A knowledge graph-oriented regulation case association reasoning method

Abstract

This invention discloses a method for legal case association reasoning oriented towards knowledge graphs, specifically in the field of audit knowledge graphs, to address the problem of lacking verifiable evidence in legal case matching. By acquiring audit doubt texts, legal clause texts, historical case texts, and rectification measure texts, a set of doubt fact elements, a set of applicable legal conditions, a set of case fact elements, and a set of rectification measure elements are generated according to a preset audit fact element template. Using the set of applicable legal conditions as a comparison benchmark, condition coverage processing is performed on the case fact element set and the doubt fact element set to generate case condition coverage records and doubt condition coverage records, which are then written into condition instance relationships to obtain a legal case association graph. Candidate legal clauses, conditions requiring supplementary evidence, candidate cases, and candidate rectification measures are identified in the legal case association graph, forming candidate reasoning paths and outputting the audit qualitative reasoning results.

Description

Technical Field

[0001] This invention relates to the field of audit knowledge graphs, and more specifically, to a method for legal case association reasoning oriented towards knowledge graphs. Background Technology

[0002] This invention relates to the processing of legal and case-related reasoning in audit knowledge graphs. In auditing work, auditors typically need to summarize issues, find legal basis, compare historical cases, and formulate rectification suggestions based on various materials such as audit reports, audit working papers, rectification ledgers, contracts, documents, and historical project data regarding anomalies. These materials come from different sources, have different text structures, and use inconsistent terminology. For example, the same fund may be described as special fund, special subsidy fund, or project subsidy fund; the same violation may be described as fictitious expenditure, expenditure without actual business support, or fictitious meeting expenses. If relying solely on manual searching or keyword retrieval, auditors need to repeatedly verify between legal provisions, historical cases, and rectification requirements, a time-consuming process easily affected by inconsistencies in wording.

[0003] Existing knowledge graph methods can integrate regulations, cases, terminology, reports, and other objects into a unified management system and establish connections through entity recognition, relation extraction, knowledge fusion, and similarity retrieval. For example, patent document CN110334212A discloses a method for constructing a domain-specific audit knowledge graph based on machine learning. This method generates an audit knowledge graph by acquiring data from multiple sources, preprocessing, identifying entities, processing relations, using topic models, and employing machine learning classification. This improves the efficiency of retrieving and comparing audit regulations and cases. Such methods can improve the organization of audit knowledge resources, enabling unified retrieval and retrieval of regulatory clauses, audit cases, and related texts.

[0004] However, in audit qualitative scenarios, simply establishing entity relationships or textual similarity is insufficient to meet the requirements for judging the applicability of regulations. Audit doubts are typically not standardized question names, but rather constitute a combination of factual elements such as the type of audit matter, the audit object, the nature of the funds, the manner of the action, the approval process, the period of occurrence, the evidentiary materials, and the consequences. The applicability of regulatory provisions does not solely depend on similar titles or keywords, but requires determining whether the current facts cover the conditions of the applicable object, the nature of the funds, the elements of the action, the procedural conditions, the timing conditions, the evidentiary requirements, and the consequences. Existing methods often remain at the level of static relationship display, keyword search, or similarity recommendations, failing to explain why regulatory provisions are applicable, why historical cases are comparable, and why corrective measures correspond to the consequences.

[0005] Especially when audit evidence is incomplete, existing methods typically only provide similar regulations or cases, failing to differentiate between situations where facts meet the applicable conditions, lack supporting evidence, or conflict with regulations. For example, if the audit report only contains invoices and payment vouchers, without approval forms or acceptance documents, recommending regulations and cases based on similar issue names can easily lead to the overlooking of evidentiary gaps. Consequently, the lack of a verifiable field-level link between regulatory basis, comparable cases, supporting evidence, and rectification suggestions affects the accuracy, consistency, and verifiability of audit findings.

[0006] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a regulatory case association reasoning method oriented towards knowledge graphs. This method acquires audit doubt text, regulatory clause text, historical case text, and rectification measure text, and generates doubt fact element sets, regulatory application condition sets, case fact element sets, and rectification measure element sets according to a preset audit fact element template. Using the regulatory application condition set as a comparison benchmark, the case fact element sets and doubt fact element sets undergo condition coverage processing to generate case condition coverage records and doubt condition coverage records, which are then written into condition instance relationships to obtain a regulatory case association graph. Candidate regulatory clauses, conditions requiring supplementary evidence, candidate cases, and candidate rectification measures are identified within the regulatory case association graph, forming candidate reasoning paths and outputting audit qualitative reasoning results to address the problems raised in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: Obtain audit doubt texts, regulatory clause texts, historical case texts, and rectification measure texts, and generate doubt fact element sets, regulatory application condition sets, case fact element sets, and rectification measure element sets according to preset audit fact element templates; Perform same-field condition coverage processing on the set of applicable regulatory conditions and the set of case fact elements under the same audit matter category, generate case condition coverage records, establish condition instance relationships, and form a regulatory case relationship graph; The set of factual elements of doubt and the set of applicable legal conditions that are consistent with the audit matter category are subjected to the same field condition coverage process to generate doubt condition coverage record. Candidate legal clauses are determined based on the coverage status of the qualitative necessary conditions and the conflict status of the time conditions in the set of applicable legal conditions. Conditions that are missing and can be supplemented are extracted from the doubt condition coverage record. In the condition instance relationship corresponding to the candidate regulatory clauses, candidate cases are selected based on the field correspondence between the doubt condition coverage record and the case condition coverage record, and candidate rectification measures are selected based on the handling consequences conditions in the candidate regulatory clauses, forming candidate reasoning paths and generating audit qualitative reasoning results.

[0009] Furthermore, the pre-set audit fact element template includes audit item category, audit object, fund attribute, behavior method, approval process, occurrence period, evidence materials, and handling consequences arranged in a fixed field order; The set of applicable conditions for regulations includes conditions for the applicable objects, conditions for the nature of funds, conditions for the composition of the behavior, procedural conditions, time conditions, conditions for evidentiary requirements, and conditions for the consequences of handling. The conditions for applicable objects, financial attributes, and behavioral composition are qualitatively necessary conditions; the conditions for procedural conditions and evidentiary requirements are supplementary conditions; the conditions for handling consequences are result-corresponding conditions; and the time point conditions are independently retained for the period of application of the regulations.

[0010] Furthermore, the same-field condition coverage processing uses the case fact element set and the doubtful fact element set as the comparison element sets, respectively, including: Using the condition fields and condition restriction values ​​in the set of applicable regulations as the comparison benchmark, find the fact values ​​of the same fields in the set of elements being compared. If no fact value for the same field is found, the corresponding coverage status will be written as condition missing; When a fact value with the same field is found, the same value judgment, inclusion judgment, period fall-in judgment and normalized synonym merge judgment are executed in sequence. If any judgment is true, the coverage status is written as the condition is covered. If none of the judgments are true, the field exclusion relationship table is used to determine whether the fact value and condition restriction value of the same field are mutually exclusive. If they are mutually exclusive, the coverage status is written as condition conflict; if they are not mutually exclusive, the coverage status is written as condition missing.

[0011] Furthermore, establishing conditional instance relationships includes: Collect case condition coverage records under the same legal clause identifier and the same case identifier; When there are no conflicts between the qualitative necessary conditions and the time conditions, the legal provisions, legal application conditions, case fact elements, and case condition coverage records corresponding to the set of applicable legal conditions are written into the condition instance relationship. When there are missing conditions but no conflicting conditions in the qualitative necessary conditions, the case fact element set is written into the condition instance relationship as an incomplete coverage instance, and the corresponding missing condition state is retained. When any necessary condition for certainty has a conditional conflict or a conditional point in time has a conditional conflict, the corresponding set of case fact elements will not be written into the conditional instance relation.

[0012] Furthermore, when generating the set of factual elements of doubt and the set of factual elements of the case, a negative semantic determination is performed before the factual values ​​of the evidence material fields are written, including: When negative terms appear before or after a noun of evidence, the noun of evidence is not written into the obtained fact value of the evidence material field, and the noun of evidence is retained as a condition missing clue. When an affirmative term appears in the text segment containing the evidence term, the evidence term is written into the evidence material field; When the same evidence term has both positive and negative descriptions, the text segment containing at least one conclusive term from among "audit opinion," "problem manifestation," "handling decision," "rectification requirement," "recovery requirement," and "order to rectify" shall be identified as the audit conclusion segment; if no text segment contains conclusive terms, the last text segment in the corresponding text source that contains the behavior method or handling consequence shall be identified as the audit conclusion segment. Determine the text paragraph distance between the affirmative and negative descriptive sentences and the audit conclusion sentences, and use the description with the smaller distance as the basis for writing; when the distances are the same, use the negative description as a clue of missing conditions, and do not write the evidence terms into the obtained fact values.

[0013] Furthermore, the generation of the set of applicable regulatory conditions includes: The text of the legal provisions is broken down into conditional statements according to the clause number, item number, and conditional connector. The conditional fields, conditional limit values, and conditional roles are determined based on the semantic function of the conditional statements. When the same conditional statement contains both procedural requirements and evidentiary requirements, the conditional statement is split into procedural conditions and evidentiary requirement conditions according to the core nouns in the conditional statement, and the conditional roles of the procedural conditions and evidentiary requirement conditions are determined as supplementary evidence conditions respectively. When the same legal provision has multiple parallel acts constituting it, multiple sets of applicable conditions are formed according to the multiple parallel acts constituting it. Each set of applicable conditions shares the same legal provision identifier and has different condition restriction values ​​for the acts constituting it. When the text of a regulation only specifies the consequences of handling without recording the evidence requirements, the conditions for handling consequences should be written into the conditions corresponding to the results, and the conditions for evidence requirements should be kept in an empty state.

[0014] Furthermore, the candidate rectification measures include: Read the corresponding relationships of processing consequences, rectification actions, rectification targets, and rectification basis in sequence from the elements of the rectification measures; When the corresponding relationship of the handling consequences directly points to the handling consequences conditions in the candidate regulations, the set of rectification measures elements will be written into the candidate rectification measures. When the correspondence between the consequences and the conditions does not directly point to the conditions for handling consequences, compare whether the rectification action and the handling result in the conditions for handling consequences belong to the same standard expression, and compare whether the object of rectification and the object pointed to by the conditions for handling consequences are consistent. When the rectification action and the handling result belong to the same standard expression and the rectification object and the handling consequence condition point to the same object, the rectification measure element set is written into the candidate rectification measures; when the rectification basis points to the same legal clause identifier as the candidate legal clause, the rectification basis is used as the matching confirmation basis. When the rectification action and the handling result do not belong to the same standard expression, or the object of rectification and the object pointed to by the handling result conditions are inconsistent, the set of rectification measures elements shall not be written into the candidate rectification measures.

