A method for reporting disassembly and semantic verification

Abstract

The present application relates to enterprise report verification, in particular to a report disassembly and semantic verification method, obtaining an original report and converting it into an original report text; obtaining the rules and guidelines of the report from the exchange and converting them into a disclosure matter set; converting the original report text into a proposition set; using a large language model to classify the proposition set to obtain an explanation class subset, a disclosure class subset and a residual item subset; performing forward matching and reverse matching on the explanation class subset, the disclosure class subset and the disclosure matter set to obtain a proposition set to be explained, a proposition set to be disclosed, an explained requirement set and a disclosed requirement set; The technical scheme provided by the present application can effectively overcome the defects of the prior art that it is difficult to accurately check the items of the enterprise report, and the traceability compliance review.

Description

Technical Field

[0001] This invention relates to enterprise report verification, specifically to a report decomposition and semantic verification method. Background Technology

[0002] As regulatory rules, industry standards, and internal corporate policies become increasingly detailed, companies are required to disclose or explain relevant regulatory requirements through environmental, social, and governance reports, as well as other types of reports such as sustainability reports, annual reports, and compliance reports. Regulatory agencies, auditing firms, and corporate compliance personnel need to verify whether the report content responds to the corresponding regulatory requirements.

[0003] Corporate reports are typically in unstructured or semi-structured document formats such as PDFs, and may contain continuous text, tables, multi-column layouts, mixed text and graphics, and content spanning multiple pages. Regulatory rules, on the other hand, are usually in structured formats such as clauses, guidelines, indicators, and lists. Therefore, corporate report verification requires establishing a correspondence between the report content and specific regulatory requirements to support compliance review, audit verification, and internal management.

[0004] Current corporate report verification typically employs a combination of manual review and machine-assisted analysis. Manual review usually involves auditors reading the report and judging whether it contains relevant content based on rules, clauses, or checklists. This method allows for contextual judgment, but it requires significant time and manpower when reports are lengthy, have numerous regulatory requirements, or need to compare multiple corporate or annual reports. Furthermore, different auditors may have different understandings and judgment standards for the same text, affecting the consistency and repeatability of the verification results.

[0005] Machine-aided analysis typically requires converting corporate reports into machine-processable text or structured text first. For complex reports with multi-column layouts, tables, and mixed text and graphics, traditional OCR or PDF parsing methods struggle to accurately reconstruct the original reading order and page structure, thus affecting the accuracy of subsequent information extraction and text comprehension.

[0006] Furthermore, individual clauses in regulatory rules may contain multiple independently verifiable information elements. For example, a governance responsibility requirement may simultaneously involve management mechanisms, assessment processes, prioritization processes, and other aspects. If the entire rule clause is used as a single matching object, it is difficult to further determine which information elements the report has responded to and which information elements have been omitted, thus hindering the acquisition of refined, item-level verification results.

[0007] Content related to a regulatory requirement in a company's report may be scattered across different sentences, paragraphs, or tables. Even within the same sentence or paragraph, information may include factual data, scope of application, calculation methods, limitations, and follow-up arrangements. Matching solely based on sentences, paragraphs, or fixed-length text fragments makes it difficult to accurately determine which content constitutes direct evidence of a specific regulatory requirement and which merely serves as limitation, explanation, or background information. This increases the difficulty of verifying regulatory rules at the item level and locating textual evidence.

[0008] Corporate reports may contain texts of different natures, including direct disclosures, explanatory texts, and background information. Existing methods, if they do not differentiate between these different types of texts, may misclassify explanatory or background information as direct disclosures, leading to inaccurate assessments of regulatory compliance. Summary of the Invention

[0009] (a) Technical problems to be solved In view of the above-mentioned shortcomings of the existing technology, the present invention provides a report decomposition and semantic verification method, which can effectively overcome the shortcomings of the existing technology in that it is difficult to conduct accurate item-level and traceable compliance review of enterprise reports.

[0010] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A report decomposition and semantic verification method includes the following steps: S1. Obtain the original report and convert it into original report text; S2. Obtain the exchange's rules and guidelines for reporting and convert them into a set of disclosure requirements; S3. Convert the original report text into a set of propositions; S4. Use a large language model to classify the proposition set to obtain the explanatory subset, the disclosure subset, and the residual item subset; S5. Perform forward and backward matching on the explanatory subset, the disclosure subset, and the disclosure item set to obtain the set of propositions to be explained, the set of propositions to be disclosed, the set of explained requirements, and the set of disclosed requirements.

[0011] Preferably, in S1, the original report is obtained and converted into original report text, including: S11. Input the enterprise's environmental, social, and corporate governance report, render the report page by page, and generate page images for each page; S12. Input all page images into the document structure parsing module. This module is based on a visual-language model and integrates layout detection, text recognition, and relationship understanding. The visual-language model performs three operations simultaneously: detects and classifies layout elements including titles, body paragraphs, tables, figure captions, headers and footers; identifies the text content corresponding to each layout element; and sorts out and determines the reading order and hierarchical relationship of all layout elements. S13. Generate hierarchical structured text through parsing and processing. This text includes a document-level heading tree, chapter levels, paragraph units, table row and column relationships, and positional associations between figure captions and corresponding body text. This text adopts a machine-processable format with hierarchical tags, preserving the relative positions and hierarchical relationships between page elements. S14. For all extracted text units, record their source location information in the original report one by one. The source location information includes page number, coordinate range, and logical chapter path. Finally, generate and output the original report text T.

