Legal eligible contract clause intelligent review method and device

Abstract

The application discloses a kind of legal qualified contract clause intelligent review method and device, it is related to artificial intelligence technical field.The method includes: obtaining the contract text to be reviewed, extracts hierarchical structure and atomic clause by structured analysis, generates clause semantic vector;Multi-dimensional legal compliance knowledge graph comprising rule condition tree structure is constructed;Atomic clause is split into the atomic fact of the form of quadruple, and the target compliance rule set is obtained by bidirectional alignment matching;Combined with the matching result, build review prompt word input large language model, obtain compliance review result;By structured analysis and confidence evaluation, generate hierarchical contract overall compliance review report.The application realizes rule condition complete coverage verification and numerical boundary accurate matching, improves contract review efficiency and result consistency, and can be widely applied to legal compliance management scene.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and apparatus for intelligent review of legally compliant contract terms. Background Technology

[0002] As enterprises deepen their digital transformation, contract review, as a core component of corporate legal and compliance management, faces increasing challenges in terms of efficiency and accuracy. Traditional contract review primarily relies on legal personnel or lawyers to manually review each clause, which is not only time-consuming and labor-intensive but also prone to overlooking risk clauses due to fatigue or negligence. Furthermore, different reviewers may have different understandings and judgment standards for the same contract clause, leading to a lack of consistency and repeatability in the review results.

[0003] In recent years, artificial intelligence technology has received widespread attention and application in the field of contract review. Some existing solutions attempt to automate contract text review using natural language processing, deep learning, and other methods. However, these solutions generally lack a deep semantic understanding of the legal compliance of contract clauses, and their rule matching methods are all one-way retrieval—either retrieving relevant information from the knowledge graph based on the contract clauses, or comparing the knowledge graph with the clauses based on their prototypes. No solution employs a two-way verification matching architecture. One-way retrieval can only discover which clauses contain information related to the rules, but cannot verify whether all conditions required by the rules are fully covered by the clauses. For example, when a clause contains "payment period of 90 days" while the rule requires "payment period ≤ 30 days," one-way retrieval might incorrectly conclude that the clause is related to the rule based on the keyword "payment period," but it cannot detect the substantial non-compliance of "90 days > 30 days." Existing solutions also have significant shortcomings in terms of review granularity, failing to further break down clauses into finer-grained atomic facts for independent matching. When a clause involves multiple legal issues simultaneously (such as a clause stipulating both delivery and payment obligations), existing solutions cannot precisely pinpoint the specific factual basis for the violation, and can only generally determine that the clause is non-compliant. Existing technical triples lack the structured capacity to carry constraint information such as time, quantity, method, and conditions. The core of legal compliance review lies in the precise comparison of numerical boundaries and conditional constraints (such as "within 30 days," "no fewer than three people," "with the other party's written consent," etc.). Triples lacking constraint fields cannot support such precise matching and can only degenerate into semantic similarity comparisons, failing to distinguish the essential difference between "30 days" and "90 days." Summary of the Invention

[0004] The present invention provides a method and apparatus for intelligent review of legally compliant contract terms, aiming to solve the problems that existing technologies cannot verify the complete coverage of rule conditions, cannot accurately locate the factual basis of specific violations, and / or cannot support the precise matching of numerical boundaries.

[0005] The first aspect of this invention provides an intelligent review method for legally compliant contract terms, comprising: Obtain the contract text to be reviewed, and perform structured parsing on the contract text to extract the hierarchical structure and atomic clauses in the contract text to be reviewed. Then, perform semantic vectorization processing on each atomic clause to obtain the clause semantic vector corresponding to each atomic clause. Construct a multi-dimensional legal compliance knowledge graph; the multi-dimensional legal compliance knowledge graph includes legal and regulatory entities, compliance rule entities, and the relationships between these entities; each compliance rule is represented in a structured form as a rule condition tree, where the internal nodes of the rule condition tree are logical operators, and the leaf nodes are atomic conditions; The atomic clauses are broken down into atomic facts, and the atomic facts are used as the smallest matching units to perform bidirectional alignment compliance rule matching on the multi-dimensional legal compliance knowledge graph to obtain the target compliance rule set corresponding to the atomic clauses. Based on the target compliance rule set, review prompt words are constructed, and the review prompt words are input into a preset large language model to obtain the clause compliance review results output by the large language model; The clause compliance review results output by the large language model are subjected to structured parsing and confidence assessment to generate a structured clause review report. Based on the hierarchical structure in the contract text to be reviewed, the clause review report is constructed into a hierarchical overall contract compliance review report.