[0015] Furthermore, the legal case relationship graph uses legal clause nodes, legal application condition nodes, case fact nodes, and rectification measure nodes as graph nodes, and condition instance relationships as graph edges; Each condition instance relationship connects at least one legal clause node, one legal application condition node, and one case fact node. When there is a set of corrective measures elements that match the conditions for handling consequences, the condition instance relationship also connects the corresponding corrective measures node. The attributes of the graph include the correspondence between case condition coverage records and processing consequences. Case condition coverage records include legal clause identifiers, case identifiers, condition fields, condition limit values, corresponding fact values, and coverage status. The sets of factual elements of doubt, the sets of applicable legal conditions, the sets of factual elements of cases, and the sets of elements of rectification measures retain the text source identifier and the text paragraph position, and the text paragraph position is associated with the text sentence or segment from which the corresponding element originates.

[0016] Furthermore, the qualitative reasoning results of the audit include the characterization of the problem, the legal basis, comparable cases, supporting evidence, and rectification recommendations, among which: Read the behavior marked as covered in the doubtful condition coverage record as the factual behavior basis for the problem characterization, and read the processing consequence conditions in the candidate legal clauses to form the processing conclusion corresponding to the factual behavior basis; The candidate regulatory clauses will be used as the basis for the regulations, and the regulatory clause identifiers and the regulatory application conditions fields covered by the set of questionable fact elements will be retained. Read the case fact element set and case condition coverage record of the candidate case, output the audit object, fund attribute, behavior mode and occurrence period fields that are compatible with the suspicious fact element set, and retain the corresponding coverage status to form comparable cases; The conditional limits of the procedural conditions and evidentiary requirements in the conditions to be supplemented are converted into evidence materials or procedural facts that need to be obtained or verified, thus forming evidence reinforcement items; Read the rectification actions, rectification targets, and rectification basis from the candidate rectification measures, form rectification suggestions, and ensure that the rectification actions correspond to the conditions of the handling consequences; The problem characterization, legal basis, comparable cases, supporting evidence, and rectification suggestions each retain source path identifiers pointing to the corresponding candidate reasoning paths.

[0017] Furthermore, when there are multiple candidate reasoning paths under the same candidate legal provision, a sorting key is generated for each candidate reasoning path; The sorting key sequentially records whether the applicable object conditions, fund attribute conditions, and behavioral composition conditions in the corresponding case condition coverage record are all covered, whether the occurrence period field of the candidate case and the suspicious fact element set belong to the same time point condition limitation period, whether the candidate rectification measures are directly connected to the handling consequence conditions through the handling consequence correspondence relationship, and the writing time sequence number of the candidate case in the legal case association graph. Output the audit qualitative reasoning results corresponding to each candidate reasoning path according to the sorting key. Each audit qualitative reasoning result retains its source candidate reasoning path. Case condition coverage records of different candidate cases are not merged, and rectification measures that have not entered the candidate reasoning path are not included in the rectification suggestions.

[0018] The technical effects and advantages of the legal case association reasoning method based on knowledge graphs in this invention are as follows: This invention unifies the texts of audit concerns, regulatory clauses, historical cases, and rectification measures into a unified set of factual elements, applicable regulatory conditions, case factual elements, and rectification measures with consistent fields. This allows the fund attributes, behavior methods, approval processes, occurrence periods, evidence materials, and consequences, which were originally scattered across different materials, to be compared under the same field definition. Simultaneously, using the set of applicable regulatory conditions as a common benchmark for judging concerns, historical cases, and rectification measures, it generates concerns condition coverage records and case condition coverage records. This transforms the applicability of regulatory clauses, the comparability of historical cases, the absence of evidence materials, and whether rectification measures correspond to the consequences into clear states such as covered conditions, missing conditions, and conflicting conditions. This reduces mismatches of regulations and misuse of cases caused by relying solely on similar keywords or issue names. When audit evidence is incomplete, this method can extract the conditions to be supplemented from procedural conditions and evidence requirements, so that missing content such as approval documents and acceptance materials can directly form evidence reinforcement items. When outputting the audit qualitative reasoning results, the problem characterization, legal basis, comparable cases, evidence reinforcement items and rectification suggestions are all derived from the candidate reasoning path. It can retain the correspondence between the doubtful facts, legal provisions, historical cases and rectification measures, so that the audit qualitative process has more stable factual support and a clearer review chain. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the multi-source audit text processing and regulatory case correlation reasoning process of the present invention; Figure 2 This is a schematic diagram illustrating the field-level condition coverage process between the regulatory application condition set, the doubtful factual element set, and the case factual element set of the present invention. Figure 3 This is a schematic diagram illustrating the reasoning path for forming audit qualitative results based on candidate regulations, historical cases, conditions requiring supplementary certification, and candidate rectification measures in the scenario of fictitious expenditures of special funds in this invention. Detailed Implementation

[0020] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 - Figure 3 This invention provides a method for legal case association reasoning based on knowledge graphs, including: This invention processes audit-related texts, regulatory clauses, historical case texts, and rectification measures. First, it uses a pre-defined audit fact element template and audit terminology standardization table to convert texts from different sources into sets of audit-related fact elements, regulatory application conditions, case fact elements, and rectification measures elements with consistent fields and comparable values. Then, using audit matter categories as group boundaries, it performs field-based condition coverage processing on the regulatory application conditions set and the case fact element set, generating case condition coverage records and forming condition instance relationships containing regulatory clauses, case fact element sets, case condition coverage records, and rectification measure element sets, thus obtaining a regulatory-case association graph. Subsequently, the audit-related fact element set is input into the regulatory-case association graph to generate audit-related condition coverage records, thereby determining candidate regulatory clauses and conditions requiring supplementary evidence. Finally, in the condition instance relationships corresponding to the candidate regulatory clauses, candidate cases are screened based on the field correspondence between the audit-related condition coverage records and the case condition coverage records, and candidate rectification measures are screened based on the processing consequence conditions, forming candidate reasoning paths and outputting the audit qualitative reasoning results.

[0022] In this implementation, the data processing server uses audit-related texts, regulatory clause texts, historical case texts, and rectification measure texts obtained from external input or the audit knowledge base as processing objects. It uses the same field comparison criteria required for subsequent condition coverage processing in step S2 as output constraints, first converting texts from different sources into structured element sets with consistent fields and comparable values. However, the original expressions of each text have inconsistencies in matter names, subject titles, funding descriptions, and evidence descriptions. Therefore, before entering the construction of the regulatory case association graph, it is necessary to complete the extraction of factual elements, condition decomposition, and terminology normalization processing, so that the set of audit-related factual elements, the set of applicable regulatory conditions, the set of case factual elements, and the set of rectification measure elements can be directly called by subsequent steps.

[0023] S101: Obtain multi-source text and form an audit fact element extraction sequence.

[0024] At the start of step S1, the data processing server receives audit suspicion text, regulatory clause text, historical case text, and rectification measure text, and records the text source identifier, text paragraph position, and initial value of audit matter category according to the text source. The text source identifier is used to distinguish whether the current text comes from suspicion, regulations, cases, or rectification measures; the text paragraph position is used to trace back the corresponding statement in subsequent coverage records; and the initial value of audit matter category is used to limit the comparison range under the same audit matter category in subsequent records.

[0025] The initial value of the audit matter category is determined in the order of explicit labeling, directory source, and keyword matching. If the text contains tags such as fiscal funds, government procurement, special funds, or state-owned assets, then those tags are used directly. If the text originates from a fixed directory in an audit report, legal database, or case database, then the audit matter category corresponding to that directory is used. If the text does not contain any matter tags or directory source, the data processing server performs keyword matching on the text segment sequence based on the audit matter category terms, and uses the category corresponding to the matched terms as the initial value of the audit matter category. If the same text matches multiple audit matter categories, the audit matter category with the more matched behavior mode fields and fund attribute fields is used as the initial value of the audit matter category.

[0026] When the same text matches multiple audit matter categories via keyword matching, the data processing server determines the initial value of the audit matter category based on the audit matter category score. For any audit matter category... Their scores are determined according to the following relationship: ; in, Category of audit matters Category scores; For the text segment sequence that matches the audit matter category term and the corresponding category is The number of times; Matches and categorizes in the fund attribute field. The number of terms; Hit and categorize in the behavior field The number of times the terms are used. The data processing server will... The highest audit matter category is written into the initial value of the audit matter category; if multiple audit matter categories have the same score, the initial value of the audit matter category is determined in the following order: explicit label, directory source, and the position of the earliest appearing text paragraph.

[0027] The data processing server uses the audit item category, audit object, fund attribute, behavior method, approval process, occurrence period, evidence materials, and handling consequences as fixed field order in a pre-defined audit fact element template. It then performs sentence segmentation, term segmentation, and keyword positioning for each type of text. Sentence segmentation uses punctuation marks, clause numbers, case fact paragraphs, and rectification requirement paragraphs as segmentation boundaries to obtain a sequence of text sentences. Term segmentation extracts main nouns, fund nouns, action verbs, program nodes, time expressions, evidence nouns, and result statements from the text sentences as candidate term sequences. Keyword positioning follows the field order of the pre-defined audit fact element template, placing terms from the candidate term sequences that point to the meaning of the fields into the confirmation position of the corresponding fields.

[0028] Sentence segmentation is performed using periods, semicolons, line breaks, clause numbers, case fact paragraph titles, and rectification requirement paragraph titles. When consecutive short sentences describe the same audit behavior and no new audit object, fund attribute, or behavior method appears, the data processing server retains the consecutive short sentences in the same text segment. When a new audit object, fund attribute, or behavior method appears in a subsequent short sentence, the data processing server segments it into a new text segment.

[0029] Terms segmentation is performed using a field dictionary matching method. The field dictionary includes a basic field dictionary and a condition-extended field dictionary. The basic field dictionary is generated from the field names, example terms, and non-standard and standard expressions in the pre-defined audit fact element template and the audit terminology standardization table. It is used in S101 to perform initial term segmentation on audit doubt texts, regulatory clause texts, historical case texts, and rectification measure texts. The condition-extended field dictionary is formed after the regulatory application condition set is generated in S103. It consists of the basic field dictionary and the condition restriction values ​​in the regulatory application condition set, and is used for comparison of the same field in S2 and S3. The basic field dictionary is denoted as... The extended field dictionary for conditions is denoted as The relationship is as follows: ; in, This is a set of field terms derived from a pre-defined audit fact element template. This is a set of non-standard and standard expressions derived from the Audit Terminology Standardization Table. This is a set of conditional restriction values ​​derived from the set of applicable legal conditions. S101 calls this when performing term segmentation. After S103 completes the generation of the set of applicable regulatory conditions, it will... Incorporation It is available for use by S2, S3 and S4.