[0012] Preferably, in S2, the exchange's rules and guidelines for reporting are obtained and converted into a set of disclosure items, including: S21. Input the exchange's rules and guidelines for corporate environmental, social, and governance reports, and identify the disclosure carrier types corresponding to each clause. The disclosure carrier types include board statements, policy documents, quantitative indicators, target commitments, and governance structures. S22. Deconstruct each clause and extract all independently verifiable information elements from the clauses. Independent information elements include action elements and information elements. S23. Rewrite the extracted information elements into self-contained verification questions. Each verification question is configured with a clear subject and predicate. The disclosure status is defined by binary verification logic. If the information element is clearly disclosed, it is determined by "yes". If the information element is not disclosed, it is determined by "no". S24. According to the preset proposition decomposition criteria, standardize and organize all verification questions, limit each disclosure item to correspond to only a single verifiable statement, eliminate compound requirement expressions, avoid multi-element nested item structures, and finally generate and output the disclosure item set R. The disclosure set R contains multiple independently verifiable binary verification questions, each accompanied by the relevant clause identifier and the original clause text.

[0013] Preferably, in S3, the original report text is converted into a set of propositions, including: S31. Input the original report text T, traverse all sentences in the text unit and perform semantic analysis to identify whether there are separable semantic components, including multiple parallel predicates, multiple purpose adverbs, and hierarchical modifiers. S32. If separable semantic components are identified, the sentence is split according to the preset propositional processing constraints. The constraints include three aspects: each proposition corresponds to a single independent semantic unit; each proposition has atomic properties and cannot be further split without destroying atomicity; each proposition adopts an independently interpretable contextual sentence structure, sets a clear subject, and retains the resolution results of the referential relationship. S33. Rewrite the splitting results into short statements to form atomic propositions; S34. Bind source location information to each atomic proposition. The source location information includes page number, paragraph, and original sentence. At the same time, record the necessary context. The necessary context includes the original paragraph where the atomic proposition is located, adjacent text, and the title of the chapter to which it belongs. Finally, generate and output the proposition set P. This step involves calling a language model with controlled semantic decomposition capabilities to complete all processing. During the operation of the language model, intermediate inference processes can be retrieved to check whether the sentence splitting process generates incorrect inferences due to deviations in understanding the original text.

[0014] Preferably, in S4, a large language model is used to classify the proposition set, resulting in an explanatory subset, a disclosure subset, and a residual item subset, including: S41. Retrieve each atomic proposition from the proposition set P one by one, bind its corresponding necessary context, and input them into the large language model. S42. Classify each atomic proposition according to a preset functional classification system, which includes three mutually exclusive proposition categories: The explanatory subset E is used to argue, limit, and explain related matters, without providing specific disclosure content for propositions, including narrative reasons for failure to achieve the goal and qualitative descriptions without quantitative indicators; Subset D of the disclosure category includes propositions that explicitly record measurable values, key performance indicators, work objectives, management systems, public commitments, governance structures, and other specific reportable matters, including numerical data, named indicators, and publicized policies. Residual propositions, which do not belong to the explanatory subset E or the disclosure subset D, include general corporate introductions, slogans, and broad statements without substantive operational content;

[0015] In this step, a large language model with instruction following and reasoning analysis capabilities is used as a classifier to output a unique category label for each atomic proposition. This classification system is limited to a three-category structure to adapt to subsequent semantic matching business needs, which is different from the general multi-level subdivision classification method.

[0016] Preferably, during the method development phase, the intermediate reasoning process of the large language model classification is retrieved and verified, and the classification failure modes of each category are statistically identified. The classification failure modes include situations where qualitative argumentation content is misjudged as quantitative disclosure content and visionary statements are misjudged as valid disclosure basis. The classification judgment criteria are iteratively calibrated and optimized based on the classification failure modes.

[0017] Preferably, in S5, forward and backward matching are performed on the explanatory subset, the disclosure subset, and the disclosure item set to obtain the set of propositions to be explained, the set of propositions to be disclosed, the set of explained requirements, and the set of disclosed requirements, including: S51. Input the set of disclosure items R, the subset of explanatory classes E, and the subset of disclosure classes D, and define the intermediate variables including the set of explanatory propositions, the set of disclosable propositions, the set of explained requirements, and the set of disclosed requirements; S52. Forward matching, the matching direction is based on the proposition pointing to regulatory disclosure requirements: S521. Traverse all propositions in the subset E of the explanation class and combine them with each disclosure item in the set R of disclosure items to form proposition-disclosure item pairs; traverse all propositions in the subset D of the disclosure class and combine them with each disclosure item in the set R of disclosure items to form proposition-disclosure item pairs. S522. Use a language model that pairs all propositions and disclosures as inputs to perform binary semantic judgment. The judgment criterion is the semantic implication relationship under category constraints, and literal similarity is not used as the judgment criterion. Two types of judgment criteria are distinguished: Explanation of matching criteria: If a proposition can explain, limit, or demonstrate the subject matter corresponding to the disclosure, the proposition shall not be used as a direct basis for disclosure; Disclosure matching criteria: The proposition fully provides the disclosure content requested for the disclosure matter, and the requested disclosure content includes policies, governance arrangements, indicators, objectives and other specific reportable matters; S523. For each proposition-disclosure item pairing, the language model only outputs a Boolean result of "yes" or "no". If the matching result is "yes", the propositions in the pairing are classified into the set of propositions to be explained or the set of propositions to be disclosed; if the matching result is "no", the pairing association is not recorded. S53. Reverse matching, matching the direction of the regulatory disclosure requirements: S531. Filter all proposition-disclosure pairings where the matching result in the forward matching is "yes"; S532. For proposition-disclosure item pairs that are classified as part of the set of propositions to be explained, the disclosure items in the pair shall be classified as part of the set of explained requirements; for proposition-disclosure item pairs that are classified as part of the set of propositions to be disclosed, the disclosure items in the pair shall be classified as part of the set of disclosed requirements. S533. Reverse matching is used to establish a source-tracing association centered on the disclosure item. For any disclosure item, all propositions that provide matching support for it can be retrieved, as well as the source location information of each proposition in the original report. Finally, a set of propositions to be explained, a set of propositions to be disclosed, a set of explained requirements, and a set of disclosed requirements are generated and output.

