the factual fragment; Exact(fk) represents the exact match of facts, used to characterize the degree of matching between the subject, time, indicator, value, unit, and statistical caliber and the content of the evidence; Scope(fk) represents the consistency of the scope of application, used to characterize the consistency between the factual fragment and the scope of application of the evidence; Cross(fk) represents the cross-modal mutual verification support, used to characterize the degree of mutual verification from evidence units from different modalities; Path(fk) represents the completeness of the evidence chain, used to characterize the completeness of the evidence chain path supporting the factual fragment; Source(fk) represents the source credibility score, used to characterize the credibility of the source corresponding to the supporting evidence; Conflict(fk) represents the degree of conflict, used to characterize the strength and quantity of evidence that contradicts the factual fragment; Uncertainty(fk) represents the degree of uncertainty, used to characterize the uncertainty caused by missing evidence, ambiguous evidence, or incomplete inference chain. δ1 to δ8 represent the weight coefficients of the corresponding scoring items, and their values ​​can be determined manually, through historical sample statistical analysis, or model training.

[0133] In step S604, based on the segment consistency score of each fact segment and the pre-set segment type weight, the overall answer consistency score of the candidate answer is calculated as the segment-level consistency back-check result of the candidate answer.

[0134] Specifically, the overall answer consistency score is calculated using a weighted summation method:

[0135] Cons(A)= Σwk·Cons(fk) / Σwk

[0136] Where Cons(fk) represents the consistency score of the fact fragment fk; wk represents the importance weight of the fact fragment fk, which is used to characterize the degree of influence of different types of fact fragments on the accuracy and credibility of the answer.

[0137] In some embodiments, different types of factual fragments correspond to different weighting strategies. For example, in a financial Q&A scenario, numerical fragments, time fragments, financial indicator fragments, and statistical caliber fragments directly affect the factual accuracy of the answer, and their corresponding weights are higher than those of general explanatory fragments; in a policy interpretation scenario, fragments concerning the applicable subject, effective time, regulatory ratio, exception clauses, and scope of application directly affect the understanding and application judgment of the policy, and their corresponding weights are higher than those of background explanation fragments.

[0138] It should be noted that steps S600 to S604 above can be completed using the segment consistency matching model trained in the offline stage.

[0139] In some embodiments, the fragment consistency matching model may be a natural language inference model, a fact verification model, a Transformer-based discriminant model, a graph neural network-based relational reasoning model, a large language model-based fact verification model, or a combination of the above models.

[0140] In this implementation, a segment-level consistency back-check mechanism is used to verify the evidence and score the consistency of each fact segment in the candidate answer. This can identify fact segments with insufficient evidence support, conflicting evidence, or inconsistent statistical standards, and provide a basis for subsequent local corrections, thereby improving the factual consistency, accuracy, and credibility of the financial question-and-answer results.

[0141] In one embodiment of this application, the financial question answer generation method further includes step S700.

[0142] In step S700, based on the fragment-level consistency back lookup results of the candidate answers, the candidate answers are locally corrected.

[0143] Please see Figure 8The diagram shows a flowchart of partial correction of candidate answers in one embodiment of this application.

[0144] like Figure 8 As shown, the local correction of the candidate answer includes the following steps S701 to S704.

[0145] In step S701, it is determined whether the segment consistency score of the fact segment is lower than the first threshold; if so, it is determined that the fact segment has insufficient evidence support, and the fact segment is partially re-searched to supplement the evidence units or evidence chains related to the fact segment; otherwise, the fact segment remains unchanged.

[0146] In step S702, it is determined whether the consistency score of the fact segment is lower than the second threshold and whether the degree of evidence conflict corresponding to the fact segment is higher than the preset conflict threshold; if so, it is determined that there is evidence conflict in the fact segment, and the fact segment is partially replaced or partially deleted; otherwise, the fact segment remains unchanged.

[0147] In step S703, when the fact fragment is the cause fragment, it is determined whether the cause fragment lacks numerical or trend evidence as corroborating evidence; if so, an uncertainty description is added to the cause fragment, or attribution content lacking supporting evidence is deleted; otherwise, the cause fragment remains unchanged.