[0030] The data processing server performs a maximum forward match according to the character order of the text segments. For the first character in the text segment... Character positions, if field dictionary If a word exists that can be matched starting from that character position, the current matching word is determined according to the following relationship: ; in, For the first The matched terms output at each character position. For the currently invoked field dictionary, retrieve from S101 Choose from S2, S3 and S4 ; For the current text segment, Indicates terms With text segments China The consecutive characters starting with each character are completely identical; For terms The character length is specified. If multiple matching terms of the same length exist, the term whose field source is consistent with the current text source is prioritized; if they still cannot be distinguished, the term written earlier in the field dictionary is used. After a successful match, the data processing server writes the matching term, its corresponding field, and the position of the text paragraph into the candidate term sequence, and continues matching from the next character after the last character of the matching term.

[0031] Keyword positioning prioritizes the field in the candidate term sequence as the primary basis for attribution, writing the candidate terms into the corresponding fields in the preset audit fact element template. If a candidate term belongs to multiple fields in the field dictionary, the data processing server reads the adjacent verbs and nouns in the text segment containing the candidate term to determine its field attribution. When adjacent verbs represent expenditure, allocation, misappropriation, interception, acceptance, or payment, the relevant terms are prioritized for writing into the behavior mode or approval process field; when adjacent nouns represent funds, subsidies, special projects, or budgets, the relevant terms are prioritized for writing into the fund attribute field; when adjacent nouns represent contracts, invoices, approval forms, payment vouchers, or acceptance documents, the relevant terms are prioritized for writing into the evidence materials field.

[0032] In one embodiment, the audit suspect text is that a unit listed special subsidy funds as meeting expenses without actual business support, and the only attachments were invoices, with no approval forms or acceptance documents. The data processing server first segmented the text into segments describing the source of funds, the behavior, and the evidence. Then, it located the special subsidy funds in the fund attribute field, the expenditure behavior without actual business support in the behavior method field, and the invoices, approval forms, and acceptance documents in the evidence materials field. Among these, the approval forms and acceptance documents are only considered missing clues and are not yet considered as factual evidence materials because the text semantics indicate that they were not seen.

[0033] After completing the above processing, the data processing server outputs the audit fact element extraction sequence. The audit fact element extraction sequence refers to the set of candidate terms arranged in the order of text source identifier and preset audit fact element template fields. It serves as the input for terminology normalization and field confirmation in S102, and preserves the text paragraph positions so that the subsequently generated element sets can maintain a traceable relationship with the original text.

[0034] S102: Generate a set of factual elements for doubts and a set of factual elements for cases based on the audit terminology standardization table.

[0035] The data processing server reads the audit fact element extraction sequence output by S101 and calls the audit terminology normalization table to perform standardized replacements on synonyms or near-synonyms. The audit terminology normalization table records the correspondence between different textual expressions and standard expressions in audit operations. The replacement process is carried out with fields as boundaries; the same candidate term can only enter its corresponding field and cannot be replaced across fields.

[0036] The audit terminology normalization table uses non-standard expressions, standard expressions, and their respective fields as record units. Normalization processing is performed only within the corresponding field. The data processing server matches candidate terms with non-standard expressions; if a match is successful, the candidate term is replaced with the corresponding standard expression; if a candidate term is already equal to a standard expression, its original value is retained. When the same textual expression exists in different fields, the corresponding normalized record is retrieved according to the field to which the candidate term currently belongs.

[0037] For the evidence materials field, the data processing server performs a negative semantic determination before writing the fact value. The negative semantic determination is based on the text segment containing the evidence term. When negative terms such as "not seen," "missing," "not provided," "unobtainable," "not attached," or "not submitted" appear before or after the evidence term, the evidence term is not written into the obtained fact value of the evidence materials field, but is retained as a missing condition clue. When positive terms such as "obtained," "provided," "attached," "seen," "included," or "submitted" appear in the text segment containing the evidence term, the evidence term is written into the evidence materials field. If the same evidence term has both positive and negative descriptions, the audit conclusion segment is determined first. The audit conclusion segment is a text segment containing conclusive terms such as "audit's opinion," "problem manifestation," "handling decision," "rectification requirements," "request for recovery," or "order to rectify." If no conclusive terms exist, the last text segment in the source text containing the method of action or the consequences of handling is used as the audit conclusion segment. The data processing server calculates the text paragraph distance between the affirmative description paragraph and the negative description paragraph and the audit conclusion paragraph, respectively. The description with the smaller distance is used as the basis for writing; if the distance is the same, the negative description is used as a condition missing clue, and the evidence term is not written into the obtained fact value.

[0038] For candidate terms corresponding to audit-related suspicious texts, the data processing server confirms them according to the field order of audit item category, audit object, fund attribute, behavior method, approval process, occurrence period, evidence materials, and handling consequences. If a candidate term is consistent with a non-standard expression in the audit terminology standardization table, the candidate term is replaced with the corresponding standard expression; if multiple candidate terms appear in the same field, the factual description is retained according to the order of the text paragraphs, and terms that can be merged into the same standard expression in the audit terminology standardization table are merged into one fact value; if no affirmative factual description appears in a field, the field remains in a null value state in the suspicious factual element set, and this null value state is used to determine the missing condition in step S3.

[0039] For candidate terms corresponding to historical case texts, the data processing server performs normalization verification using the same field order and additionally retains the correspondence between case identifiers and the final processing results. The fact values ​​in the case fact element set are derived solely from the factual content already described in the historical case texts; the issue name or case title is only used to assist in determining the audit matter category and is not used as a substitute fact value for the behavior mode field. Therefore, historical cases can participate in condition coverage processing in subsequent step S2 based on fact fields, rather than simply entering the regulatory case association graph based on title similarity.

[0040] In one embodiment, a historical case document records that a project expensed project subsidies as fictitious meeting expenses. An audit obtained contracts, invoices, and payment vouchers, and requested the recovery of funds. The data processing server consolidates the project subsidies into a special fund using audit terminology, merges the fictitious meeting expense into fictitious expenditures, writes the contracts, invoices, and payment vouchers into the evidence material field, and writes the recovered funds into the processing consequences field, generating a corresponding case fact element set. This case fact element set is used in step S2 to compare with the applicable legal conditions set in the same fields, forming a case condition coverage record.

[0041] After execution, the data processing server outputs a set of suspicious fact elements and a set of case fact elements. The set of suspicious fact elements serves as the input for generating suspicious condition coverage records in step S3, while the set of case fact elements serves as the input for generating case condition coverage records and condition instance relationships in step S2. Both use the same preset audit fact element template and the same audit terminology unification table to ensure consistency in the field definitions of subsequent comparison objects.

[0042] S103: Deconstruct the text of the regulations and generate a set of applicable conditions for the regulations.

[0043] The data processing server uses the text segment sequence corresponding to the legal clause text in S101 as the starting point for processing, and breaks down the legal clause text into conditional statements according to the clause number, payment number, and conditional connectors. The conditional connectors include semantic connectors that point to the applicable subject, scope of funds, prohibited behaviors, approval procedures, applicable period, evidence requirements, and consequences of processing. The data processing server determines the conditional fields, conditional limit values, and conditional roles based on the position and semantic function of the conditional statements in the legal clauses.

[0044] Conditional connectors and conditional fields correspond semantically. Conditional statements pointing to units, departments, or responsible entities should include applicable object conditions; conditional statements pointing to fiscal funds, special funds, subsidies, or operating funds should include fund attribute conditions; conditional statements pointing to prohibitions, restrictions, violations, fictitious listings, withholding, or misappropriation should include behavioral composition conditions; conditional statements pointing to approvals, authorizations, filings, acceptance, or payment review should include procedural conditions; conditional statements pointing to periods, years, implementation dates, or validity periods should include point-in-time conditions; conditional statements pointing to contracts, invoices, approval forms, payment vouchers, or acceptance documents should include evidence requirement conditions; and conditional statements pointing to orders to rectify, fund recovery, penalties, or system improvement should include consequences conditions.

[0045] The condition fields correspond to the fields in the preset audit fact element template. Condition limit values ​​are taken from the regulatory clauses' limitations on subjects, funds, actions, procedures, periods, evidence, or consequences, and converted into standard expressions comparable to factual values ​​using an audit terminology standardization table. Condition roles distinguish between qualitatively necessary conditions, supplementary evidence conditions, and result-corresponding conditions. Applicable object conditions, fund attribute conditions, and behavioral constitutive conditions enter into qualitatively necessary conditions; procedural conditions and evidentiary requirement conditions enter into supplementary evidence conditions; and handling consequence conditions enter into result-corresponding conditions. Point-in-time conditions independently retain their applicable regulatory period and are used in step S3 to determine if there are any conflicts with point-in-time conditions.

[0046] When the same conditional statement contains both procedural requirements and evidentiary requirements, the data processing server splits it into procedural conditions and evidentiary requirement conditions according to the core terms in the conditional statement, and writes supplementary evidence conditions separately.

[0047] During the decomposition process, the data processing server first determines the audit matter category to which the regulatory clause belongs, and then establishes at least one set of applicable conditions under the same regulatory clause identifier. If the same regulatory clause has multiple parallel behaviors, the data processing server forms multiple sets of applicable conditions according to the parallel behaviors. Each set of applicable conditions shares the same regulatory clause identifier, but the condition limit values ​​of the behavior-constituting conditions are different. If the same regulatory clause only stipulates the consequences of handling without directly describing the evidence requirements, the conditions for handling consequences are still written into the corresponding conditions of the results, and the conditions for evidence requirements remain in a null state, without generating condition items without a source.

[0048] In one embodiment, the regulatory clause stipulates that special funds should be used for their approved purposes and that no fictitious expenditures, misappropriation, or embezzlement is permitted. Violations will result in an order to rectify the situation and the recovery of funds. The data processing server writes the special funds into the fund attribute conditions, maps the use for approved purposes to the procedural conditions, the prohibition of fictitious expenditures into the behavioral constitutive conditions, and the order to rectify the situation and recover funds into the handling consequence conditions. If the clause further limits its application to fund usage behaviors occurring within a certain period, the data processing server writes that period into the time point conditions. The generated set of applicable regulatory conditions serves as the comparison benchmark for the case fact element set in step S2 and as the comparison benchmark for the doubtful fact element set in step S3.

[0049] After execution, the data processing server outputs a set of applicable regulatory conditions. Each set of applicable regulatory conditions is bound to a regulatory clause identifier and an audit matter category, and uses condition fields, condition limit values, and condition roles as the smallest processing unit for subsequent comparisons of the same field. This enables step S2 to generate case condition coverage records, and step S3 to generate doubtful condition coverage records and conditions requiring supplementary verification.

[0050] S104: Extract the text of rectification measures and generate a set of rectification measure elements.