[0018] Preferably, in S51, intermediate variables including the set of explanatory propositions, the set of propositions to be disclosed, the set of explained requirements, and the set of disclosed requirements are defined, including: The set of explanatory propositions should be defined as a subset of explanatory propositions that match at least one disclosure under forward matching. The set of disclosure propositions: a subset of disclosure propositions that match at least one disclosure matter under forward matching; Explained requirements set: A subset of disclosures supported by corresponding explanatory proposition matching under reverse matching; The set of disclosed requirements is a subset of disclosure items that can be supported by matching corresponding disclosure propositions under reverse matching.

[0019] Preferably, based on the bidirectional semantic matching results output by S5, three verification indicators are generated through set operations and numerical calculations to quantify the completeness of the enterprise report's response to the disclosed matters: 1) The set of silent requirements N: ; Where N is the set of silent requirements, representing the set of regulatory disclosure requirements that have neither disclosure-based nor explanation-based support. This represents the set union operation. This represents the set difference operation; 2) Percentage of SDRs disclosed without prior notice: ; SDR represents the percentage of silent disclosures; a smaller SDR value indicates fewer unresponsive regulatory disclosure requirements. To determine the number of elements in a set N that is to be silent, The number of elements in the set R of disclosures; 3) Explain the dependency ratio (EDR): ; Here, EDR represents the percentage of reliance on explanatory responses. The smaller the EDR value, the lower the reliance on explanatory responses and the higher the tendency to provide substantive disclosures rather than narrative arguments. This represents the set of disclosures that receive only explanatory support but not disclosure support.

[0020] Preferably, the method ultimately outputs the response status of each disclosure, the propositions supporting each disclosure and their source location information, the silent disclosure ratio (SDR), the explanation dependency ratio (EDR), and a traceable chain of evidence, which are used for compliance review, audit verification, and internal management.

[0021] (III) Beneficial Effects

[0022] Compared with the prior art, the report decomposition and semantic verification method provided by the present invention has the following beneficial effects:

[0023] 1) Adopting a dual-sided atomized decomposition of text and regulatory requirements, the analysis granularity is refined to eliminate ambiguity in judgment caused by the mixing of complex elements.

[0024] This invention atomizes both the text of the verification report and the regulatory disclosure requirements: on the one hand, it breaks down the original text of the enterprise report into atomic propositions with independent semantics; on the other hand, it breaks down the complex disclosure requirements in regulatory rules and guidelines into single, independently verifiable atomic disclosure items. Based on the results of this dual-sided atomization, this solution transforms the coarse-grained comparison mode of the original text and complex regulatory requirements in traditional technologies into a fine-grained comparison mode of one-to-one correspondence between atomic propositions and atomic disclosure items. This effectively avoids problems such as recognition ambiguity and judgment bias caused by the mixing and nesting of multiple semantic contents and multiple requirement elements within a single processing unit, significantly improving the processing accuracy and analysis granularity of the entire process of text recognition, proposition classification, and semantic matching.

[0025] 2) Contextual aids are used to complete proposition recognition and classification, avoiding the shortcomings of isolated interpretation and improving the reliability of classification results.

[0026] This invention, in the process of semantic recognition and functional classification of atomic propositions, does not rely on isolated text fragments for judgment, but rather conducts collaborative analysis by combining necessary contextual information such as the original paragraph to which the atomic proposition belongs, adjacent texts, and the title of the chapter to which it belongs. By using contextual information to assist a large language model in accurately analyzing the actual semantics, referential relationships, limiting conditions, expressive purpose, and contextual attributes of atomic propositions, it can effectively distinguish text content with different functional attributes, accurately classifying disclosure propositions used to provide substantive disclosure information, explanatory propositions used to explain arguments and conditions, and redundant general text content without effective supporting function. This solution effectively solves the technical problems of ignoring contextual semantics, referential ambiguity, and missing limitations in the interpretation of single text fragments, significantly improving the accuracy and stability of the functional classification results of atomic propositions.

[0027] 3) Matching is based on semantic implication, moving beyond literal constraints and improving the accuracy and robustness of matching complex text.

[0028] This invention abandons the limitations of traditional keyword matching, word frequency comparison, literal overlap comparison, and fixed template matching. Instead, it combines atomic proposition semantics, contextual information, proposition function classification results, and the core requirements of atomic disclosure items, using semantic implication as the core judgment criterion to complete the matching verification between propositions and regulatory requirements. For propositions and disclosure items with different sentence structures, vocabulary, and expression forms but consistent core semantics, a legitimate matching relationship can be effectively established. For content with similar keywords but differences in core elements such as subject, applicable object, coverage, time dimension, hierarchical logic, and semantic tendency, erroneous matching can be accurately avoided. This adapts to the complex narrative and flexible expression characteristics of corporate reports, significantly improving the accuracy and anti-interference ability of matching results between unstructured narrative text and standardized regulatory requirements.

[0029] 4) Construct a two-way semantic matching and association system to achieve full-link traceability and verification, and reduce the risk of matching omissions.