[0148] In step S704, based on the corrected fact fragments, a corrected financial question and answer answer is generated, and the cited evidence, evidence chain path, evidence modality type, fragment consistency score and local correction record corresponding to the financial question and answer answer are output simultaneously.

[0149] In this embodiment, the local correction process does not regenerate the candidate answers as a whole. Instead, it retains the factual fragments that have passed the consistency verification and only performs targeted corrections on factual fragments that are not supported by sufficient evidence, have conflicting evidence, lack corroborating evidence, or have a consistency score lower than a preset threshold.

[0150] After completing the local correction, the system recalculates the segment consistency score Cons(fk) for each fact segment and the overall answer consistency score Cons(A) for each candidate answer, and performs consistency verification again based on the updated scores. When the overall answer consistency score reaches the preset output threshold, the corrected financial question-and-answer result is output; when the preset maximum number of corrections is reached and the preset output threshold is still not met, a conservative answer with risk warnings or uncertainty explanations is output.

[0151] In this implementation, by adopting a fragment-level local correction mechanism, problems such as semantic drift, fact overwriting, and the addition of illusory content that may occur during the overall regeneration process are avoided. While maintaining the overall semantic stability of the answer, the credibility and auditability of the financial question-and-answer results are improved.

[0152] Please see Figure 9 The above is a flowchart of a method for generating answers to financial questions in another embodiment of this application.

[0153] like Figure 9 As shown, the financial question answer generation method described in this application further includes the following steps S801 to S804.

[0154] In step S801, self-supervised training samples are constructed based on the correspondence that naturally exists within the financial documents.

[0155] Specifically, the correspondence that naturally exists within the financial document corresponds to the relationship edges in the evidence unit relationship graph.

[0156] In one embodiment of this application, the self-supervised training samples include positive samples and negative samples.

[0157] The positive samples include at least: the correspondence between management discussion paragraphs and corresponding financial statement items; the correspondence between chart titles and chart areas; the correspondence between table summary paragraphs and table sub-areas; the correspondence between chart captions and chart trends; the correspondence between page titles, footnotes, table notes and body explanations; and the correspondence between the same indicator in different modalities for the same period.

[0158] The negative samples include at least: table areas from different years but with the same indicator name, report items from different companies but with similar businesses, numerical areas with the same subject but different statistical calibers, chart areas with similar trends but different time intervals, numerical areas with different units but which could not be converted, and evidence units that are visually adjacent on the same page but do not belong to the same title or footnote.

[0159] This embodiment can obtain training data without requiring a large amount of manually labeled question-and-answer data, thereby reducing the cost of constructing training samples and improving the consistency between training data and actual financial document scenarios.

[0160] In step S802, based on the self-supervised training samples, evidence recall matching loss, evidence chain integrity loss, fragment consistency loss, and score calibration loss are constructed.

[0161] Specifically, the evidence recall matching loss is used to bring relevant evidence units closer together and negative evidence units further apart; the evidence chain integrity loss is used to constrain the integrity of the relationship path between evidence units in the same evidence chain; the fragment consistency loss is used to constrain the consistency between the answer fact fragment and the evidence chain; and the score calibration loss is used to calibrate the deviation between the model output score and the actual degree of evidence support.

[0162] In step S803, the total loss function is obtained by combining the evidence recall matching loss, the evidence chain integrity loss, the fragment consistency loss, and the score calibration loss.

[0163] In one embodiment of this application, the total loss function is expressed as:

[0164] L = L_retrieval + λ1·L_chain + λ2·L_consistency + λ3·L_calibration;

[0165] Where L_retrieval represents the evidence retrieval matching loss; L_chain represents the evidence chain integrity loss; L_consistency represents the fragment consistency loss; and L_calibration represents the score calibration loss.

[0166] In step S804, with the goal of minimizing the total loss function, the evidence recall model, the evidence chain ranking model, and the fragment consistency matching model are iteratively updated respectively.

[0167] After training, the evidence retrieval model, the evidence chain ranking model, and the fragment consistency matching model are deployed to the online question answering module for evidence retrieval, evidence chain ranking, and consistency verification during the financial question answer generation process.

[0168] In one embodiment of this application, the financial question answer generation method further includes converting the fragment-level consistency back-check results generated during the online question-and-answer process into self-supervised feedback samples, wherein the self-supervised feedback samples are used to continuously optimize the evidence recall model, the evidence chain ranking model, and the fragment consistency matching model.