[0051] The data processing server reads the text segment sequence corresponding to the rectification measures text in S101, and uses the processing consequence conditions in the set of applicable legal conditions generated in S103 as a reference to extract the correspondence between the rectification object, rectification action, rectification basis, rectification materials, and processing consequence in the rectification measures text. The rectification measures text can come from independent rectification measures texts or rectification portions of historical case texts; when the rectification measures text comes from the rectification portion of a historical case, the rectification measures element set is simultaneously bound to the corresponding case identifier.

[0052] During extraction, the data processing server first locates action terms in the rectification measures text, writing actions such as recovery, accounting adjustment, system improvement, formalities re-processing, and transfer for handling into the rectification actions. Then, it locates the recipients of these actions, writing funds, accounts, approval documents, system processes, or responsible entities into the rectification objects. Subsequently, based on the legal provisions cited, consequences described, or case results in the rectification measures text, it establishes a correspondence between the rectification basis and the consequences conditions. If the rectification measures text does not directly specify legal provisions cited, but its rectification actions and rectification objects correspond to the same fields in S103, the data processing server establishes a consequences correspondence between the rectification measures element set and the corresponding consequences conditions.

[0053] The process of extracting corrective measures first identifies the corrective actions, then the corrective targets, and finally the basis for corrective action. Corrective actions originate from verbal terms in the corrective measure text; corrective targets originate from noun phrases following the corrective actions; and the basis for corrective action originates from legal citations or statements of consequences appearing in the corrective measure text. If the corrective measure text does not directly cite legal citations, the data processing server matches the corrective actions and corrective targets against the consequences, establishing a corresponding relationship between consequences upon successful matching.

[0054] In one embodiment, the rectification measures text instructs the audited entity to recover the illegally disbursed special funds, supplement approval documents, and improve the project fund payment review system. The data processing server writes the audited entity as the rectification target, includes the recovery, supplementation of approval documents, and improvement of systems as rectification actions, establishes referential relationships between the special funds and approval documents and the fund attribute conditions and procedural conditions, and establishes a correspondence between the recovered illegally disbursed special funds and the recovery-related processing consequences conditions in the set of applicable legal conditions. This set of rectification measures elements is subsequently written with condition instance relationships in step S2 and serves as the source of candidate rectification measures in step S4.

[0055] After execution, the data processing server synchronizes the audit matter categories of the suspicious fact element set, the legal application condition set, the case fact element set, and the rectification measure element set. The synchronization process uses the audit matter category as the group boundary, marking the legal application condition set, case fact element set, and rectification measure element set under the same audit matter category as objects eligible for step S2 condition coverage processing, while retaining the suspicious fact element set as the input object for step S3. The synchronization result does not change the field values ​​of each element set; it only writes the audit matter category consistency flag required for subsequent calls.

[0056] In step S1, the data processing server transforms multi-source audit text into a structured set of elements, forming a set of suspicious factual elements, a set of applicable legal conditions, a set of case factual elements, and a set of corrective measures elements. It also writes subsequent retrieval markers such as legal clause identifiers, case identifiers, audit matter categories, and corresponding handling consequences into each element set. These results enable step S2 to generate case condition coverage records and condition instance relationships based on the set of applicable legal conditions, and allow step S3 to directly call the set of suspicious factual elements to generate suspicious condition coverage records. Furthermore, it provides a unified field specification for step S4 to screen candidate cases and candidate corrective measures.

[0057] Step S1 has generated a set of factual elements related to the doubts, a set of applicable legal conditions, a set of factual elements related to the cases, and a set of elements related to corrective measures. These elements have been incorporated with the corresponding legal clause identifiers, case identifiers, audit matter categories, and consequences. Step S2 continues processing based on the structured results. However, whether case facts can serve as factual instances of legal clauses still needs to be determined item by item according to the applicable legal conditions. This step is responsible for tasks such as covering conditions in the same field, generating case condition coverage records, matching corrective measures, and constructing a legal-case relationship graph.

[0058] S201: Organize the sets of applicable legal conditions and case fact elements by audit matter category.

[0059] In this embodiment, step S2 is executed by the data processing server. The data processing server takes the set of applicable regulatory conditions, the set of case fact elements, and the set of rectification measures elements output in step S1 as the starting point for processing. It first reads the correspondence between the audit matter category, regulatory clause identifier, case identifier, and processing consequences carried by each element set, and forms grouped objects for condition coverage processing.

[0060] The data processing server uses audit matter categories as group boundaries, placing the set of applicable regulatory conditions and the set of case fact elements into the same comparison group. Within each comparison group, the data processing server reads the set of applicable regulatory conditions one by one according to the regulatory clause identifier, and uses the set of case fact elements under the same audit matter category as the comparison object. If a regulatory clause corresponds to multiple sets of applicable regulatory conditions, they are compared one by one in the order of the sets of applicable regulatory conditions formed in step S1, and each set of applicable regulatory conditions corresponds to one independent condition coverage process.

[0061] After the comparison group is formed, the data processing server establishes a condition coverage input sequence. A condition coverage input sequence is an ordered processing unit consisting of a set of applicable regulatory conditions and a set of case fact elements. Within this unit, the condition fields and condition limit values ​​from the set of applicable regulatory conditions serve as the comparison benchmark, while the fact values ​​of the same fields from the set of case fact elements are used as the comparison values. Only after the condition coverage input sequence is formed does the subsequent coverage status determination proceed, ensuring that each determination is limited to the same audit matter category and the same field definition.

[0062] In one embodiment, step S1 categorizes regulatory clause A into the special fund audit item category and normalizes project subsidies, fictitious expenditures, missing approval documents, and recovered funds from historical case B into a case fact element set. The data processing server places the regulatory application condition set of regulatory clause A and the case fact element set of historical case B into the same comparison group, and sequentially adds the fund attribute conditions, behavioral composition conditions, procedural conditions, evidence requirement conditions, and handling consequence conditions to the condition coverage input sequence.

[0063] S202: Perform same-field condition coverage processing and generate case condition coverage records.

[0064] The data processing server reads the condition coverage input sequence formed by S201 and performs a same-field condition coverage determination on each processing unit. The same-field condition coverage determination refers to, under the same condition field, comparing the condition restriction values ​​in the set of applicable legal conditions with the fact values ​​in the set of case fact elements for same value, inclusion, period fall-in, and normalized synonym merging, and outputting the coverage status.

[0065] Condition coverage processing is executed in the following order: same field value, coverage judgment, and conflict judgment. The data processing server first reads the condition fields and condition restriction values ​​in the set of applicable legal conditions, and then searches for fact values ​​in the same field in the set of fact elements. If no fact value in the same field is found, the coverage status is written as condition missing. If a fact value in the same field is found, the same value judgment, inclusion judgment, period inclusion judgment, and normalized synonym merging judgment are executed in sequence. If any judgment is true, the condition is written as covered. If none of the above judgments are true, the condition conflict judgment is executed. If the fact value and the condition restriction value are mutually exclusive under the same field, it is written as condition conflict; otherwise, it is written as condition missing.

[0066] In practice, the data processing server first reads the condition fields and then searches for fact values ​​in the same field within the case fact element set. If no fact value is found, the coverage status of the corresponding condition is set to "condition missing." If a fact value is found, the process follows the order of same-value judgment, inclusion judgment, period inclusion judgment, and normalized synonym merging judgment. When the fact value is exactly the same as the condition limit value, the coverage status is set to "condition covered." When the fact value falls within the range defined by the condition limit value, the coverage status is set to "condition covered." When the fact value in the occurrence period field falls within the period defined by the point-in-time condition, the coverage status is set to "condition covered." When the fact value, after being processed by the audit terminology normalization table, belongs to the same standard expression as the condition limit value, the coverage status is set to "condition covered."

[0067] The same value judgment is performed when the fact value and the conditional constraint value are completely identical at the standard expression level. Inclusion judgments are performed based on the standard expression hierarchy in the audit terminology standardization table. The standard expression hierarchy records the inclusion relationship between higher-level and lower-level expressions under the same condition field. For condition fields... The fact value is recorded as The condition limit value is denoted as The condition constraint value is in the condition field. The set of corresponding subordinate representations is denoted as .when When true, the covering state is written as the condition being covered; if If there is no set of subordinate expressions, then Only contains itself.

[0068] For the "Period of Occurrence" field, the data processing server converts the "Period of Occurrence" into a "Start Date" and "End Date," and converts the "Point-in-Time Condition" into a "Start Date" and "End Date." The period inclusion / exclusion criteria are performed according to the following relationship: ; in, The "Date of Occurrence" field is derived from the set of factual elements of doubt or the set of factual elements of the case. The conditions are defined within a specific period and are derived from the set of applicable legal conditions at a specific point in time. The date on which the event began; The end date of the period during which the event occurred; The start date of the conditional period; The end date of the conditional period. When true, the coverage status of the field during the occurrence period is written as the condition is covered; when or When the occurrence period is missing a start date or end date and a comparison cannot be completed, the occurrence period field's coverage status is written as condition conflict; when the occurrence period is missing a start date or end date and a comparison cannot be completed, the occurrence period field's coverage status is written as condition missing.

[0069] Normalized synonym merge checks are performed within the same condition field, and are evaluated according to the following relationships: ; in, For fact value In the condition field Standardized expressions obtained from the Unified Table of Internal Audit Terminology; Conditional limit value In the condition field The standard expression obtained from the Internal Audit Terminology Standardization Table. If the above equation holds true, the coverage status is written as "condition covered".

[0070] Condition conflict determination is performed based on the field exclusion table. The field exclusion table uses condition fields, primary standard statements, and secondary standard statements as record units, and is used to record statements that cannot be true simultaneously under the same field. Condition fields The conflict function is denoted as follows: When the field excludes the relationship table or When recording, ,otherwise The conditional coverage state function is determined according to the following relationship: ; in, For condition fields Coverage status; This indicates that the fact value corresponding to this condition field does not exist in the fact element set. The above coverage status only takes values ​​between condition being covered, condition missing, and condition conflict.

[0071] After the data processing server completes the field determination for one set of applicable legal conditions and one set of case factual elements, it generates a case condition coverage record. The case condition coverage record records the legal clause identifier, case identifier, condition field, condition limit value, corresponding factual value, and coverage status item by item, and maintains a binding relationship with the corresponding set of applicable legal conditions. In the special fund example, project subsidy funds in the case factual element set are consolidated into special funds through an audit terminology unification table, and the conditions for coverage are formed by the fund attribute conditions; fictitious expenditures and behavioral constitutive conditions are covered; missing approval documents correspond to missing procedural conditions; and recovered funds and consequences conditions are covered. The above determination results are written into the same case condition coverage record.

[0072] S203: Establish condition instance relations based on the results of qualitative necessary condition conflicts.