[0006] Further, the contract text to be reviewed is obtained, and the contract text to be reviewed is subjected to structured parsing to extract the hierarchical structure and atomic clauses in the contract text to be reviewed, including: Obtain the contract text to be reviewed, and perform a first preprocessing on the contract text to be reviewed to obtain the contract text to be reviewed after the first preprocessing. Using the number of indentation characters, numbering pattern, relative font size, and type of the first line starting character of each paragraph in the contract text to be reviewed after the first preprocessing as input features, a Transformer-based sequence labeling clause structure recognition model is used to obtain the hierarchical labels of each paragraph. A hierarchical relationship tree of clauses is constructed based on the hierarchical tags, and the text content under each level in the hierarchical relationship tree is delineated to determine the atomic clauses; the atomic clause refers to the lowest level complete semantic unit in the hierarchical relationship tree that does not contain any child clause numbers.

[0007] Further, each of the atomic clauses undergoes semantic vectorization processing to obtain a clause semantic vector corresponding to each atomic clause, including: The atomic clauses are subjected to a second preprocessing to obtain atomic clauses after the second preprocessing. A pre-trained legal domain semantic encoding model is used to encode the atomic clauses after the second preprocessing to generate a fixed-dimensional semantic vector, thereby obtaining the clause semantic vector corresponding to the atomic clause. Iterate through each atomic clause to obtain the clause semantic vector corresponding to each atomic clause.

[0008] Furthermore, a multi-dimensional legal compliance knowledge graph will be constructed, including: Collect legal and regulatory documents related to the contract, and perform entity identification and relationship extraction on the legal and regulatory documents to extract legal and regulatory clauses and compliance rule entities, as well as the relationships between the entities; each compliance rule is represented in a structured form as a rule condition tree, where the internal nodes of the rule condition tree are logical operators and the leaf nodes are atomic conditions; The extracted entities and relationships are stored in the Neo4j graph database as triples to construct a multi-dimensional legal compliance knowledge graph.

[0009] Furthermore, the atomic facts are represented as quadruples, specifically including: ; in, The fact type is determined by a standardized type selected from a predefined fact type dictionary. It represents the core legal subject or object involved in the atomic fact, and its value is the entity name after being mapped through legal terminology standardization. The core predicates representing atomic facts are taken from standard predicates selected from a predefined legal predicate dictionary; This represents the set of constraints on the atomic fact, where each constraint is a constraint attribute, a constraint operator, or a constraint value.

[0010] Furthermore, using the atomic facts as the smallest matching unit, bidirectional alignment compliance rule matching is performed on the multi-dimensional legal compliance knowledge graph to obtain the target compliance rule set corresponding to the atomic clauses, including: For each compliance rule in the multi-dimensional legal compliance knowledge graph, extract all leaf nodes in the condition tree corresponding to the compliance rule, and in the atomic conditions corresponding to each leaf node, search for whether there is a matching item in all atomic facts of the atomic clause, obtain the matching result, and generate the rule fact matching degree based on the matching result; For each atomic fact of the atomic clause, a search is performed in all rule condition leaf nodes of the multi-dimensional legal compliance knowledge graph to obtain candidate rules, and atomic condition matching is performed on the candidate rules to generate fact rule matching degree; Based on the rule fact matching degree and the fact rule matching degree, a comprehensive matching score is calculated, and the target compliance rule set is determined based on the comprehensive matching score.

[0011] Further, review prompts are constructed based on the target compliance rule set, and these prompts are input into a preset large language model to obtain the clause compliance review results output by the large language model, including: The original text of the atomic clause, the atomic facts, the target compliance rule set, the two-way alignment matching result summary, and the preset review instruction template are combined to generate review prompt words; Input the review prompts into a preset large language model and obtain the clause compliance review results output by the large language model.

[0012] Furthermore, the compliance review results of the terms output by the large language model are subjected to structured parsing and confidence assessment to generate a structured terms review report, including: The compliance review results of the terms output by the large language model are subjected to format verification and structured parsing to extract fields such as compliance judgment result, risk level, violation basis clause and / or modification suggestion, and obtain structured parsing results; Simultaneously, a multi-factor weighted model was used to obtain confidence assessment; Based on the structured analysis results and confidence assessment, a structured clause review report is generated.

[0013] Furthermore, based on the hierarchical structure of the contract text to be reviewed, the clause review report is constructed into a hierarchical overall contract compliance review report, including: The structured clause review report corresponding to the atomic clause is organized according to the hierarchical structure in the contract text to be reviewed, generating a hierarchical overall contract compliance review report.