[0030] This invention establishes a bidirectional semantic matching mechanism of forward and reverse matching, forming a two-way closed-loop association. Forward matching, centered on atomic propositions, iterates through and matches each atomic disclosure item to determine the regulatory disclosure requirements that a single proposition can support and correspond to. Reverse matching, centered on atomic disclosure items, backtracks and matches all corresponding atomic propositions, simultaneously associating the propositions with page numbers, paragraphs, chapters, and other traceability information in the original report. This bidirectional semantic matching mechanism supports both locating corresponding regulatory compliance requirements from the report text perspective and tracing back to the corresponding original report basis and context from the regulatory requirement perspective. It effectively avoids information omissions and missing associations caused by one-way matching, providing complete and traceable data support for compliance verification, manual review, result verification, and audit tracing, thereby improving the standardization and credibility of the overall verification process. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the process for constructing the set of disclosure items in this invention; Figure 3 This is a schematic diagram of the process for constructing a proposition set in this invention; Figure 4 This is a flowchart illustrating the process of classifying the proposition set in this invention; Figure 5 This is a flowchart illustrating the bidirectional semantic matching process for the classification results of the proposition set in this invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] The following describes the specific process of the report decomposition and semantic verification method provided by this invention, using concrete examples (such as...). Figure 1 (as shown) and technical effects.

[0035] S1. Obtain the original report and convert it into original report text, including: S11. Input the company's environmental, social, and governance report (including but not limited to PDF files, scanned images, and electronic documents), render the report page by page, and generate page images for each page; S12. Input all page images into the document structure parsing module. This module is based on a visual-language model and integrates layout detection, text recognition, and relationship understanding. The visual-language model performs three operations simultaneously: detects and classifies layout elements including titles, body paragraphs, tables, figure captions, headers and footers; identifies the text content corresponding to each layout element; and sorts out and determines the reading order and hierarchical relationship of all layout elements. S13. Generate hierarchical structured text through parsing and processing. This text includes a document-level heading tree, chapter levels, paragraph units, table row and column relationships, and positional associations between figure captions and corresponding body text. This text adopts a machine-processable format with hierarchical tags, preserving the relative positions and hierarchical relationships between page elements. S14. For all extracted text units, record their source location information in the original report one by one. The source location information includes page number, coordinate range, and logical chapter path. Finally, generate and output the original report text T.

[0036] Traditional methods based on PDF object stream or optical character recognition (OCR) tend to disrupt the reading order and hierarchy, or fragment table content into disordered segments, when handling multi-column layouts, tables, and mixed text and images. This step employs a visual-language model to perform end-to-end understanding of the page image, directly generating structured text from the image that preserves the layout structure and element relationships. This reduces semantic bias caused by the loss of layout information in subsequent proposition extraction and matching stages.

[0037] S2. Obtain the exchange's rules and guidelines for reporting and convert them into a set of disclosure requirements, such as... Figure 2 As shown, it includes: S21. Input the exchange's rules and guidelines for corporate environmental, social, and governance reports (including but not limited to regulatory rules, industry standards, corporate internal systems, and indicator lists), and identify the disclosure carrier types corresponding to each clause. Disclosure carrier types include board statements, policy documents, quantitative indicators, target commitments, and governance structures. S22. Deconstruct each clause and extract all independently verifiable information elements from the clauses. Independent information elements include action elements and information elements (such as "assessment", "priority determination", "management process" etc.). S23. Rewrite the extracted information elements into self-contained verification questions. Each verification question is configured with a clear subject and predicate. The disclosure status is defined by binary verification logic. If the information element is clearly disclosed, it is determined by "yes". If the information element is not disclosed, it is determined by "no". S24. According to the preset proposition decomposition criteria, standardize and organize all verification questions, limit each disclosure item to correspond to only a single verifiable statement, eliminate compound requirement expressions, avoid multi-element nested item structures, and finally generate and output the disclosure item set R. The disclosure set R contains multiple independently verifiable binary verification questions, each accompanied by the relevant clause identifier and the original clause text.

[0038] Regulatory requirements often appear in complex forms. If a complete clause is used as a single matching object, it is difficult to further determine which information elements the report has responded to and which information elements have been omitted. This step decomposes complex clauses into multiple independent binary verification items through atomic decomposition, improving the granularity of subsequent matching judgments and reducing matching ambiguities caused by complex requirements.

[0039] S3. Convert the original report text into a set of propositions, such as Figure 3 As shown, it includes: S31. Input the original report text T, traverse all sentences in the text unit and perform semantic analysis to identify whether there are separable semantic components, including multiple parallel predicates, multiple purpose adverbs, and hierarchical modifiers. S32. If separable semantic components are identified, the sentence is split according to the preset propositional processing constraints. The constraints include three aspects: each proposition corresponds to a single independent semantic unit; each proposition has atomic properties and cannot be further split without destroying atomicity; each proposition adopts an independently interpretable contextual sentence structure, sets a clear subject, and retains the resolution results of the referential relationship. S33. Rewrite the splitting results into short statements to form atomic propositions; S34. Bind source location information to each atomic proposition. The source location information includes page number, paragraph, and original sentence. At the same time, record the necessary context. The necessary context includes the original paragraph where the atomic proposition is located, adjacent text, and the title of the chapter to which it belongs. Finally, generate and output the proposition set P. This step involves calling a language model with controlled semantic decomposition capabilities to complete all processing. During the operation of the language model, intermediate inference processes can be retrieved to check whether the sentence splitting process generates incorrect inferences due to deviations in understanding the original text.