[0169] Specifically, for answer fragments that are stably supported by the target evidence chain and whose fragment consistency score is higher than the preset positive feedback threshold, the system constructs the corresponding fact fragments, evidence units, and evidence chain paths as positive feedback samples.

[0170] For answer segments that have conflicting evidence, whose segment consistency score is lower than the preset negative feedback threshold, are partially corrected or replaced, or are deleted due to lack of supporting evidence, the system will construct the corresponding fact segments, evidence units, and evidence chain paths as negative feedback samples.

[0171] This implementation reduces reliance on large-scale manually labeled question-and-answer data, lowering model training and maintenance costs. Simultaneously, the system can continuously generate training and feedback samples using newly added financial reports, announcements, fund reports, regulatory documents, and other financial documents, enabling continuous model optimization and iterative updates.

[0172] Given the common discrepancies in financial documents, such as net profit attributable to the parent company versus net profit, operating revenue versus main business revenue, quarterly value versus cumulative value, year-on-year versus quarter-on-quarter, million yuan versus hundred million yuan, parent company caliber versus consolidated caliber, and adjusted caliber versus original caliber, the financial question answer generation method described in this application further includes: performing financial caliber normalization and pre-screening processing in the candidate evidence recall stage, evidence chain fusion and sorting stage, and consistency review stage.

[0173] Specifically, in the candidate evidence recall stage, by performing standardization processing, evidence units that do not match the problem profile are pre-screened and filtered to reduce the probability of inconsistent evidence entering the candidate set; in the evidence chain fusion and ranking stage, by comparing the statistical standard consistency of each evidence unit in the candidate evidence chain, the ranking weight of consistent evidence chains is increased, and evidence chains with conflicting standards are penalized; in the consistency review stage, by performing standard alignment verification on the factual fragments in the candidate answers and supporting evidence, problems such as numerical inconsistencies, trend inconsistencies, or subject scope inconsistencies caused by differences in statistical standards are identified.

[0174] In this implementation, by introducing a standardization mechanism in multiple processing stages, the impact of statistical standard differences on evidence recall, evidence chain ordering, and answer consistency verification can be reduced, thereby improving the accuracy, credibility, and auditability of financial question-and-answer results.

[0175] The present application will be further described below with reference to specific embodiments. It should be understood that the following embodiments are only used to explain the present application and are not intended to limit the scope of protection of the present application. Equivalent substitutions or modifications made by those skilled in the art without departing from the essential concept of the present application should fall within the scope of protection of the present application.

[0176] Example 1: Q&A on financial statement attribution of listed companies.

[0177] A user asked: "What are the main reasons for a company's revenue growth but net profit decline in 2024?"

[0178] First, the problem is analyzed to identify the target entity as a listed company, the time frame as 2024, the financial indicators as operating revenue and net profit, the comparative relationship as operating revenue growth and net profit decline, the problem type as cause explanation, and a corresponding financial problem profile is formed.

[0179] Subsequently, based on the characteristics of cause-of-fact questions, an initial modal recall strategy is generated, focusing on the management discussion and analysis section, income statement, period expense statement, asset impairment loss related tables, and operating revenue and net profit trend charts. Answer constraints are set for the evidence chain of "cause text + table values ​​+ trend chart".

[0180] During the recall phase, the textual evidence index retrieves management's explanations regarding declining gross profit margins, increased period expenses, or increased asset impairment; the tabular evidence index retrieves the income statement, expense details, and asset impairment loss statement; and the chart evidence index retrieves the revenue and net profit trend graphs. Subsequently, a pre-constructed evidence unit relationship diagram is invoked, organizing the management discussion paragraphs, income statement sub-areas, and net profit trend graph areas into candidate evidence chains according to relationships such as "textual explanation table relationships" and "relationships with the same indicator trends."

[0181] During the ranking phase, candidate evidence chains are merged and ranked. If the description of "declining gross profit margin" in the text can be supported by the corresponding data in the profit statement, and the net profit trend chart also shows a downward trend, then the cross-modal mutual support and evidence chain integrity scores of the evidence chain are improved; conversely, if a certain cause description only appears in the text and lacks supporting tables or charts, then its merged ranking score is reduced.