[0014] A second aspect of the present invention provides an intelligent review device for legally compliant contract terms, comprising: The contract parsing module is used to obtain the contract text to be reviewed, and to perform structured parsing on the contract text to be reviewed in order to extract the hierarchical structure and atomic clauses in the contract text to be reviewed, and to perform semantic vectorization processing on each atomic clause to obtain the clause semantic vector corresponding to each atomic clause. The graph construction module is used to construct a multi-dimensional legal compliance knowledge graph. The multi-dimensional legal compliance knowledge graph includes legal and regulatory entities, compliance rule entities, and the relationships between these entities. Each compliance rule is represented in a structured form as a rule condition tree, where the internal nodes of the rule condition tree are logical operators and the leaf nodes are atomic conditions. The rule matching module is used to break down the atomic clauses into atomic facts, and use the atomic facts as the smallest matching unit to perform bidirectional alignment compliance rule matching on the multi-dimensional legal compliance knowledge graph to obtain the target compliance rule set corresponding to the atomic clauses; The compliance review module is used to construct review prompts based on the target compliance rule set, input the review prompts into a preset large language model, and obtain the clause compliance review results output by the large language model. The report output module is used to perform structured parsing and confidence assessment on the clause compliance review results output by the large language model, generate a structured clause review report, and construct the clause review report into a hierarchical overall contract compliance review report based on the hierarchical structure in the contract text to be reviewed.

[0015] Beneficial effects: This invention provides an intelligent review method and apparatus for legally compliant contract clauses. The method includes: acquiring the contract text to be reviewed; extracting hierarchical structure and atomic clauses through structured parsing to generate clause semantic vectors; constructing a multi-dimensional legal compliance knowledge graph containing a rule condition tree structure; splitting atomic clauses into atomic facts in the form of four-tuples; obtaining a target compliance rule set through bidirectional alignment matching; constructing a large language model for review prompt words based on the matching results to obtain compliance review results; and generating a hierarchical overall contract compliance review report through structured parsing and confidence assessment. This invention achieves complete rule condition coverage verification and precise numerical boundary matching, improving contract review efficiency and result consistency, and can be widely applied in legal compliance management scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an intelligent review method for legally compliant contract terms proposed in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an intelligent review device for legally compliant contract terms proposed in an embodiment of the present invention; Among them, 201-Contract parsing module, 202-Graph construction module, 203-Rule matching module, 204-Compliance review module, and 205-Report output module. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0019] like Figure 1 As shown, this embodiment of the invention provides an intelligent review method for legally compliant contract terms, including: S101. Obtain the contract text to be reviewed, and perform structured parsing on the contract text to be reviewed to extract the hierarchical structure and atomic clauses in the contract text to be reviewed, and perform semantic vectorization processing on each atomic clause to obtain the clause semantic vector corresponding to each atomic clause. S102. Construct a multi-dimensional legal compliance knowledge graph; the multi-dimensional legal compliance knowledge graph includes legal and regulatory entities, compliance rule entities, and the relationships between these entities; each compliance rule is represented in a structured form as a rule condition tree, where the internal nodes of the rule condition tree are logical operators, and the leaf nodes are atomic conditions; S103. The atomic clause is broken down into atomic facts, and the atomic facts are used as the smallest matching unit to perform bidirectional alignment compliance rule matching on the multi-dimensional legal compliance knowledge graph to obtain the target compliance rule set corresponding to the atomic clause. S104. Construct review prompt words based on the target compliance rule set, and input the review prompt words into a preset large language model to obtain the clause compliance review results output by the large language model; S105. Perform structured parsing and confidence assessment on the clause compliance review results output by the large language model to generate a structured clause review report, and construct the clause review report into a hierarchical overall contract compliance review report based on the hierarchical structure in the contract text to be reviewed.

[0020] In this embodiment of the invention, the contract text to be reviewed is obtained, and the contract text to be reviewed is subjected to structured parsing to extract the hierarchical structure and atomic clauses in the contract text to be reviewed, including: Obtain the contract text to be reviewed, and perform a first preprocessing on the contract text to be reviewed to obtain the contract text to be reviewed after the first preprocessing. Using the number of indentation characters, numbering pattern, relative font size, and type of the first line starting character of each paragraph in the contract text to be reviewed after the first preprocessing as input features, a Transformer-based sequence labeling clause structure recognition model is used to obtain the hierarchical labels of each paragraph. Construct a hierarchical relationship tree of clauses according to the hierarchical labels, perform boundary division on the text content under each level in the hierarchical relationship tree, and determine atomic clauses; the atomic clause refers to the complete semantic unit at the lowest level in the hierarchical relationship tree that does not contain any sub-clause numbers.

[0021] For example, a contract document to be reviewed is obtained, and the contract document includes but is not limited to the formats of Word document, PDF document, and scanned image. Format unification processing is performed on the contract document: for the scanned image, the PaddleOCR optical character recognition engine is used for text recognition and extraction; for the PDF document, the PDFBox parsing library is used to extract text content; for the Word document, the Apache POI parsing library is used to extract text content. Contract documents in all formats are uniformly converted into a plain text format with paragraph marks, so as to realize the first preprocessing.