[0040] Example of Atomic Proposition Decomposition: For the original text "Company A is a leading global internet technology company, dedicated to creating innovative products and services to enrich people's lives," this step breaks it down into three atomic propositions: 1) "Company A is a leading global internet technology company"; 2) "Company A is dedicated to creating innovative products and services"; 3) "Company A aims to enrich people's lives." This example illustrates how a sentence containing multiple semantic components can be decomposed into multiple atomic statements while preserving the original meaning and presenting it in an independently verifiable form.

[0041] When using sentences, paragraphs, or fixed-length text fragments as matching units, the same unit may contain a mixture of factual data, applicable scope, calculation methods, and restrictive conditions, making it difficult to accurately determine which content constitutes direct evidence for a specific requirement. This step, through propositional processing, transforms the original text into semantically atomized, independently judgeable units, providing a stable target for subsequent fine-grained matching.

[0042] S4. Using a large language model, classify the proposition set to obtain subsets of explanatory, disclosure, and residual items, such as... Figure 4 As shown, it includes: S41. Retrieve each atomic proposition from the proposition set P one by one, bind its corresponding necessary context, and input them into the large language model. S42. Classify each atomic proposition according to a preset functional classification system, which includes three mutually exclusive proposition categories: The explanatory subset E is used to argue, limit, and explain related matters, without providing specific disclosure content for propositions, including narrative reasons for failure to achieve the goal and qualitative descriptions without quantitative indicators; Subset D of the disclosure category includes propositions that explicitly record measurable values, key performance indicators, work objectives, management systems, public commitments, governance structures, and other specific reportable matters, including numerical data, named indicators, and publicized policies. Residual propositions, which do not belong to the explanatory subset E or the disclosure subset D, include general corporate introductions, slogans, and broad statements without substantive operational content; In this step, a large language model with instruction following and reasoning analysis capabilities is used as a classifier to output a unique category label for each atomic proposition. This classification system is limited to a three-category structure to adapt to subsequent semantic matching business needs, which is different from the general multi-level subdivision classification method.

[0043] For the above steps, during the method development phase, the intermediate reasoning process of the large language model classification is retrieved and verified, and classification failure modes of each category are statistically identified. Classification failure modes include situations where qualitative argumentation content is misjudged as quantitative disclosure content and visionary statements are misjudged as valid disclosure basis. Based on the classification failure modes, the classification judgment criteria are iteratively calibrated and optimized.

[0044] Report texts may simultaneously contain direct disclosures, explanatory content, and background information. Failure to differentiate between these different types of text may lead to misclassification of explanatory or background information as direct disclosures, resulting in inaccurate assessments of required responses. This step categorizes propositions into explanatory and disclosure categories based on their function, enabling subsequent matching to be performed within category constraints and distinguishing between "providing substantive information" and "providing explanations or background." Furthermore, by incorporating necessary context into the classification, it avoids judging solely based on isolated text fragments while ignoring semantic context, thus improving classification reliability.

[0045] S5. Perform forward and backward matching on the explanatory subset, the disclosure subset, and the set of disclosure items to obtain the set of propositions to be explained, the set of propositions to be disclosed, the set of explained requirements, and the set of disclosed requirements, such as... Figure 5 As shown, it includes: S51. Input the set of disclosure items R, the subset of explanatory classes E, and the subset of disclosure classes D, and define the intermediate variables including the set of explanatory propositions, the set of disclosable propositions, the set of explained requirements, and the set of disclosed requirements; S52. Forward matching, the matching direction is based on the proposition pointing to regulatory disclosure requirements: S521. Traverse all propositions in the subset E of the explanation class and combine them with each disclosure item in the set R of disclosure items to form proposition-disclosure item pairs; traverse all propositions in the subset D of the disclosure class and combine them with each disclosure item in the set R of disclosure items to form proposition-disclosure item pairs. S522. Use a language model that pairs all propositions and disclosures as inputs to perform binary semantic judgment. The judgment criterion is the semantic implication relationship under category constraints, and literal similarity is not used as the judgment criterion. Two types of judgment criteria are distinguished: Explanation of matching criteria: If a proposition can explain, limit, or demonstrate the subject matter corresponding to the disclosure, the proposition shall not be used as a direct basis for disclosure; Disclosure matching criteria: The proposition fully provides the disclosure content requested for the disclosure matter, and the requested disclosure content includes policies, governance arrangements, indicators, objectives and other specific reportable matters; S523. For each proposition-disclosure item pairing, the language model only outputs a Boolean result of "yes" or "no". If the matching result is "yes", the propositions in the pairing are classified into the set of propositions to be explained or the set of propositions to be disclosed; if the matching result is "no", the pairing association is not recorded. S53. Reverse matching, matching the direction of the regulatory disclosure requirements: S531. Filter all proposition-disclosure pairings where the matching result in the forward matching is "yes"; S532. For proposition-disclosure item pairs that are classified as part of the set of propositions to be explained, the disclosure items in the pair shall be classified as part of the set of explained requirements; for proposition-disclosure item pairs that are classified as part of the set of propositions to be disclosed, the disclosure items in the pair shall be classified as part of the set of disclosed requirements. S533. Reverse matching is used to establish a source-tracing association centered on the disclosure item. For any disclosure item, all propositions that provide matching support for it can be retrieved, as well as the source location information of each proposition in the original report. Finally, a set of propositions to be explained, a set of propositions to be disclosed, a set of explained requirements, and a set of disclosed requirements are generated and output.

[0046] Specifically, S51 defines intermediate variables corresponding to the set of explanatory propositions, the set of disclosable propositions, the set of explained requirements, and the set of disclosed requirements, including: The set of explanatory propositions should be defined as a subset of explanatory propositions that match at least one disclosure under forward matching. The set of disclosure propositions: a subset of disclosure propositions that match at least one disclosure matter under forward matching; Explained requirements set: A subset of disclosures supported by corresponding explanatory proposition matching under reverse matching; The set of disclosed requirements is a subset of disclosure items that can be supported by matching corresponding disclosure propositions under reverse matching.