[0073] The data processing server uses the case condition coverage record generated in S202 as the starting point for processing, and reads the coverage status of the qualitative necessary conditions. The qualitative necessary conditions adopt the condition roles of the set of applicable legal conditions in step S1, including applicable object conditions, financial attribute conditions, and behavioral composition conditions; the time point conditions participate in the determination as conflict verification conditions.

[0074] The data processing server aggregates case condition coverage records under the same regulatory clause identifier and the same case identifier, checks for conflicts in qualitative necessary conditions, and checks for conflicts in point-in-time conditions. When there are no conflicts in the applicable object conditions, funding attribute conditions, and behavioral constitutive conditions, and there are no conflicts in the point-in-time conditions, the data processing server writes the corresponding regulatory clause, case fact element set, and case condition coverage record into the condition instance relationship. If there are missing qualitative necessary conditions but no conflicts, the data processing server writes the case fact element set as an incomplete coverage instance into the condition instance relationship, and retains the corresponding missing condition status in the condition instance relationship. In step S4, the incomplete coverage instance can only serve as an auxiliary source of explanation for candidate cases and cannot replace the case fact element set with complete coverage of qualitative necessary conditions as a source of qualitative basis.

[0075] When establishing conditional instance relationships, the data processing server organizes the connections in the following order: regulatory clauses, sets of applicable regulatory conditions, sets of case factual elements, and case condition coverage records. Regulatory clauses serve as the source node, sets of applicable regulatory conditions serve as the benchmark node, sets of case factual elements serve as instance nodes, and case condition coverage records are written as a connection attribute between sets of applicable regulatory conditions and sets of case factual elements. This connection attribute preserves the coverage status of each condition field, enabling subsequent step S4 to filter candidate cases based on the field correspondence between the suspected condition coverage records and the case condition coverage records.

[0076] In the special fund implementation example, historical case B had conditions covered in both the fund attribute conditions and the behavioral constitutive conditions. The occurrence period fell within the period limited by legal clause A. Although the approval materials were missing conditions, they were conditions that could be supplemented. The data processing server wrote legal clause A, the set of applicable legal conditions, the set of case fact elements of historical case B, and the corresponding case condition coverage record into the condition instance relationship, and retained the missing procedural conditions corresponding to the approval materials in the connection attribute.

[0077] S204: Match the elements of rectification measures and generate a legal case association map.

[0078] After the data processing server establishes the conditional instance relationship in S203, it continues to read the set of rectification measures elements output in step S1, and uses the processing consequence conditions in the set of applicable regulatory conditions as the matching benchmark. The rectification measure matching process is executed in the order of processing consequence conditions, rectification actions, rectification objects, and rectification basis. First, it compares the correspondence between the processing consequence conditions and the processing consequences in the set of rectification measures elements. Then, it compares whether the rectification actions or rectification objects are consistent with the processing results pointed to by the processing consequence conditions. Finally, it compares whether the rectification basis points to the same regulatory clause identifier.

[0079] When the set of rectification measures elements matches the conditions for handling consequences, the data processing server writes the set of rectification measures elements as a measure node into the condition instance relationship formed in S203. When the set of rectification measures elements originates from the rectification portion of a historical case, the measure node also retains the case identifier, enabling the measure node to return to the corresponding case fact element set. When the set of rectification measures elements originates from an independent rectification measure text, the measure node is connected to the corresponding legal clause through the corresponding relationship for handling consequences. After matching is completed, the condition instance relationship consists of the legal clause, the set of applicable legal conditions, the set of case fact elements, the case condition coverage record, and the set of rectification measures elements.

[0080] The data processing server then generates a legal case association graph. The graph uses legal clause nodes, legal application condition nodes, case fact nodes, and rectification measure nodes as nodes, and condition instance relationships as edges. Edge attributes include case condition coverage records and the correspondence between processing consequences. Legal application condition nodes are located between legal clause nodes and case fact nodes, and rectification measure nodes are connected to the processing results in legal clause nodes or case fact nodes through processing consequence conditions. After the graph is generated, step S3 inputs the set of questionable factual elements into the legal case association graph and invokes the same condition coverage rule; step S4 extracts candidate cases and candidate rectification measures along the condition instance relationships.

[0081] Each condition instance relationship connects at least one legal clause node, one legal application condition node, and one case fact node; when a matching set of rectification measures elements exists, the corresponding rectification measure node is then connected.

[0082] In the special fund implementation example, the set of rectification measures includes recovering illegally spent funds and supplementing approval documents. The processing consequence conditions of Article A of the regulations include the recovery of funds. The data processing server matches the recovery of illegally spent funds with the processing consequence conditions of fund recovery, and writes the set of rectification measures into the condition instance relationship corresponding to Article A of the regulations and historical case B, ultimately forming a graph path from Article A of the regulations to the set of applicable conditions of the regulations, historical case B, case condition coverage records, and the set of rectification measures.

[0083] In step S2, the data processing server completes the condition coverage processing between the set of applicable regulatory conditions and the set of case fact elements under the same audit matter category, generates case condition coverage records, and writes the set of case fact elements that do not conflict with the necessary qualitative conditions into the condition instance relationship. The regulatory case association graph consists of regulatory clause nodes, regulatory application condition nodes, case fact nodes, condition instance relationships, and rectification measure nodes. It can be used in step S3 for doubtful condition coverage processing and in step S4 for screening candidate cases, candidate rectification measures, and candidate reasoning paths.

[0084] Step S1 has already formed the set of factual elements of the doubt and the set of applicable legal conditions. Step S2 has generated a legal case relationship graph using the set of applicable legal conditions as an intermediate comparison benchmark, and retained the condition instance relationships and case condition coverage records in the graph. However, whether the current audit doubt can be included in the legal case comparison requires first determining the coverage status of the doubt facts with the applicable legal conditions. Step S3, on top of the legal case relationship graph, performs same-field condition coverage processing on the set of factual elements of the doubt and the set of applicable legal conditions, generating doubt condition coverage records, and determining candidate legal clauses and conditions requiring supplementary evidence based on these records.

[0085] S301: Input the set of suspicious factual elements into the legal case association graph and construct a suspicious coverage input sequence.

[0086] The following steps are all performed by the data processing server. The data processing server uses the set of suspicious fact elements output in step S1 and the regulatory case association graph obtained in step S2 as the starting point. It first reads the audit matter categories from the set of suspicious fact elements, and then searches the regulatory case association graph for regulatory clause nodes and regulatory application condition nodes that match the audit matter categories. The audit matter category serves as an entry condition, limiting the comparison of suspicious facts to a set of regulatory application conditions only within the same business area.

[0087] After completing the node lookup, the data processing server reads the sets of applicable conditions for each regulation according to the regulatory clause identifier, and combines the set of questionable factual elements with each set of applicable conditions to form a questionable coverage input sequence. A questionable coverage input sequence is an ordered processing unit consisting of one set of questionable factual elements and one set of applicable conditions. In this unit, the condition fields and condition limit values ​​in the set of applicable conditions serve as the comparison benchmark, and the factual values ​​of the same fields in the set of questionable factual elements serve as the comparison values. If the same regulatory clause corresponds to multiple sets of applicable conditions, each set of applicable conditions forms a separate questionable coverage input sequence, retaining the same regulatory clause identifier.

[0088] When constructing the input sequence for suspicious point coverage, the data processing server synchronously reads the condition instance relationships connected to the regulatory application condition nodes in step S2. However, these condition instance relationships are only saved as reachable path markers for subsequent candidate case screening in this sub-step and do not participate in the suspicious point coverage status determination. Therefore, the set of suspicious point factual elements first establishes a direct comparison relationship with the set of regulatory application conditions. After the candidate regulatory clauses are determined, step S4 searches for candidate cases and candidate rectification measures along the condition instance relationships.

[0089] In one embodiment, the suspicious fact element set records that an entity listed special funds as meeting expenses without actual business support during 2023, and the only evidence materials are invoices and payment vouchers. The data processing server reads the audit matter category of this suspicious fact element set as special funds, and locates the regulatory clause nodes belonging to special funds in the regulatory case association graph, forming a suspicious fact element set with the corresponding regulatory application condition set to form a suspicious coverage input sequence.

[0090] S302: Apply condition coverage processing to the set of factual elements of doubt and the set of applicable legal conditions.

[0091] The data processing server uses the suspicious point coverage input sequence formed by S301 as the processing object and performs condition coverage processing according to the condition fields in the set of applicable legal conditions. For each condition field processed, the data processing server first searches for the fact value of the same field in the set of suspicious point fact elements, then compares the found fact value with the condition limit value, and outputs the coverage status.

[0092] The comparison process follows the condition coverage rules of step S2. If the fact value corresponding to the condition field does not exist in the set of fact elements of suspicion, the coverage status is written as "condition missing"; if the same fact value exists, the same value judgment, inclusion judgment, period inclusion judgment, and normalized synonym merging judgment are executed in sequence. When the fact value is completely consistent with the condition limit value, the coverage status is written as "condition covered"; when the fact value is within the range of the condition limit value, the coverage status is written as "condition covered"; when the fact value of the occurrence period field falls within the period limited by the time point condition, the coverage status is written as "condition covered"; when the fact value, after being processed by the audit terminology normalization table, belongs to the same standard expression as the condition limit value, the coverage status is written as "condition covered".

[0093] The condition coverage processing in step S3 follows the rules in S202 regarding same field value, coverage judgment, conflict judgment, inclusion judgment, period fall-in judgment and normalized synonym merging judgment. Under this rule, the suspicious fact element set only replaces the case fact element set as the source of fact value, and the coverage status is still limited to condition covered, condition missing and condition conflict.

[0094] When suspicious factual elements contain factual values ​​in the same field but do not form a covering relationship, the data processing server continues to determine whether the factual values ​​and conditional limits are mutually exclusive. If the audit object type is inconsistent with the applicable object conditions, the fund attribute is inconsistent with the fund attribute conditions, the behavior mode is contrary to the behavior constituting conditions, or the occurrence period exceeds the time limit specified by the time point conditions, the coverage status is written as condition conflict; if the factual values ​​in the same field do not meet the conditional limits but do not form a covering relationship, the coverage status is written as condition missing. The coverage status only takes values ​​between condition covered, condition missing, and condition conflict.

[0095] In the special fund implementation example, the set of applicable regulatory conditions requires that the fund attribute be special fund, the behavioral condition be fictitious expenditure, the procedural condition be used for the approved purpose, the evidentiary condition be approval documents and acceptance materials, and the time condition be within the applicable period of the regulations. The data processing server compares the special fund in the set of suspicious factual elements with the fund attribute conditions and finds that the conditions are covered; it merges the meeting expenses without actual business support into fictitious expenditures through the audit terminology standardization table and compares them with the behavioral condition, finding that the conditions are covered; since no affirmative factual value of approval documents and acceptance materials appears in the set of suspicious factual elements, the procedural condition and the evidentiary condition are written as missing; if 2023 falls within the applicable period of the regulations, the time condition is written as covered.