[0022] Subsequently, a preset clause structure recognition model is used to perform structured parsing on the to-be-reviewed contract text. The clause structure recognition model is trained based on a deep learning algorithm, and is used to recognize the hierarchical structures of chapter, section, article, paragraph, and item in the contract text. Specifically, the clause structure recognition model adopts a Transformer-based sequence labeling architecture, and the input features include the number of indented characters of each paragraph (an integer with a value from 0 to 8), the numbering pattern (regular matching recognizes patterns such as "Chapter X", "Article X", "(X)", "X.", and outputs the corresponding hierarchical label), the font size (a relative value, taking the font size of the main text of the contract as the benchmark 1.0, and the font size of the title is usually 1.2 to 2.0), and whether the first line of the paragraph has typical starting characters of legal documents such as "第", "(", "一" (a Boolean value). The model training dataset is 50,000 pieces of various contract texts with labeled clause hierarchical structures. Under the condition of standard digital text (that is, non-OCR recognized text, which is a directly editable electronic document format), the F1 value of boundary recognition of the model on the test set reaches 0.92; for OCR recognized text, the F1 value is 0.87. The F1 value is a performance indicator in the model evaluation stage, while the boundary recognition confidence is the probability value predicted by the model for the boundary of each specific atomic clause (the maximum probability value of the softmax output layer, with a value range of [0, 1]). The F1 value reflects the overall generalization performance of the model, and the confidence reflects the reliability of the model for a single prediction result. When the boundary recognition confidence of a single atomic clause is lower than 0.90, a boundary confirmation interaction can be automatically triggered, the corresponding clause is pushed to the user interface in a highlighted form, and the subsequent steps are performed after waiting for manual click confirmation or correction.

[0023] Through structured parsing, the contract text is decomposed into a hierarchical clause tree structure, and the smallest indivisible clause unit is extracted to obtain the atomic clause. The atomic clause is defined as: the lowest level complete semantic unit in the clause tree that no longer contains any child clause numbers (i.e., no longer contains sub-numbers such as "(X)" or "X.").

[0024] Optionally, when the boundary identification confidence level of any atomic clause is below 0.90, the text range of that atomic clause is highlighted in the user interface, and a confirmation dialog box is generated, waiting for the user to click to confirm that the boundary is correct or manually drag to adjust the boundary. Only after user confirmation is the atomic clause allowed to proceed to step S102. If the user does not complete the confirmation within a preset time of 60 minutes, the system automatically marks the clause as "awaiting manual confirmation" and skips the review, listing it separately in the final report.

[0025] In this embodiment of the invention, semantic vectorization processing is performed on each atomic clause to obtain the clause semantic vector corresponding to each atomic clause, including: The atomic clauses are subjected to a second preprocessing to obtain atomic clauses after the second preprocessing. A pre-trained legal domain semantic encoding model is used to encode the atomic clauses after the second preprocessing to generate a fixed-dimensional semantic vector, thereby obtaining the clause semantic vector corresponding to the atomic clause. Iterate through each atomic clause to obtain the clause semantic vector corresponding to each atomic clause.

[0026] For example, the second preprocessing of the atomic clause may include: removing redundant whitespace characters (merging multiple consecutive spaces or line breaks into one space), unifying Chinese punctuation (converting English commas and periods into Chinese commas and periods, and converting full-width numbers and letters into half-width numbers), and performing standardized replacement of legal terms (constructing a legal term synonym mapping table, for example, uniformly mapping "Party A", "Buyer", and "Purchaser" to "Buyer"; uniformly mapping "Party B", "Seller", and "Supplier" to "Seller"; and uniformly mapping "Payment", "Price", and "Contract Price" to "Price").

[0027] The pre-trained legal domain semantic encoding model Law-BERT can be used to encode the text content of each atomic clause, generating a 768-dimensional clause semantic vector. The Law-BERT model is based on the BERT-base architecture (12 Transformer layers, 768-dimensional hidden layers, and 12 attention heads), and is further pre-trained in a domain-adaptive manner on a large-scale legal text corpus (including 300,000 judgment documents, 50,000 legal provisions, and 100,000 contract templates) (100,000 pre-training steps, a learning rate of 2e-5, and a batch size of 32) to improve the model's ability to capture the semantics of legal texts.

[0028] In this embodiment of the invention, a multi-dimensional legal compliance knowledge graph is constructed, including: Collect legal and regulatory documents related to the contract, and perform entity identification and relationship extraction on the legal and regulatory documents to extract legal and regulatory clauses, compliance rule entities, and relationships between various entities; each compliance rule is represented in a structured form as a rule condition tree, where the internal nodes of the rule condition tree are logical operators and the leaf nodes are atomic conditions; The extracted entities and relationships are stored in the Neo4j graph database as triples to construct a multi-dimensional legal compliance knowledge graph.