[0047] Example of judgment criteria: For the disclosure matter "whether the board of directors discloses its oversight responsibility for environmental matters", the proposition "the ESG working group reports twice a year to the corporate governance committee" can be judged as an explanatory valid match, because it reveals the formal reporting relationship between the ESG-related institutions and the governance committee, providing contextual evidence for including environmental matters in governance oversight; while the proposition "the ESG working group manages and coordinates the company's environmental work" should be judged as a mismatch for the same disclosure matter, because it describes an operational-level function and does not establish a substantial semantic connection with the oversight responsibility for environmental matters at the board or committee level.

[0048] Existing methods that rely solely on one-way matching may suffer from omission risks and struggle to trace back to the corresponding text from the requirement side. This step combines forward and reverse matching to establish a bidirectional relationship between the text and requirement sides. This allows for both locating the corresponding requirement from the text and tracing back to the corresponding text and its context from the requirement, reducing the omission risks of one-way matching and providing traceable evidence for manual review, result verification, and auditing. Furthermore, matching is based on semantic implications rather than keywords, word frequency, or fixed templates. This allows for matching of propositions and requirements with different vocabulary and sentence structures but the same semantic meaning, and avoids incorrect matching of content with similar keywords but inconsistent subjects, objects, scope, time, or levels.

[0049] In the technical solution of this application, based on the bidirectional semantic matching results output by S5, three verification indicators are generated through set operations and numerical calculations to quantify the completeness of the enterprise report's response to the disclosed matters: 1) The set of silent requirements N: ; Where N is the set of silent requirements, representing the set of regulatory disclosure requirements that have neither disclosure-based nor explanation-based support. This represents the set union operation. This represents the set difference operation; 2) Percentage of SDRs disclosed without prior notice: ; SDR represents the percentage of silent disclosures; a smaller SDR value indicates fewer unresponsive regulatory disclosure requirements. To determine the number of elements in a set N that is to be silent, The number of elements in the set R of disclosures; 3) Explain the dependency ratio (EDR): ; Here, EDR represents the percentage of reliance on explanatory responses. The smaller the EDR value, the lower the reliance on explanatory responses and the higher the tendency to provide substantive disclosures rather than narrative arguments. This represents the set of disclosures that receive only explanatory support but not disclosure support.

[0050] This method ultimately outputs the overall response status of each disclosure (including disclosure response, explanation response, silence, etc.), supports the proposition and source location information of each disclosure, the percentage of silent disclosure (SDR), the percentage of explanation dependence (EDR), and a traceable chain of evidence, which can be used for compliance review, audit verification, and internal management.

[0051] This invention can be implemented on computing devices equipped with a central processing unit (CPU), a graphics processing unit (GPU), or a tensor processor (TPU). For the visual-language model inference stage, it is recommended to equip a GPU acceleration device with at least 16GB of video memory; for the large language model inference stage, it is recommended to equip a GPU acceleration device with at least 32GB of video memory or to adopt a multi-GPU parallel architecture. The computing device should also include memory for storing report files, regulatory requirement texts, and intermediate processing results.

[0052] The operating system can be a Linux-based operating system; the programming language can be Python; the deep learning framework can be PyTorch or a framework with equivalent functionality; the model inference service can be an inference engine that supports model weight loading and configurable inference parameters. The language model used in this invention is an open-weighted inference language model that supports instruction following and allows for inspection of the inference process. Its inference behavior is configurable during inference and is suitable for controlled semantic decomposition and pairwise judgment tasks.

[0053] To better illustrate the technical solution of this application, the following uses a company's reports for three consecutive years and a regulatory guideline as examples to illustrate the complete implementation process of this method.

[0054] (a) Data preparation

[0055] Obtain the company's environmental, social, and governance (ESG) reports for the first to third years, as well as the exchange's rules and guidelines for ESG reports. For each report document, use a visual-language document parsing module to render and parse it page by page, obtaining a hierarchical raw text containing a title tree, chapter levels, paragraph units, and table row and column relationships. Each text unit records the page number and chapter path. For regulatory guidance texts, perform atomic decomposition according to step S2, identify the disclosure carriers and information elements, and rewrite composite clauses into independent binary verification questions, resulting in M ​​disclosure items (e.g., 192), each with an identifier for the relevant clause.

[0056] (II) Proposition Extraction and Classification

[0057] For each report's original text, propositional processing is performed to identify sentences containing multiple parallel predicates, multiple purpose adverbs, or hierarchical modifiers and break them down into atomic propositions. The source location and necessary context (paragraph and adjacent text) of each proposition are recorded. For each atomic proposition, its necessary context is input into a language model as a classifier, and it is classified into three categories: explanatory, disclosure, and residual, resulting in an explanatory subset E and a disclosure subset D.

[0058] (III) Bidirectional semantic matching

[0059] For each proposition in the explanatory subset and the disclosure subset, a proposition-disclosure item pair is formed with M disclosure items. The input language model is used as a binary semantic judge for judgment. The judgment is based on semantic implication under category constraints: for explanatory matching, the proposition must meaningfully describe the matters involved in the disclosure item; for disclosure matching, the proposition must explicitly provide the disclosure content requested by the disclosure item. A Boolean judgment result is output for each proposition-disclosure item pair. When the result is "yes", the matching relationship is recorded, and the corresponding disclosure item is simultaneously recorded in either the explanatory requirement set or the disclosure requirement set. The proposition's source position is also retained to support backtracking.