[0182] Subsequently, candidate answers are generated based on the top-ranked target evidence chains. After the candidate answers are generated, the answers are broken down into multiple factual fragments, including fragments such as "2024 operating revenue growth", "net profit decline", "main reasons include declining gross profit margin", and "increased asset impairment losses", and the corresponding evidence chains are checked back for each fragment.

[0183] If the "increase in asset impairment losses" segment cannot be supported by both tabular and textual evidence units, then the segment is corrected by deletion, replacement, or the addition of uncertainty explanations, without rewriting the entire answer. After correction, the system recalculates the segment consistency score and the overall consistency score, and outputs the final answer, corresponding evidence source, evidence path, modality type, and correction record. Simultaneously, the results of this review and correction are transformed into feedback samples for subsequent self-supervised training.

[0184] Example 2: Q&A on the reasons for fund product volatility.

[0185] A user asked: "What are the main reasons for the increased volatility in the fund's net asset value over the past year?"

[0186] First, the problem is identified as a composite problem of trend analysis and cause explanation. The target subject is determined to be a certain fund, the time range is the past year, the financial indicator is the fund's net asset value fluctuation, the problem type is trend change cause analysis, and a corresponding financial problem profile is generated.

[0187] Subsequently, based on the needs of complex questions, priority was given to allocating recall quotas to net asset value trend charts, fund manager operation analysis sections, industry allocation tables of holdings, tables of changes in the top ten holdings, and risk exposure explanations.

[0188] Next, based on chart evidence, we retrieved the net asset value trend chart; based on table evidence, we retrieved the industry allocation table and the table of changes in the top ten holdings; and based on text evidence, we retrieved the fund manager's operational analysis paragraphs and risk exposure explanations.

[0189] Subsequently, based on the pre-established evidence unit relationship diagram, the “net asset value trend chart - holdings allocation table - operation analysis text” were organized into a candidate evidence chain using the same indicator relationship, the same period relationship, and the text interpretation relationship.

[0190] Subsequently, the candidate evidence chains are merged and ranked. If the net asset value trend chart shows a significant expansion in the fluctuation range, the industry allocation table shows an increase in the allocation ratio of high-volatility industries, and the fund manager's operational analysis text explains a market style shift or portfolio structure adjustment, then the cross-modal mutual support, relationship path completeness, and overall fusion ranking score of the evidence chain will be improved. If a conclusion is supported only by a single text fragment and not by the holding table or trend chart, then the ranking score of the corresponding evidence chain will be reduced. Based on the ranking results, candidate answers regarding the increased net asset value volatility of the fund over the past year are output.

[0191] During the consistency verification phase, each causal segment in the candidate answers is examined for support, exact match, consistency of terminology, degree of multimodal support, and degree of conflict. For example, if a candidate answer states "longer bond duration leads to increased volatility," but the portfolio structure and operational analysis do not reflect information related to bond duration, this segment has low support and cross-modal support, while its degree of conflict or uncertainty is high. In this case, the segment is identified as a weakly supported segment and deleted, or replaced with a more conservative statement supported by evidence, such as "Based on existing evidence, increased volatility is more likely related to increased industry concentration in equity holdings and a shift in market style." Finally, the revised final answer is output, and the replaced or deleted segments are used as negative feedback samples for subsequent model optimization.

[0192] Example 3: Q&A on the interpretation of regulatory policies.

[0193] A user asked: "What changes have been made to the provision coverage ratio requirements for consumer finance companies under the new regulations?"

[0194] First, identify that the question belongs to the policy interpretation category, determine that the target entity is consumer finance companies, the financial indicator is the provision coverage ratio, and extract the answer constraints from the question, including the applicable objects, effective time, regulatory ratio, and changes.

[0195] Subsequently, based on the characteristics of policy interpretation questions, an initial modal recall strategy was generated, primarily targeting the main clauses, appendices, footnotes, and scope of application of regulatory documents.

[0196] During the recall phase, priority is given to recalling candidate evidentiary units such as the main text clauses, appendices, footnotes, and scope descriptions related to the provision coverage ratio requirements, and forming a candidate evidentiary chain.