[0029] For example, legal compliance data can be collected from national legal databases, Supreme People's Court judicial interpretation databases, and industry standard and specification databases, covering relevant laws and regulations such as the Contract Law of the Civil Code of the People's Republic of China, the Government Procurement Law of the People's Republic of China, and the Bidding Law of the People's Republic of China, as well as corresponding judicial interpretations, thereby obtaining legal and regulatory documents related to contracts.

[0030] Then, entity recognition and relationship extraction are performed on the collected data to extract legal and regulatory clauses, compliance rules entities, and the reference relationships, interpretation relationships, and refinement relationships between entities; Compliance rules include at least the following types: mandatory rules (rules containing obligations expressed as "shall" or "must"), prohibitive rules (rules containing prohibitions expressed as "shall not" or "prohibited"), and authorizing rules (rules containing rights expressed as "may" or "have the right").

[0031] Each compliance rule is structured and stored in the form of a rule condition tree. The rule condition tree is defined as a binary tree where internal nodes are logical operators (AND, OR, NOT), and leaf nodes are atomic conditions. Each atomic condition is represented as a quintuple (...). ),in: The value represents the legal fact attribute involved in the condition. It is a standardized attribute name selected from a predefined legal fact attribute dictionary. This dictionary contains 50 commonly used legal fact attributes, such as "contract type", "price agreement status", "payment period in days", "delivery period in days", "subject matter status", and "acceptance status". This represents the comparison operators, with values ​​{=, ≠, >, ≥, <, ≤, exist,} }; Indicates the target value for comparison; The data type representing the target value has the following values: {NUMERIC, DATE, ENUM, BOOLEAN, STRING}. This field indicates the specific legal provision number upon which the atomic condition is based. Legal provisions use this field to provide an anchor for the validity and retrospective basis of all compliant rules, forming the legal cornerstone of full interpretability. Specifically, `exists` is used to check whether the provision includes this fact. Used to check if the terms do not specify this fact; NUMERIC is used to represent continuous or discrete numbers that can be compared and performed arithmetic operations; DATE is used to represent a specific point in time or the end of a time period; ENUM values ​​are limited to a specific option in a predefined finite set, do not have size comparison attributes, and only support equal to or not equal to comparisons; BOOLEAN represents Boolean data, containing only two logical states: TRUE (existence / yes) or FALSE (non-existence / no); STRING represents unstructured free text.

[0032] For example, Article 626 of the Civil Code of the People's Republic of China (Rules for Payment of Price) is represented as the following condition tree (root node is AND): Leaf node 1: (Contract type, =, "Sales contract", ENUM, "Article 595 of the Civil Code of the People's Republic of China"); Leaf node 2: (Price agreement status, exists, TRUE, BOOLEAN, "Article 626 of the Civil Code of the People's Republic of China"); Leaf node 3: (Payment period in days, ≤, 30, NUMERIC, "Article 626 of the Civil Code of the People's Republic of China") The extracted entities and relations are stored in the Neo4j graph database in the form of (head entity, relation, tail entity) triples to construct a legal compliance knowledge graph, which contains approximately 150,000 nodes and 420,000 relations. Optionally, every Monday at 2:00 AM, the latest legal and regulatory revisions and new judicial interpretations can be retrieved from official data sources to dynamically update the knowledge graph. During each update, if the content of a rule changes, the existing rule node is not overwritten; instead, a new rule node is created and assigned a new version number (in the format "v_" + 8-digit date, such as "v_20260720"). Simultaneously, the "Expiration Date" field of the old rule node is updated to the change date, and the "Effective Date" field of the new rule node is set to the change date.

[0033] In this embodiment of the invention, the atomic fact is represented as a quadruple, specifically including: ; in, The fact type is determined by a standardized type selected from a predefined fact type dictionary. It represents the core legal subject or object involved in the atomic fact, and its value is the entity name after being mapped through legal terminology standardization. The core predicates representing atomic facts are taken from standard predicates selected from a predefined legal predicate dictionary; This represents the set of constraints on the atomic fact, where each constraint is a constraint attribute, a constraint operator, or a constraint value.

[0034] For example, atomic clauses can be broken down into atomic facts. An atomic fact is defined as the smallest indivisible factual proposition that constitutes a compliance judgment, represented as a quadruple. ,in: This indicates the fact type, and its value is a standardized type selected from a predefined fact type dictionary. This dictionary contains six types: "obligation", "right", "prohibited behavior", "authorized behavior", "fact description", and "definition". This indicates the core legal subject or object involved in the fact, and its value is the entity name after being mapped through standardized legal terminology. The core predicate (action or state) representing this fact takes values ​​from a predefined legal predicate dictionary containing 80 standard predicates, including "payment", "delivery", "guarantee", "indemnification", "notification", "acceptance", "assignment", and "license". This represents the set of conditions that constrain the fact, with each condition being a triple (qualifying attribute, qualifying operator, qualifying value).