[0096] S303: Generate a record of suspicious condition coverage and identify candidate regulatory clauses After the data processing server completes the coverage status determination of all condition fields in a single suspected coverage input sequence in step S302, it generates a suspected condition coverage record. This record includes the regulatory clause identifier, condition fields, condition limit values, suspected fact values, and coverage status, and is bound to the corresponding set of applicable regulatory conditions. The suspected fact value originates from the fact value of the same field in the suspected fact element set. If the fact value of the same field does not exist, the suspected fact value remains null, and the coverage status is set to "condition missing."

[0097] After generating the suspected condition coverage record, the data processing server aggregates the coverage status under the same set of applicable conditions according to the legal clause identifier, and reads the applicable object condition, fund attribute condition, behavioral constitutive condition, and time point condition. The determination of candidate legal clauses uses the set of applicable conditions as the smallest unit: when the applicable object condition, fund attribute condition, and behavioral constitutive condition are all marked as covered, and the time point condition is not marked as conflicting, the legal clause is determined as a candidate legal clause. If the time point condition does not exist in the set of applicable conditions, then the legal clause does not constitute a conflict determination regarding the time point condition.

[0098] If the set of applicable conditions for regulations does not specify applicable object conditions, then the application object conditions are not required to be covered when determining candidate regulatory clauses; if the set of applicable conditions for regulations specifies applicable object conditions, then those applicable object conditions must be marked as covered. Fund attribute conditions and behavioral constitutive conditions are mandatory conditions for candidate regulatory clauses. If any condition is marked as missing or conflicting, the corresponding set of applicable conditions for regulations will not be considered as a basis for selecting candidate regulatory clauses. For regulatory clauses to be determined... Its corresponding first The set of applicable conditions for each regulation is denoted as follows: The set of candidate regulatory provisions is denoted as The candidate regulatory provisions are determined according to the following relationship: ; in, For legal provisions The corresponding set of applicable legal conditions; For the first The determination of the coverage of applicable conditions in a set of applicable conditions for a regulation is made if there are applicable conditions in that set of applicable conditions. This indicates that the condition is marked as overridden; if no applicable object condition exists, then... The default setting is true; For the determination of coverage of fund attribute conditions, it means that the fund attribute conditions are marked as covered. The determination of whether a behavior constitutes a condition being covered indicates that the condition for the behavior has been marked as covered. This indicates that a time-point condition is marked as a conflict. This indicates that there are no conflicting conditions at any given point in time. When the above relationship is satisfied, the legal provisions... Write into the candidate regulatory clause set And retain the identifier of the set of applicable legal conditions that satisfy the relationship.

[0099] During the determination of candidate regulatory clauses, if procedural conditions and evidentiary requirements are marked as missing, the determination result of the candidate regulatory clause remains unchanged. The data processing server retains this missing status in the questionable condition coverage record and continues to read the next set of applicable regulatory conditions. If the same regulatory clause corresponds to multiple sets of applicable regulatory conditions, as long as one of the sets of applicable regulatory conditions meets the criteria for determining the candidate regulatory clause, the regulatory clause is written into the candidate regulatory clause, and the identifier of the set of applicable regulatory conditions that meets the criteria is retained for step S4 to search for candidate cases along the corresponding condition instance relationship.

[0100] In the special fund implementation example, the set of factual elements of doubt covers both special funds and fictitious expenditures, the occurrence period of which does not conflict with the applicable period of regulations, but approval documents and acceptance data are missing. Based on the doubt condition coverage record, the data processing server determines the corresponding regulatory clause as a candidate regulatory clause and retains the missing status of the approval documents and acceptance data in the doubt condition coverage record of that candidate regulatory clause.

[0101] S304: Extract the conditions to be supplemented from the candidate regulatory clauses and form the results of subsequent invocation.

[0102] The data processing server takes the candidate legal clauses and doubtful condition coverage records determined in S303 as the starting point for processing. It reads all condition fields belonging to the candidate legal clauses according to the legal clause identifiers and filters out applicable legal conditions whose coverage status is missing and whose condition roles belong to supplementary evidence conditions. Supplementary evidence conditions use the condition roles formed in step S1, specifically corresponding to procedural conditions and evidentiary requirement conditions.

[0103] The extraction order of conditions to be supplemented is as follows: first, the legal clause identifier is limited to candidate legal clauses; second, the coverage status is limited to condition missing; and finally, the condition field is limited to procedural conditions or evidentiary requirements. Conditions conforming to this order are written into the conditions to be supplemented, and the condition field, condition limit value, and legal clause identifier are retained. These conditions to be supplemented represent procedural or evidentiary facts in the current set of questionable factual elements that have not yet been obtained or described, but do not affect the preliminary entry of candidate legal clauses into subsequent reasoning.

[0104] The conditions for supplementary certificates are extracted according to the following relationship: ; in, Candidate regulatory provisions The corresponding set of conditions to be supplemented; Candidate regulatory provisions The corresponding suspicious condition coverage record contains the condition items; Candidate regulatory provisions The corresponding set of records covering suspicious conditions; For conditional items Coverage status in the suspicious condition coverage record; For conditional items Conditional role within the set of applicable legal conditions. Only when and When both conditions are met, the condition item Write the set of conditions to be supplemented .

[0105] If a procedural condition or evidentiary requirement condition in the doubtful condition coverage record is covered, the data processing server will not write it into the supplementary evidence condition. If a procedural condition or evidentiary requirement condition is conflicting, although the corresponding legal provision may have entered the candidate legal provision judgment process, the conflict status is still retained in the doubtful condition coverage record for further reading in step S4 when judging candidate reasoning paths. The supplementary evidence condition only records missing conditions and does not replace the coverage status in the doubtful condition coverage record.

[0106] When a procedural condition or evidentiary requirement is marked as a condition conflict, it is not written into the supplementary evidence conditions; this conflict state is entered into step S4 as an exclusionary marker in the doubt condition coverage record. In step S4, when forming candidate reasoning paths, if the conflict state directly affects the correspondence of the processing consequence conditions or the case comparability, the corresponding candidate reasoning path is not retained.

[0107] In the special fund implementation example, candidate regulatory provisions require the existence of approval documents and acceptance data, while the factual elements of doubt only record invoices and payment vouchers. The data processing server extracts the procedural conditions corresponding to the approval documents and the evidentiary requirements corresponding to the acceptance data from the doubtful condition coverage records of the candidate regulatory provisions, and writes them into the conditions to be supplemented. In step S4, the conditions to be supplemented are converted into evidence strengthening items and enter the candidate reasoning path screening together with candidate cases and candidate rectification measures.

[0108] In step S3, the data processing server completes the condition coverage processing between the set of doubtful factual elements and the set of applicable legal conditions in the legal case association graph, generating a doubtful condition coverage record. Based on the applicable object conditions, funding attribute conditions, behavioral composition conditions, and timing conditions, candidate legal clauses are determined. Simultaneously, procedural conditions and evidentiary requirement conditions with missing coverage status are extracted from the candidate legal clauses, forming conditions requiring supplementary evidence. The doubtful condition coverage record, candidate legal clauses, and conditions requiring supplementary evidence are output to step S4 for screening candidate cases, candidate rectification measures, and forming candidate reasoning paths.

[0109] Step S1 has generated a set of factual elements for suspected issues, a set of factual elements for cases, and a set of elements for corrective measures. Step S2 has obtained a regulatory case relationship graph. Step S3 has output records of suspected condition coverage, candidate regulatory clauses, and conditions requiring supplementary evidence. However, after regulatory clauses enter the candidate scope, it is still necessary to further determine the comparability of historical cases and the correspondence of corrective measures in the condition instance relationships. Step S4, based on the above results, filters candidate cases and candidate corrective measures, forms candidate reasoning paths, and generates audit qualitative reasoning results.

[0110] S401: Read the condition instance relationship along the candidate regulatory clauses.

[0111] In this embodiment, step S4 is executed by the data processing server. The data processing server takes the candidate legal clauses, doubtful condition coverage records, and conditions requiring supplementary certification output in step S3 as the starting point for processing, and reads the legal case association graph generated in step S2. In the legal case association graph, it searches for the corresponding legal clause node according to the legal clause identifier of the candidate legal clause.

[0112] After the data processing server locates the legal clause node corresponding to the candidate legal clause, it continues to read the condition instance relationship along the legal application condition node connected to that legal clause node. The reading order is: legal clause node, legal application condition node, condition instance relationship, case fact node, and rectification measure node. Among them, the case condition coverage record in the condition instance relationship serves as the comparison basis for candidate case screening, and the rectification measure element set bound to the rectification measure node serves as the source for candidate rectification measure screening. This reading process is only performed within the scope of candidate legal clauses, so that subsequent screening does not re-traverse all legal clauses.

[0113] The data processing server will assemble each set of candidate regulatory clauses, applicable regulatory conditions, condition instance relationships, case factual element sets, case condition coverage records, and rectification measure element sets into a path unit to be filtered. This path unit is a newly formed intermediate object in this step, representing a computable data unit that starts from one candidate regulatory clause, proceeds through one condition instance relationship to one case factual element set and corresponding rectification measure element set. Only after the path unit to be filtered is formed will the case field compatibility determination begin.

[0114] In the special fund implementation example, step S3 has identified a certain special fund fictitious expenditure clause as a candidate regulatory clause and generated a corresponding suspicious condition coverage record for the clause. The data processing server reads the condition instance relationship connected to the clause in the regulatory case association graph, obtains the case fact element set, case condition coverage record, and rectification measure element set connected to the conditions of the consequences of recovering funds for historical case B, thereby forming one path unit to be screened.

[0115] S402: Filter candidate cases based on field correspondence.

[0116] The data processing server uses the path unit to be filtered formed by S401 as the processing object, reads the suspicious condition coverage record and the case condition coverage record, and establishes a field correspondence relationship according to the condition fields under the same applicable legal conditions. The field correspondence relationship means that the condition fields in the suspicious condition coverage record are aligned item by item with the same condition fields in the case condition coverage record, which is used to determine whether the suspicious facts and historical cases cover the same applicable legal conditions.

[0117] Field correspondences are established only under the same set of applicable legal conditions, and not across legal clauses or audit matter categories.

[0118] After the field correspondence is established, the data processing server first checks whether there are any missing or conflicting conditions in the qualitative necessary conditions for the candidate legal provisions in the set of suspicious factual elements. The objects of the check are the applicable object conditions, fund attribute conditions, and behavioral constituent conditions in the suspicious condition coverage records; if all of the above conditions are covered, and there are no condition conflicts in the point-in-time conditions corresponding to the occurrence period, then the case compatibility determination is initiated. The case compatibility determination is performed in the order of audit object, fund attribute, behavioral mode, and occurrence period, comparing the coverage status and corresponding fact values ​​of the suspicious condition coverage records and the case condition coverage records item by item.