[0197] Next, the candidate chains of evidence are merged and ranked. When the main clause only provides general requirements, while the appendix clearly lists specific regulatory ratios, the system prioritizes the explicit values ​​in the appendix and supplements the main clause with explanations of the applicable objects and effective dates, thereby improving the ranking score of the chains of evidence that are more consistent with the user's question, have more credible sources, and are more accurately described.

[0198] Next, candidate answers are generated based on the top-ranked chain of evidence. After generation, the candidate answers are broken down into multiple factual fragments, such as "applicable to consumer finance companies," "the provision coverage ratio is required to be a certain percentage," "effective from a certain date," and "not applicable to other types of institutions," and each fragment is reviewed against the main text, appendices, and footnotes of the regulatory document. If a fragment references regulatory requirements for other institutions that are not applicable to consumer finance companies, the system determines that the fragment conflicts with the scope of application.

[0199] At this point, by standardizing financial terminology, the applicable entities, effective dates, regulatory ratios, and scope of application are verified, distinguishing between those that can and cannot be standardized. Regulatory requirements not applicable to consumer finance companies are not forcibly merged. Subsequently, the conflicting information is deleted or corrected to accurately correspond to the regulatory requirements for consumer finance companies. Finally, a consistent policy interpretation is output, simultaneously recording the correction trajectory and scope of application conflict information for subsequent model iteration and optimization.

[0200] Example 4: Recall of chart evidence in scanned annual reports.

[0201] In this embodiment, the object being processed is a scanned annual report.

[0202] First, structure-preserving multimodal parsing is used to detect and parse the title area, body text area, table area, chart area, and caption area of ​​the page, preserving the page coordinates, layout relationships, and reading order information of each area. Compared to processing methods that simply convert the entire scanned page into plain text, this embodiment preserves the chart structure, table structure, and page layout information during the parsing stage.

[0203] Subsequently, the parsed results are transformed into multimodal evidence units with a unified metadata structure. For chart areas, in addition to saving the OCR text, axis names, legends, data points, trend directions, and inflection point information are further extracted and written into the index as chart evidence units. At the same time, multimodal indexes are established for titles, body text, tables, charts, and captions, and evidence unit relationship diagrams are constructed based on title attribution, relationships with the same indicator, relationships within the same period, and page adjacency relationships.

[0204] When a user asks the question "What is the trend of the company's R&D investment in the past three years?", the target entity is identified as a company, the time range is the past three years, the financial indicator is R&D investment, and the question type is trend analysis.

[0205] Subsequently, based on the requirements of trend analysis questions, the R&D investment description paragraph, R&D expense table, and R&D investment trend chart were prioritized as recall targets. Next, the R&D investment description paragraph was recalled using textual evidence indexing, the R&D expense table was recalled using tabular evidence indexing, and the R&D investment trend chart was recalled using graphical evidence indexing. Relationship expansion was then completed using an evidence relationship diagram, forming a candidate evidence chain among the R&D description text, R&D expense table, and R&D trend chart.

[0206] In subsequent processing, the cross-modal corroboration relationship between textual descriptions, table values, and trend graph directions is evaluated, prioritizing the retention of mutually supporting evidence chains. Then, candidate answers regarding the changing trends of R&D investment over the past three years are generated based on the target evidence chain.

[0207] Next, the trend and numerical segments in the candidate answers are reviewed and verified. If the OCR misinterprets a certain numerical value, but the trend chart and R&D expense table both show an overall increase in R&D investment, the system reduces the impact of a single OCR error on the answer through a cross-modal mutual verification mechanism. If necessary, the affected numerical segment is locally corrected without rewriting the entire trend conclusion, thus ensuring the stability and controllability of the output answer. Finally, the verified and corrected final answer and its evidence path are output.

[0208] Please see Figure 10 The image shown is a schematic diagram of the structure of an electronic device according to an embodiment of this application.

[0209] like Figure 10As shown, this application provides an electronic device, including:

[0210] A memory for storing computer programs.

[0211] A processor, the processor being configured to execute a computer program stored in the memory, so as to cause the electronic device to perform the method described in any of the preceding descriptions.

[0212] Preferably, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The memory can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0213] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, or methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules / units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or units may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or modules or units may be electrical, mechanical, or other forms.