[0035] The quadruple of this invention adds a constraints field, which can carry constraint information such as time, quantity, method, and conditions (e.g., "within 30 days", "after acceptance", "in written form"), making precise matching of numerical boundaries possible. Taking "payment period shall not exceed 30 days" as an example, the triple of the prior art can usually only be represented as (buyer, payment, price), losing the key constraint information of "30 days"; the quadruple of this invention can be represented as (obligation, buyer, payment, {term ≤ 30 days}), completely preserving all semantic information of the rule condition.

[0036] The method for extracting atomic facts is as follows: An extraction pipeline based on dependency parsing is employed. First, the HanLP dependency parser is used to perform syntactic parsing on the atomic clause text, identifying all verbs in the text as candidate predicates. For each candidate predicate, its agent, patient, time, manner, and conditional arguments are filled in according to syntactic dependency relations (including subject-verb (SBV), verb-object (VOB), attributive-head (ATT), and adverbial-head (ADV). The <predicate, argument 1, argument 2, ...> structure is then transformed into the aforementioned four-tuple structure. When all fields of a four-tuple cannot be completely filled from the syntactic parsing, the extraction result is marked as an incomplete fact and does not participate in subsequent matching calculations, but it is listed in the missing information alert section of the final report.

[0037] In this embodiment of the invention, using the atomic fact as the smallest matching unit, bidirectional alignment compliance rule matching is performed on the multi-dimensional legal compliance knowledge graph to obtain the target compliance rule set corresponding to the atomic clause, including: For each compliance rule in the multi-dimensional legal compliance knowledge graph, extract all leaf nodes in the condition tree corresponding to the compliance rule, and in the atomic conditions corresponding to each leaf node, search for whether there is a matching item in all atomic facts of the atomic clause, obtain the matching result, and generate the rule fact matching degree based on the matching result; For each atomic fact of the atomic clause, a search is performed in all rule condition leaf nodes of the multi-dimensional legal compliance knowledge graph to obtain candidate rules, and atomic condition matching is performed on the candidate rules to generate fact rule matching degree; Based on the rule fact matching degree and the fact rule matching degree, a comprehensive matching score is calculated, and the target compliance rule set is determined based on the comprehensive matching score.

[0038] For example, for each compliance rule in the knowledge graph, a depth-first traversal is performed on the rule's condition tree to extract all leaf nodes (i.e., all atomic conditions). For each atomic condition, a match is searched among all atomic fact quadruples of that atomic clause. The matching criterion is: the atomic facts... With atomic conditions The required type is consistent, atomic facts and With atomic conditions ( The cosine similarity after semantic encoding is no less than 0.82. This similarity threshold of 0.82 was obtained by performing grid search optimization on 10,000 legal and contractual clause comparison data labeled "match / non-match", achieving an optimal balance point with an F1 score of 0.89 on the validation set.

[0039] In particular, when the atomic conditions are... When the value is NUMERIC (e.g., payment period in days, amount, quantity), the matching determination does not use semantic similarity comparison, but instead extracts the atomic facts. The numerical type limitation in the above, and the atomic condition in the above. Perform numerical comparisons, that is, according to The specified comparison operators (≤, ≥, <, >, =, ≠) are used to determine numerical size, rather than semantic similarity. For example, when the atomic condition is (payment period in days, ≤, 30, NUMERIC, ...) and the constraint condition of the atomic fact is {duration = 60 days}, the system performs a numerical comparison "60≤30", which results in 0. This directly determines that the atomic condition does not match, avoiding the problem of pure semantic similarity matching misjudging "60 days" and "30 days" as semantically similar.

[0040] The rule fact matching degree of this rule when all leaf nodes are marked as matched. The value is 1.0; if there are unmatched leaf nodes, then... This represents the number of matched leaf nodes divided by the total number of leaf nodes.

[0041] For each atomic fact quadruple of atomic clauses, its ( , , After concatenation into text, the data is encoded into a 768-dimensional vector. A KNN search is then performed on the leaf node vectors of all rules in the knowledge graph, with K=10, and cosine distance is used as the retrieval distance metric. For each retrieved candidate rule, atomic condition matching is performed, and the matching degree of each candidate rule is recorded. .

[0042] For each candidate rule, calculate the comprehensive matching score. The calculation formula is: ; in, The value range is [0, 1]; The value range is [0, 1], and the weights for both directions are 0.5.