[0119] In determining case compatibility, if a candidate case and the set of factual elements of doubt both cover the same applicable legal conditions under the same field, or if the factual value of one party, after being merged into the audit terminology standardization table, belongs to the same standard expression as the other party, then that field is determined to be compatible; if the occurrence period falls within the same time point condition limitation period, then the occurrence period field is determined to be compatible; if any field of audit object, fund attribute, behavior mode, or occurrence period is marked as condition conflict in the case condition coverage record, or is mutually exclusive with the corresponding factual value in the doubt condition coverage record, then the path unit to be screened will not be retained as a candidate case path.

[0120] Candidate cases are selected based on field compatibility. For the fields of audit object, fund attribute, behavior mode, and occurrence period, the data processing server reads the coverage status from the suspicious condition coverage record and the case condition coverage record, respectively. If any field is in conflict in the case condition coverage record, or if the fact value corresponding to the field is mutually exclusive with the suspicious fact value according to the field exclusion relationship table, the corresponding path unit to be screened is not retained. If there is no conflict in the above fields, and the case fact element set covers the same applicable legal conditions in the candidate legal provisions, the corresponding case fact element set is written into the candidate case.

[0121] Candidate case compatibility is determined according to the following relationship: ; in, Set of factual elements of doubt With candidate cases The compatibility determination results between them; This is a set of key fields, including the audit object, fund attributes, behavior pattern, and period of occurrence; Set of factual elements of doubt In the field The following are the questionable factual values; Candidate cases The corresponding case fact element set in the field The following factual value; The field conflict function defined for S202. When When the case condition coverage record corresponding to the candidate case indicates that it covers the same applicable conditions of the candidate regulation, the data processing server writes the case fact element set into the candidate case; when When this happens, the corresponding path unit to be filtered is not retained.

[0122] When there are no conflicting conditions in the fields of audit object, fund attribute, behavior mode, and occurrence period for the path unit to be screened, and the case condition coverage record indicates that the case fact element set covers the same applicable conditions of the candidate regulations, the data processing server determines the case fact element set as a candidate case and retains the corresponding case condition coverage record as the description attribute of the candidate case. In the special fund example, the fund attribute of historical case B is special fund, the behavior mode is fictitious expenditure, the occurrence period falls within the same applicable period of the regulations, and there are no conflicting conditions in the key fields of the case condition coverage record, the data processing server determines historical case B as a candidate case.

[0123] S403: Screen candidate rectification measures and form candidate reasoning paths.

[0124] The data processing server takes the candidate cases identified in S402 and their respective filterable path units as the starting point, reads the handling consequence conditions from the candidate legal clauses, and reads the set of rectification measures elements from the filterable path units. The selection of candidate rectification measures is based on the handling consequence conditions as the matching benchmark, and is executed in the order of handling consequence correspondence, rectification action, rectification object, and rectification basis.

[0125] The selection of candidate rectification measures is based primarily on the matching criteria of the consequences conditions. The data processing server first reads the correspondence between the consequences and the rectification measures in the element set. If the correspondence directly points to the consequences conditions in the candidate legal clauses, the element set of rectification measures is written into the candidate rectification measures. If no direct correspondence exists, the data processing server compares the rectification action and the rectification object: when the rectification action and the result in the consequences conditions belong to the same standard expression, and the rectification object is consistent with the object pointed to by the consequences conditions, the element set of rectification measures is written into the candidate rectification measures; when the rectification basis points to the same legal clause identifier, it is used as the matching confirmation basis. The matching result between the candidate rectification measures and the consequences conditions is determined according to the following relationship: ; in, For the collection of elements of rectification measures Conditions for handling consequences The matching results; Derived from the set of elements of rectification measures; The consequences of handling are derived from the set of applicable conditions of the candidate legal provisions. To handle the direct connection between consequences and judgments, if the set of corrective measures elements... The processing consequence correspondence has been linked to the processing consequence conditions. ,but Established; For the collection of elements of rectification measures The rectification actions in the process; Conditions for handling consequences The processing results are described in the text; Map the audit terms within the rectification action field to a unified table; For the collection of elements of rectification measures The targets for rectification; Conditions for handling consequences The object it refers to is derived from the limitations on funds, responsible parties, or objects of handling in the conditions for handling the consequences. This maps audit terms within the rectification target fields to a unified table. If... Then the set of rectification measures elements Included in the candidate rectification measures.

[0126] In specific processing, the data processing server first determines whether the set of rectification measures elements has been connected to the handling consequence conditions of the candidate regulatory clauses through the corresponding relationship of handling consequences. If it has been connected, the set of rectification measures elements is written into the candidate rectification measures. If the corresponding relationship of handling consequences does not directly point to the candidate regulatory clauses, the data processing server continues to compare whether the rectification action and the handling consequence conditions point to the same handling result, and compares whether the rectification object corresponds to the fund attributes and audit objects in the set of suspicious fact elements or the set of case fact elements. When the rectification basis points to the same regulatory clause identifier, or the rectification action and the rectification object can form a corresponding relationship with the handling consequence conditions, the set of rectification measures elements is determined as a candidate rectification measure.

[0127] After screening both candidate cases and candidate rectification measures, the data processing server constructs candidate reasoning paths. These paths are connected in the order of the set of factual elements of doubt, candidate legal clauses, conditional instance relationships, candidate cases, and candidate rectification measures. The path attributes include doubt condition coverage records, case condition coverage records, and conditions requiring supplementary evidence. A candidate reasoning path is represented as follows: ; in, Candidate reasoning paths; A set of factual elements indicating doubt; Candidate regulatory provisions; Candidate regulatory provisions Corresponding conditional instance relationships; Candidate cases; Candidate rectification measures. Candidate reasoning path. The conditions for its validity are: a set of factual elements of doubt. Covering candidate regulatory provisions The same applicable legal conditions, candidate cases Set of Factual Elements of Doubt There are no conflicts in the fields of audit object, fund attribute, behavior mode, and period of occurrence, and the candidate rectification measures are... With candidate regulatory provisions The conditions for handling the consequences correspond to those conditions.

[0128] In the special fund implementation example, the processing consequence condition of the candidate regulatory clause is to order rectification and recover funds. The rectification measure element set of candidate case B includes the recovery of the illegally listed special funds and supplementary approval documents. The data processing server matches the recovery of the illegally listed special funds with the processing consequence condition of fund recovery, and determines the rectification measure element set as a candidate rectification measure. Among them, the supplementary approval documents are retained as a rectification action in the rectification measure element set, and are presented together with the missing approval documents in the condition of pending supplementary certification when generating rectification suggestions, but are not used as an independent screening basis for the candidate rectification measure entry path. The candidate reasoning path is formed by the suspicious fact element set, the candidate regulatory clause, historical case B, and the rectification measure element set.

[0129] S404: Generate audit qualitative reasoning results based on candidate reasoning paths.

[0130] The data processing server uses the candidate reasoning path formed by S403 as the starting point for processing. It reads candidate regulatory clauses, records of suspicious conditions coverage, candidate cases, conditions requiring supplementary evidence, and candidate rectification measures from the path, and generates audit qualitative reasoning results according to the structured result fields. The audit qualitative reasoning results include problem characterization, regulatory basis, comparable cases, evidence strengthening items, and rectification suggestions. Each result field is directly obtained or converted from the objects in the candidate reasoning path.

[0131] The problem characterization is generated from the behavioral constitutive conditions and handling consequence conditions in the candidate regulatory clauses. The data processing server reads the behavioral constitutive conditions marked as covered in the suspicious condition coverage record, uses these behavioral constitutive conditions as the factual behavioral basis for problem characterization, and then reads the handling consequence conditions in the candidate regulatory clauses to form a handling conclusion corresponding to the factual behavior. The regulatory basis is directly taken from the candidate regulatory clauses, and the regulatory clause identifier and the regulatory application condition fields covered by the suspicious factual element set are retained, enabling auditors to trace back the source of characterization along the candidate reasoning path.

[0132] Comparable cases are generated from candidate cases. The data processing server reads the case fact element set and case condition coverage record of the candidate cases, outputs the audit object, fund attributes, behavior mode, and occurrence period fields that are compatible with the suspicious fact element set, and retains the corresponding coverage status. Evidence reinforcement items are converted from the conditions to be supplemented. During the conversion, the procedural conditions and evidence requirement conditions in the conditions to be supplemented are read, and the condition limit values ​​are rewritten into the evidence materials or procedural facts that need to be obtained or verified. Rectification suggestions are generated from candidate rectification measures. During the generation, the rectification action, rectification object, and rectification basis are read, and the rectification action is kept consistent with the handling consequence conditions.

[0133] Each result field in the audit qualitative reasoning results retains a source path identifier, which points to the corresponding candidate reasoning path.

[0134] If multiple candidate reasoning paths exist under the same candidate regulatory clause, the data processing server outputs the audit qualitative reasoning results corresponding to each path according to the sorting key. Each audit qualitative reasoning result retains its source candidate reasoning path, does not merge case condition coverage records from different candidate cases, and does not include rectification measures that did not enter the candidate reasoning path in the rectification recommendations.

[0135] When multiple candidate reasoning paths exist under the same candidate regulatory clause, the data processing server outputs the audit qualitative reasoning results corresponding to each candidate reasoning path according to the sorting key. (Regarding candidate reasoning paths...) Its sort key is: ; in, Candidate reasoning paths Sort key; This indicates whether the applicable object conditions, funding attribute conditions, and behavioral composition conditions in the case condition coverage record are all covered. If true, it is set to 1; otherwise, it is set to 0. The "Date of Occurrence" field indicates that the candidate case and the set of suspicious fact elements belong to the same time period under the condition limit. If the condition is true, the value is 1; otherwise, the value is 0. This indicates that the candidate rectification measures have been directly connected to the conditions for handling consequences through the correspondence between handling consequences; the value is 1 if the condition is met, and 0 otherwise. Candidate cases In the legal case association graph, the write time sequence number is used, with earlier write times resulting in smaller values. The data processing server follows... The candidate reasoning paths are output in ascending lexicographical order; each audit qualitative reasoning result retains its source candidate reasoning path, does not merge case condition coverage records of different candidate cases, and does not include rectification measures that have not entered the candidate reasoning path in the rectification suggestions.