[0214] The modules / units described as separate components may or may not be physically separate. The components shown as modules / units may or may not be physical modules; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules / units can be selected to achieve the objectives of the embodiments of this application, depending on actual needs. For example, the functional modules / units in the various embodiments of this application may be integrated into one processing module, or each module / unit may exist physically separately, or two or more modules / units may be integrated into one module / unit.

[0215] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0216] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the methods described in any of the above embodiments. Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing a processor. The program can be stored in a computer-readable storage medium, which is a non-transitory medium, such as random access memory, read-only memory, flash memory, hard disk, solid-state drive, magnetic tape, floppy disk, optical disk, and any combination thereof. The storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital video disc (DVD)), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0217] The descriptions of the processes or structures corresponding to the above-mentioned figures each have their own emphasis. For parts of a process or structure that are not described in detail, please refer to the relevant descriptions of other processes or structures.

[0218] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A method for generating answers to financial questions, characterized in that, include: Obtain financial Q&A requests input by users, and parse the financial Q&A requests to generate financial Q&A profiles; The financial problem profile includes one or more of the following: target subject, time range, financial indicators, statistical scope, unit, statistical period, problem type, and answer constraints. Based on the aforementioned financial problem profile, multimodal evidence units are recalled in parallel as candidate evidence units; the multimodal evidence units include text evidence units, table evidence units, chart evidence units, image region evidence units, and page layout evidence units; Using the candidate evidence unit as a seed node, the relationship edge along the pre-constructed evidence unit relationship graph is extended in one or more hops to form multiple candidate evidence chains, wherein each candidate evidence chain includes at least a main evidence unit, a circumstantial evidence unit, and a relationship path connecting the main evidence unit and the circumstantial evidence unit. The multiple candidate evidence chains are merged and sorted to determine the target evidence chain; In the fusion ranking process, the ranking result of each candidate evidence chain is determined based on the problem matching degree, the completeness of the evidence chain, the cross-modal mutual evidence support degree, and the conflict penalty term. Based on the target evidence chain, candidate answers corresponding to the financial question-and-answer request are generated.

2. The method according to claim 1, characterized in that, The steps for constructing the evidence unit relationship diagram include: Obtain financial documents; The financial document is subjected to structure-preserving multimodal parsing to obtain multimodal evidence information; the multimodal evidence information includes text evidence information, table evidence information, chart evidence information, image region evidence information, and page layout evidence information; The multimodal evidence information is transformed into multimodal evidence units with a unified metadata structure, and an evidence index is established for the multimodal evidence units to obtain a set of evidence units; Based on the set of evidence units, the relationship diagram of the evidence units is constructed; the relationship diagram of the evidence units is represented as follows: G=(E, R), Wherein G is the evidence unit relationship diagram; E is the evidence unit set; R is the relationship edge set, used to record the association relationships between each evidence unit in the evidence unit set; the association relationships include at least: same-page adjacency relationship, cross-page continuation relationship, title attribution relationship, figure caption attribution relationship, table caption attribution relationship, and footnote explanation relationship representing page structure association; main text explanation table relationship and main text explanation chart relationship representing the association between main text and content explanation; same subject relationship, same time relationship, same indicator relationship, same statistical caliber relationship, and unit conversion relationship representing financial semantic association; numerical support relationship, trend support relationship, and conflict relationship representing cross-evidence support and conflict association.

3. The method according to claim 2, characterized in that, Based on the aforementioned financial problem profile, the parallel recall of multimodal evidence units as candidate evidence units includes: Based on the problem types in the financial problem profile, an initial modal recall strategy is generated; the initial modal recall strategy is used to determine the recall priority, recall weight, and recall quota of each modal evidence unit; The target subject, time range, financial indicators, statistical caliber, unit, statistical period, and answer constraints in the financial problem profile are transformed into structured filtering conditions. Based on the initial modality recall strategy and the structured filtering conditions, a modality-based parallel retrieval is performed in the evidence unit set to obtain candidate evidence units for each modality.