[0043] When the rule-fact matching degree is high, but the fact-rule matching degree is low (i.e., the rule's conditions are all expressed in the clause, but the rule cannot be retrieved from the clause's facts), it indicates that a rule may be missing, triggering a deep search. When the fact-rule matching degree is high, but the rule-fact matching degree is low (i.e., the rule was retrieved from the clause's facts, but some of the rule's conditions are not covered in the clause), it indicates that the clause has missing conditions, and it is marked as non-compliant. Through cross-validation in both directions, the bidirectional alignment mechanism of this invention can reduce the false positive rate of keyword matching but numerical non-compliance.

[0044] when When the MatchScore is ≥0.75, the rule is included in the target compliance rule set for that atomic clause. This threshold of 0.75 was also obtained through grid search optimization on the labeled dataset, achieving a balance between precision (0.88) and recall (0.87) on the validation set. When the number of rules in the target compliance rule set exceeds 10, the top 10 rules are selected as the final target compliance rule set, sorted from highest to lowest MatchScore.

[0045] In this embodiment of the invention, review prompts are constructed based on the target compliance rule set, and the review prompts are input into a preset large language model to obtain the clause compliance review results output by the large language model, including: The original text of the atomic clause, the atomic facts, the target compliance rule set, the two-way alignment matching result summary, and the preset review instruction template are combined to generate review prompt words; Input the review prompts into a preset large language model and obtain the clause compliance review results output by the large language model.

[0046] For example, the review prompts can be constructed as follows: the original text of the atomic clause, the clause semantic vector (injected as contextual information), and the target compliance rule set (each rule contains the full text of the rule condition tree and its corresponding...). , , The details are combined with the preset review instruction template.

[0047] Optionally, the review instruction template may include one or more of the following elements: Character description: "You are a professional contract law review expert, proficient in contract law and related laws and regulations"; A structured list of extracted atomic facts (listed in quadruplets). A list of matched rule conditions (listed in indented text, with each leaf node indicating its matching status). Summary of bidirectional alignment matching results ( and (Specific numerical values ​​and difference analysis); Review task description: "Based on the above matched structured facts and rules data, please complete the following tasks one by one." Output format requirements: "Please output the review results in JSON format, including compliance judgments (field names)." Values ​​can be 'compliant', 'non-compliant', or 'requires further review', and the risk level (field name) is... Values ​​can be 'high', 'medium', or 'low'; violation criteria (field name) If a violation exists, fill in the specific description of the violated rule conditions; otherwise, leave null. Modification suggestions (field name: suggestion; if the rule is not compliant, generate a modification plan based on the constraint values ​​in the matched target compliant rule; otherwise, leave null). A brief description of the reasoning process (field name: reasoning).

[0048] The large language model of this invention does not undertake the core tasks of fact extraction and rule matching. All factual content (atomic fact quadruples, rule condition trees, and matching states) is provided by structured processing, eliminating the risk of fabricating legal basis or numerical values ​​in the large language model.

[0049] The constructed review prompts are input into a large language model, which is a generative large language model fine-tuned from legal domain instructions. Specifically, it is a version of the ChatGLM3-6B-Base model fine-tuned on a legal domain instruction dataset (containing 100,000 legal question-and-answer pairs and contract review instructions). The fine-tuning method is LoRA (low-rank adaptation, rank r=8, learning rate 3e-4, training epochs 3). The temperature parameter is set to 0.1 during inference to reduce generation randomness.

[0050] In this embodiment of the invention, the clause compliance review results output by the large language model are subjected to structured parsing and confidence assessment to generate a structured clause review report, including: The compliance review results of the terms output by the large language model are subjected to format verification and structured parsing to extract fields such as compliance judgment result, risk level, violation basis clause and / or modification suggestion, and obtain structured parsing results; Simultaneously, a multi-factor weighted model was used to obtain confidence assessment; Based on the structured analysis results and confidence assessment, a structured clause review report is generated.

[0051] Optionally, the confidence level assessment uses a multi-factor weighted model, calculated using the following formula: in, Set to the comprehensive matching score MatchScore, which reflects the semantic alignment between the rule and the fact, and the value range is [0, 1]. This represents the proportion of leaf nodes in the total number of leaf nodes of all rule conditions in the target compliance rule set that are covered (i.e., successfully matched) by the clause atomic facts. It reflects the completeness of the coverage of the rule conditions by the clause facts, and its value ranges from [0, 1]. The bidirectional alignment consistency score is represented by the following formula: The smaller the difference in matching degree between the rule fact and the fact rule, the higher the score, with a value range of [0, 1]. LLMConfidence represents the cumulative probability value of the large language model corresponding to its generated answer in the softmax output layer, that is, the overall confidence of the model in its generated result. After normalization, the value is between [0, 1], and it is directly obtained from the output layer during model inference. The default values ​​of the weights are respectively =0.30, =0.25, =0.25, =0.20, the sum of the four is 1.0.