[0060] (iv) Calculation of verification results

[0061] Calculate the company's set of silent requirements N, silent disclosure ratio (SDR), and explanation dependence ratio (EDR) for each year. Taking the data in this embodiment as an example, the SDR was 6.77% and the EDR was 16.67% in the first year; the SDR rose to 11.98% and the EDR rose to 21.35% in the second year; and the SDR fell back to 6.77% and the EDR fell back to 12.50% in the third year. The direction of change in these indicators is consistent with the direction of change in the company's external rating: the deterioration of the indicators in the second year corresponds to a downgrade in the external rating, and the improvement of the indicators in the third year corresponds to an upgrade in the external rating. This result demonstrates that the indicators generated by the method of this invention can reflect the year-on-year changes in the quality of report response and have traceability, meaning that the response status of each disclosure can be traced back to the specific proposition and its position in the original document.

[0062] (v) Results Output

[0063] The output includes the following verification results: a list of disclosures and their response status; the title and source location (page number, chapter path) for each disclosure; SDR and EDR indicator values; and a traceable chain of evidence. This output can be used for compliance review, audit verification, and internal management.

[0064] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for report decomposition and semantic verification, characterized in that: Includes the following steps: S1. Obtain the original report and convert it into original report text; S2. Obtain the exchange's rules and guidelines for reporting and convert them into a set of disclosure requirements; S3. Convert the original report text into a set of propositions; S4. Use a large language model to classify the proposition set to obtain the explanatory subset, the disclosure subset, and the residual item subset; S5. Perform forward and backward matching on the explanatory subset, the disclosure subset, and the disclosure item set to obtain the set of propositions to be explained, the set of propositions to be disclosed, the set of explained requirements, and the set of disclosed requirements.

2. The report decomposition and semantic verification method according to claim 1, characterized in that: The original report is retrieved in S1 and converted into original report text, including: S11. Input the enterprise's environmental, social, and corporate governance report, render the report page by page, and generate page images for each page; S12. Input all page images into the document structure parsing module. This module is based on a visual-language model and integrates layout detection, text recognition, and relationship understanding. The visual-language model performs three operations simultaneously: detects and classifies layout elements including titles, body paragraphs, tables, figure captions, headers and footers; identifies the text content corresponding to each layout element; and sorts out and determines the reading order and hierarchical relationship of all layout elements. S13. Generate hierarchical structured text through parsing and processing. This text includes a document-level heading tree, chapter levels, paragraph units, table row and column relationships, and positional associations between figure captions and corresponding body text. This text adopts a machine-processable format with hierarchical tags, preserving the relative positions and hierarchical relationships between page elements. S14. For all extracted text units, record their source location information in the original report one by one. The source location information includes page number, coordinate range, and logical chapter path. Finally, generate and output the original report text T.

3. The report decomposition and semantic verification method according to claim 2, characterized in that: S2 retrieves the exchange's rules and guidelines for reporting and converts them into a set of disclosure requirements, including: S21. Input the exchange's rules and guidelines for corporate environmental, social, and governance reports, and identify the disclosure carrier types corresponding to each clause. The disclosure carrier types include board statements, policy documents, quantitative indicators, target commitments, and governance structures. S22. Deconstruct each clause and extract all independently verifiable information elements from the clauses. Independent information elements include action elements and information elements. S23. Rewrite the extracted information elements into self-contained verification questions. Each verification question is configured with a clear subject and predicate. The disclosure status is defined by binary verification logic. "Yes" indicates that the information element has been clearly disclosed, and "No" indicates that the information element has not been disclosed or is only an implicit statement. S24. According to the preset proposition decomposition criteria, standardize and organize all verification questions, limit each disclosure item to correspond to only a single verifiable statement, eliminate compound requirement expressions, avoid multi-element nested item structures, and finally generate and output the disclosure item set R. The disclosure set R contains multiple independently verifiable binary verification questions, each accompanied by the relevant clause identifier and the original clause text.

4. The report decomposition and semantic verification method according to claim 3, characterized in that: In S3, the original report text is converted into a set of propositions, including: S31. Input the original report text T, traverse all sentences in the text unit and perform semantic analysis to identify whether there are separable semantic components, including multiple parallel predicates, multiple purpose adverbs, and hierarchical modifiers. S32. If separable semantic components are identified, the sentence is split according to the preset propositional processing constraints. The constraints include three aspects: each proposition corresponds to a single independent semantic unit; each proposition has atomic properties and cannot be further split without destroying atomicity; each proposition adopts an independently interpretable contextual sentence structure, sets a clear subject, and retains the resolution results of the referential relationship. S33. Rewrite the splitting results into short statements to form atomic propositions; S34. Bind source location information to each atomic proposition. The source location information includes page number, paragraph, and original sentence. At the same time, record the necessary context. The necessary context includes the original paragraph where the atomic proposition is located, adjacent text, and the title of the chapter to which it belongs. Finally, generate and output the proposition set P. This step involves calling a language model with controlled semantic decomposition capabilities to complete all processing. During the operation of the language model, intermediate inference processes can be retrieved to check whether the sentence splitting process generates incorrect inferences due to deviations in understanding the original text.