[0136] In the special fund implementation example, the data processing server, based on candidate reasoning paths, outputs a problem characterized as fictitious expenditure of special funds. The legal basis is the corresponding candidate legal clauses, the comparable case is historical case B, the supporting evidence is supplementary approval documents and acceptance materials, and the rectification suggestions are to recover the illegally disbursed funds, supplement the approval materials, and improve the fund payment review process. The problem characterization in the above output comes from the conditions constituting the behavior and the conditions of the consequences, the comparable case comes from the candidate case, the supporting evidence comes from the conditions to be supplemented, and the rectification suggestions come from the candidate rectification measures.

[0137] In step S4, the data processing server filters candidate cases and candidate rectification measures within the conditional instance relationships corresponding to the candidate regulatory clauses. This forms a candidate reasoning path that connects the set of suspicious factual elements, candidate regulatory clauses, conditional instance relationships, candidate cases, and candidate rectification measures, and outputs the audit qualitative reasoning results. These results retain the source relationships of suspicious condition coverage records, case condition coverage records, and conditions requiring supplementary evidence, demonstrating the corresponding chain between problem characterization, regulatory basis, comparable cases, supporting evidence, and rectification recommendations.

[0138] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A legal case association reasoning method oriented towards knowledge graphs, characterized in that, include: Obtain audit doubt texts, regulatory clause texts, historical case texts, and rectification measure texts, and generate doubt fact element sets, regulatory application condition sets, case fact element sets, and rectification measure element sets according to preset audit fact element templates; Perform same-field condition coverage processing on the set of applicable regulatory conditions and the set of case fact elements under the same audit matter category, generate case condition coverage records, establish condition instance relationships, and form a regulatory case relationship graph; The set of factual elements of doubt and the set of applicable legal conditions that are consistent with the audit matter category are subjected to the same field condition coverage process to generate doubt condition coverage record. Candidate legal clauses are determined based on the coverage status of the qualitative necessary conditions and the conflict status of the time conditions in the set of applicable legal conditions. Conditions that are missing and can be supplemented are extracted from the doubt condition coverage record. In the condition instance relationship corresponding to the candidate regulatory clauses, candidate cases are selected based on the field correspondence between the doubt condition coverage record and the case condition coverage record, and candidate rectification measures are selected based on the handling consequences conditions in the candidate regulatory clauses, forming candidate reasoning paths and generating audit qualitative reasoning results.

2. The legal case association reasoning method based on knowledge graphs according to claim 1, characterized in that: The pre-set audit fact element template includes audit item category, audit object, fund attribute, behavior method, approval process, occurrence period, evidence materials, and handling consequences, arranged in a fixed field order; The set of applicable conditions for regulations includes conditions for the applicable objects, conditions for the nature of funds, conditions for the composition of the behavior, procedural conditions, time conditions, conditions for evidentiary requirements, and conditions for the consequences of handling. The conditions for applicable objects, financial attributes, and behavioral composition are qualitatively necessary conditions; the conditions for procedural conditions and evidentiary requirements are supplementary conditions; the conditions for handling consequences are result-corresponding conditions; and the time point conditions are independently retained for the period of application of the regulations.

3. The legal case association reasoning method based on knowledge graphs according to claim 2, characterized in that, The same-field condition coverage processing uses the case fact element set and the doubtful fact element set as the comparison element sets, respectively, including: Using the condition fields and condition restriction values ​​in the set of applicable regulations as the comparison benchmark, find the fact values ​​of the same fields in the set of elements being compared. If no fact value for the same field is found, the corresponding coverage status will be written as condition missing; When a fact value with the same field is found, the same value judgment, inclusion judgment, period fall-in judgment and normalized synonym merge judgment are executed in sequence. If any judgment is true, the coverage status is written as the condition is covered. If none of the judgments are true, the field exclusion relationship table is used to determine whether the fact value and condition restriction value of the same field are mutually exclusive. If they are mutually exclusive, the coverage status is written as condition conflict; if they are not mutually exclusive, the coverage status is written as condition missing.

4. The legal case association reasoning method based on knowledge graphs according to claim 3, characterized in that, Establishing conditional instance relationships includes: Collect case condition coverage records under the same legal clause identifier and the same case identifier; When there are no conflicts between the qualitative necessary conditions and the time conditions, the legal provisions, legal application conditions, case fact elements, and case condition coverage records corresponding to the set of applicable legal conditions are written into the condition instance relationship. When there are missing conditions but no conflicting conditions in the qualitative necessary conditions, the case fact element set is written into the condition instance relationship as an incomplete coverage instance, and the corresponding missing condition state is retained. When any necessary condition for certainty has a conditional conflict or a conditional point in time has a conditional conflict, the corresponding set of case fact elements will not be written into the conditional instance relation.

5. The legal case association reasoning method based on knowledge graphs according to claim 1, characterized in that, When generating the set of factual elements of doubt and the set of factual elements of the case, a negative semantic judgment is performed before the factual values ​​of the evidence material fields are written, including: When negative terms appear before or after a noun of evidence, the noun of evidence is not written into the obtained fact value of the evidence material field, and the noun of evidence is retained as a condition missing clue. When an affirmative term appears in the text segment containing the evidence term, the evidence term is written into the evidence material field; When the same evidence term has both positive and negative descriptions, the text segment containing at least one conclusive term from among "audit opinion," "problem manifestation," "handling decision," "rectification requirement," "recovery requirement," and "order to rectify" shall be identified as the audit conclusion segment; if no text segment contains conclusive terms, the last text segment in the corresponding text source that contains the behavior method or handling consequence shall be identified as the audit conclusion segment. Determine the text paragraph distance between the affirmative and negative descriptive sentences and the audit conclusion sentences, and use the description with the smaller distance as the basis for writing; when the distances are the same, use the negative description as a clue of missing conditions, and do not write the evidence terms into the obtained fact values.

6. The legal case association reasoning method based on knowledge graphs according to claim 1, characterized in that, The generated set of applicable regulatory conditions includes: The text of the legal provisions is broken down into conditional statements according to the clause number, item number, and conditional connector. The conditional fields, conditional limit values, and conditional roles are determined based on the semantic function of the conditional statements. When the same conditional statement contains both procedural requirements and evidentiary requirements, the conditional statement is split into procedural conditions and evidentiary requirement conditions according to the core nouns in the conditional statement, and the conditional roles of the procedural conditions and evidentiary requirement conditions are determined as supplementary evidence conditions respectively. When the same legal provision has multiple parallel acts constituting it, multiple sets of applicable conditions are formed according to the multiple parallel acts constituting it. Each set of applicable conditions shares the same legal provision identifier and has different condition restriction values ​​for the acts constituting it. When the text of a regulation only specifies the consequences of handling without recording the evidence requirements, the conditions for handling consequences should be written into the conditions corresponding to the results, and the conditions for evidence requirements should be kept in an empty state.

7. The legal case association reasoning method based on knowledge graphs according to claim 1, characterized in that, The candidate rectification measures include: Read the corresponding relationships of processing consequences, rectification actions, rectification targets, and rectification basis in sequence from the elements of the rectification measures; When the corresponding relationship of the handling consequences directly points to the handling consequences conditions in the candidate regulations, the set of rectification measures elements will be written into the candidate rectification measures. When the correspondence between the consequences and the conditions does not directly point to the conditions for handling consequences, compare whether the rectification action and the handling result in the conditions for handling consequences belong to the same standard expression, and compare whether the object of rectification and the object pointed to by the conditions for handling consequences are consistent. When the rectification action and the handling result belong to the same standard expression and the rectification object and the handling consequence condition point to the same object, the rectification measure element set is written into the candidate rectification measures; when the rectification basis points to the same legal clause identifier as the candidate legal clause, the rectification basis is used as the matching confirmation basis. When the rectification action and the handling result do not belong to the same standard expression, or the object of rectification and the object pointed to by the handling result conditions are inconsistent, the set of rectification measures elements shall not be written into the candidate rectification measures.

8. The legal case association reasoning method based on knowledge graphs according to claim 1, characterized in that: The legal case relationship graph uses legal clause nodes, legal application condition nodes, case fact nodes, and rectification measure nodes as graph nodes, and condition instance relationships as graph edges. Each condition instance relationship connects at least one legal clause node, one legal application condition node, and one case fact node. When there is a set of corrective measures elements that match the conditions for handling consequences, the condition instance relationship also connects the corresponding corrective measures node. The attributes of the graph include the correspondence between case condition coverage records and processing consequences. Case condition coverage records include legal clause identifiers, case identifiers, condition fields, condition limit values, corresponding fact values, and coverage status. The sets of factual elements of doubt, the sets of applicable legal conditions, the sets of factual elements of cases, and the sets of elements of rectification measures retain the text source identifier and the text paragraph position, and the text paragraph position is associated with the text sentence or segment from which the corresponding element originates.

9. The legal case association reasoning method based on knowledge graphs according to claim 1, characterized in that, The qualitative reasoning results of the audit include the characterization of the problem, the legal basis, comparable cases, supporting evidence, and rectification recommendations, among which: Read the behavior marked as covered in the doubtful condition coverage record as the factual behavior basis for the problem characterization, and read the processing consequence conditions in the candidate legal clauses to form the processing conclusion corresponding to the factual behavior basis; The candidate regulatory clauses will be used as the basis for the regulations, and the regulatory clause identifiers and the regulatory application conditions fields covered by the set of questionable fact elements will be retained. Read the case fact element set and case condition coverage record of the candidate case, output the audit object, fund attribute, behavior mode and occurrence period fields that are compatible with the suspicious fact element set, and retain the corresponding coverage status to form comparable cases; The conditional limits of the procedural conditions and evidentiary requirements in the conditions to be supplemented are converted into evidence materials or procedural facts that need to be obtained or verified, thus forming evidence reinforcement items; Read the rectification actions, rectification targets, and rectification basis from the candidate rectification measures, form rectification suggestions, and ensure that the rectification actions correspond to the conditions of the handling consequences; The problem characterization, legal basis, comparable cases, supporting evidence, and rectification suggestions each retain source path identifiers pointing to the corresponding candidate reasoning paths.

10. The legal case association reasoning method based on knowledge graphs according to claim 9, characterized in that, When there are multiple candidate reasoning paths under the same candidate legal clause, a sorting key is generated for each candidate reasoning path. The sorting key sequentially records whether the applicable object conditions, fund attribute conditions, and behavioral composition conditions in the corresponding case condition coverage record are all covered, whether the occurrence period field of the candidate case and the suspicious fact element set belong to the same time point condition limitation period, whether the candidate rectification measures are directly connected to the handling consequence conditions through the handling consequence correspondence relationship, and the writing time sequence number of the candidate case in the legal case association graph. Output the audit qualitative reasoning results corresponding to each candidate reasoning path according to the sorting key. Each audit qualitative reasoning result retains its source candidate reasoning path. Case condition coverage records of different candidate cases are not merged, and rectification measures that have not entered the candidate reasoning path are not included in the rectification suggestions.

Images