4. The method according to claim 1, characterized in that, Also includes: The candidate answers are broken down into multiple factual fragments; each factual fragment includes at least numerical fragments, trend fragments, causal fragments, time fragments, and policy constraint fragments. Each fact fragment is traversed, and the corresponding back-lookup modality is determined according to the type of the fact fragment; the back-lookup modality is used to characterize the priority back-lookup evidence modality corresponding to the type of the fact fragment and its back-lookup order; Based on the back lookup mode, evidence units related to the fact fragment are retrieved in the target evidence chain or the candidate evidence chain, and entity matching relationship, evidence support relationship and evidence conflict relationship between the fact fragment and the retrieved evidence units are determined. Based on the entity matching relationship, the evidence support relationship, and the evidence conflict relationship, calculate the fragment consistency score of the fact fragment; Based on the segment consistency score of each fact segment and the pre-set segment type weight, the overall answer consistency score of the candidate answer is calculated as the segment-level consistency backcheck result of the candidate answer.

5. The method according to claim 4, characterized in that, Determining the corresponding lookup mode based on the type of the fact fragment includes: When the factual fragment is the numerical fragment, the table evidence unit is checked back first, and the unit, currency and statistical period are verified. When the factual fragment is the trend fragment, the chart evidence unit is checked back first, and the trend direction, range, and inflection point are verified. When the factual fragment is the cause fragment, the textual evidence unit is checked back first, and the tabular evidence unit or the chart evidence unit is used as auxiliary support. When the fact segment is the time segment, prioritize checking the title, header, table header, reporting period field, or effective date of the regulation; When the factual fragment is a policy constraint fragment, prioritize reviewing the main text of the regulation, appendices, and footnotes.

6. The method according to claim 4, characterized in that, Also includes: Based on the fragment-level consistency back lookup results of the candidate answers, the candidate answers are locally corrected; The local correction of the candidate answers includes: Determine whether the segment consistency score of the factual segment is lower than a first threshold; If so, the factual fragment is determined to be insufficiently supported by evidence, and a partial re-search is performed on the factual fragment to supplement the evidence units or evidence chains related to the factual fragment; otherwise, the factual fragment remains unchanged. Determine whether the consistency score of the factual fragment is lower than a second threshold, and whether the degree of evidence conflict corresponding to the factual fragment is higher than a preset conflict threshold; If so, the factual fragment is determined to have a conflict of evidence, and the factual fragment is partially replaced or partially deleted; otherwise, the factual fragment remains unchanged. When the factual fragment is the cause fragment, determine whether the cause fragment lacks numerical or trend evidence as corroborating evidence; If so, add an uncertainty explanation to the cause segment, or delete the attribution content that lacks supporting evidence; otherwise, keep the cause segment unchanged. Based on the corrected factual fragments, a corrected financial question and answer is generated, and the corresponding cited evidence, evidence chain path, evidence modality type, fragment consistency score, and local correction record are output simultaneously.

7. The method according to claim 1, characterized in that, The multiple candidate evidence chains are fused and sorted to determine the target evidence chain, which includes: Traverse each candidate evidence chain and calculate the basic matching score, evidence chain integrity score, cross-modal mutual evidence support, and conflict penalty item corresponding to the candidate evidence chain. The basic matching scores mentioned above include semantic matching score, keyword matching score, subject matching score, time matching score, indicator matching score, unit, period and statistical caliber matching score, source credibility score, parsing quality score and page adjacency or structural affiliation support score. The evidence chain integrity score is used to characterize the degree to which the candidate evidence chain covers the target subject, time range, financial indicators, statistical caliber, unit, statistical period, question type, and answer constraints. The cross-modal mutual evidence support is used to characterize the degree of mutual support between each modal evidence unit in the candidate evidence chain; The conflict penalty term is used to characterize the degree of conflict between modal evidence units in the candidate evidence chain; Based on the basic matching score, the evidence chain integrity score, the cross-modal mutual evidence support, and the conflict penalty term, calculate the fusion ranking score corresponding to the candidate evidence chain; Based on the fusion ranking score of each candidate evidence chain, the multiple candidate evidence chains are sorted in descending order to obtain the candidate evidence chain ranking result. One or more evidence chains are selected as the target evidence chain based on the candidate evidence chain ranking results.

8. A financial question answer generation system, characterized in that, include: The offline graph building module is used to perform multimodal parsing of financial documents to construct evidence unit sets, evidence indexes, and evidence unit relationship graphs. An online question-and-answer module, connected to the offline mapping module, is used to perform the method as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, include: The memory is used to store computer programs; A processor for executing a computer program stored in the memory to cause the electronic device to perform the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images