5. The report decomposition and semantic verification method according to claim 4, characterized in that: S4 uses a large language model to classify the proposition set, resulting in a subset of explanatory items, a subset of disclosure items, and a subset of residual items, including: S41. Retrieve each atomic proposition from the proposition set P one by one, bind its corresponding necessary context, and input them into the large language model. S42. Classify each atomic proposition according to a preset functional classification system, which includes three mutually exclusive proposition categories: The explanatory subset E is used to argue, limit, and explain related matters, without providing specific disclosure content for propositions, including narrative reasons for failure to achieve the goal and qualitative descriptions without quantitative indicators; Subset D of the disclosure category includes propositions that explicitly record measurable values, key performance indicators, work objectives, management systems, public commitments, governance structures, and other specific reportable matters, including numerical data, named indicators, and publicized policies. Residual propositions, which do not belong to the explanatory subset E or the disclosure subset D, include general corporate introductions, slogans, and broad statements without substantive operational content; In this step, a large language model with instruction following and reasoning analysis capabilities is used as a classifier to output a unique category label for each atomic proposition. This classification system is limited to a three-category structure to adapt to subsequent semantic matching business needs, which is different from the general multi-level subdivision classification method.

6. The report decomposition and semantic verification method according to claim 5, characterized in that: During the method development phase, the intermediate reasoning process of the large language model classification was retrieved and verified, and classification failure modes for each category were statistically identified. Classification failure modes include situations where qualitative argumentation content is misjudged as quantitative disclosure content and visionary statements are misjudged as valid disclosure basis. Based on the classification failure modes, the classification judgment criteria were iteratively calibrated and optimized.

7. The report decomposition and semantic verification method according to claim 5, characterized in that: S5 performs forward and backward matching on the explanatory subset, the disclosure subset, and the set of disclosure items to obtain the set of propositions to be explained, the set of propositions to be disclosed, the set of explained requirements, and the set of disclosed requirements, including: S51. Input the set of disclosure items R, the subset of explanatory classes E, and the subset of disclosure classes D, and define the intermediate variables including the set of explanatory propositions, the set of disclosable propositions, the set of explained requirements, and the set of disclosed requirements; S52. Forward matching, the matching direction is based on the proposition pointing to regulatory disclosure requirements: S521. Traverse all propositions in the subset E of the explanation class and combine them with each disclosure item in the set R of disclosure items to form proposition-disclosure item pairs; traverse all propositions in the subset D of the disclosure class and combine them with each disclosure item in the set R of disclosure items to form proposition-disclosure item pairs. S522. Use a language model that pairs all propositions and disclosures as inputs to perform binary semantic judgment. The judgment criterion is the semantic implication relationship under category constraints, and literal similarity is not used as the judgment criterion. Two types of judgment criteria are distinguished: Explanation of matching criteria: If a proposition can explain, limit, or demonstrate the subject matter corresponding to the disclosure, the proposition shall not be used as a direct basis for disclosure; Disclosure matching criteria: The proposition fully provides the disclosure content requested for the disclosure matter, and the requested disclosure content includes policies, governance arrangements, indicators, objectives and other specific reportable matters; S523. For each proposition-disclosure item pairing, the language model only outputs a Boolean result of "yes" or "no". If the matching result is "yes", the propositions in the pairing are classified into the set of propositions to be explained or the set of propositions to be disclosed. If the matching result is "no", the pairing association is not recorded. S53. Reverse matching, matching the direction of the regulatory disclosure requirements: S531. Filter all proposition-disclosure pairings where the matching result in the forward matching is "yes"; S532. For proposition-disclosure item pairs that are classified as part of the set of propositions to be explained, the disclosure items in the pair shall be classified as part of the set of explained requirements; for proposition-disclosure item pairs that are classified as part of the set of propositions to be disclosed, the disclosure items in the pair shall be classified as part of the set of disclosed requirements. S533. Reverse matching is used to establish a source-tracing association centered on the disclosure item. For any disclosure item, all propositions that provide matching support for it can be retrieved, as well as the source location information of each proposition in the original report. Finally, a set of propositions to be explained, a set of propositions to be disclosed, a set of explained requirements, and a set of disclosed requirements are generated and output.

8. The report decomposition and semantic verification method according to claim 7, characterized in that: S51 defines intermediate variables corresponding to the set of explanatory propositions, the set of disclosable propositions, the set of explained requirements, and the set of disclosed requirements, including: The set of explanatory propositions should be defined as a subset of explanatory propositions that match at least one disclosure under forward matching. The set of disclosure propositions: a subset of disclosure propositions that match at least one disclosure matter under forward matching; Explained requirements set: A subset of disclosures supported by corresponding explanatory proposition matching under reverse matching; The set of disclosed requirements is a subset of disclosure items that can be supported by matching corresponding disclosure propositions under reverse matching.

9. The report decomposition and semantic verification method according to claim 7, characterized in that: Based on the bidirectional semantic matching results output by S5, three verification indicators are generated through set operations and numerical calculations to quantify the completeness of the enterprise's report response to the disclosed matters: 1) The set of silent requirements N: ； Where N is the set of silent requirements, representing the set of regulatory disclosure requirements that have neither disclosure-based nor explanation-based support. This represents the set union operation. This represents the set difference operation; 2) Percentage of SDRs disclosed without prior notice: ； SDR represents the percentage of silent disclosures; a smaller SDR value indicates fewer unresponsive regulatory disclosure requirements. To determine the number of elements in a set N that is to be silent, The number of elements in the set R of disclosures; 3) Explain the dependency ratio (EDR): ； Here, EDR represents the percentage of reliance on explanatory responses. The smaller the EDR value, the lower the reliance on explanatory responses and the higher the tendency to provide substantive disclosures rather than narrative arguments. This represents the set of disclosures that receive only explanatory support but not disclosure support.

10. The report decomposition and semantic verification method according to claim 9, characterized in that: The method ultimately outputs the response status of each disclosure, supporting propositions and their source locations, silent disclosure ratio (SDR), explanation dependency ratio (EDR), and traceable evidence chain, which are used for compliance review, audit verification, and internal management.

Images