[0052] The analyzed and evaluated results are organized according to a pre-set review report template to generate a structured clause review report containing fields such as the original clause text, compliance judgment, risk level, applicable rules (including specific clause number and version number), modification suggestions, and confidence level (including confidence level value and grading labels: high / medium / low, with grading standards as follows: ≥0.85 for high, ≥0.60 and <0.85 for medium, and <0.60 for low).

[0053] In this embodiment of the invention, based on the hierarchical structure of the contract text to be reviewed, the clause review report is constructed into a hierarchical overall contract compliance review report, including: The structured clause review report corresponding to the atomic clause is organized according to the hierarchical structure in the contract text to be reviewed, generating a hierarchical overall contract compliance review report.

[0054] For example, the review reports of all atomic clauses are organized according to the clause hierarchy to generate a hierarchical overall contract compliance review report. The overall compliance review report includes: contract overview information (contract name, contract type, review time, snapshot of the knowledge graph rule version used), review details of each clause (displayed one by one according to the clause hierarchy tree, each clause is expanded to include a step-by-step mapping from rule conditions, atomic facts, and the original text of the clause), risk summary statistics (classified by risk level to count the number of high-risk clauses, medium-risk clauses, low-risk clauses, clauses requiring further review, clauses awaiting manual confirmation, and the total number of clauses reviewed), and one or more of the following overall compliance conclusions.

[0055] Optionally, the overall compliance review report also includes a visual risk heatmap, arranging all atomic clauses of the contract hierarchically in a two-dimensional plane. Each atomic clause is represented by a rectangular color block, with the RGB color values ​​determined according to the risk level: high risk is RGB(255, 50, 50) red, medium risk is RGB(255, 165, 0) orange, low risk is RGB(50, 200, 50) green, requiring further review is RGB(255, 255, 0) yellow, and awaiting manual confirmation is RGB(200, 200, 200) gray. The clause number is marked in the upper left corner of the color block, and the confidence percentage value is displayed in white text inside the color block. Reviewers can intuitively grasp the overall risk distribution of the contract through this heatmap.

[0056] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses cross-validation in two directions. When the matching degree between the rule and the fact is high but the matching degree between the fact and the rule is low, a deep search is triggered. When the matching degree between the fact and the rule is high but the matching degree between the rule and the fact is low, it is marked as non-compliant. This effectively avoids misjudgment caused by rule omissions or fact extraction bias in one-way search.

[0057] 2. The quadruple of the present invention completely preserves constraint information such as time, quantity, method, and condition through independent segments, and performs numerical comparison rather than semantic similarity comparison for NUMERIC type constraints (such as direct numerical judgment of "60 ≤ 30"), thereby achieving accurate matching of rule conditions with numerical boundaries and solving the technical defect in the prior art that semantic similarity cannot handle numerical comparison.

[0058] 3. This invention breaks down contracts into atomic clause levels, and further breaks down each atomic clause into atomic fact quadruples, using atomic facts as the smallest matching unit, so that the review results can be accurate to the detailed clause provisions, rather than generally determining that the clause is non-compliant.

[0059] 4. This invention does not rely on large language models for fact extraction and rule matching. All factual content is provided by structured processing. The large language model only performs grammatical paraphrasing and does not create new values ​​or time limits on its own, thus fundamentally eliminating the risk of large language models fabricating legal basis or values.

[0060] 5. This invention automates the entire process from contract parsing, clause decomposition, fact extraction, two-way alignment matching to review report generation, significantly improving review efficiency compared to purely manual review. Based on a unified knowledge graph, a unified two-way alignment matching algorithm, and a unified confidence calculation model, it ensures the consistency and repeatability of review results across different contracts and batches.

[0061] like Figure 2 As shown, this embodiment of the invention also provides an intelligent review device for legally compliant contract terms, comprising: The contract parsing module 201 is used to obtain the contract text to be reviewed, and to perform structured parsing on the contract text to be reviewed in order to extract the hierarchical structure and atomic clauses in the contract text to be reviewed, and to perform semantic vectorization processing on each atomic clause to obtain the clause semantic vector corresponding to each atomic clause. The graph construction module 202 is used to construct a multi-dimensional legal compliance knowledge graph. The multi-dimensional legal compliance knowledge graph includes legal entities, compliance rule entities, and the relationships between these entities. Each compliance rule is represented in a structured form as a rule condition tree, where the internal nodes of the rule condition tree are logical operators and the leaf nodes are atomic conditions. Rule matching module 203 is used to split the atomic clause into atomic facts, and use the atomic facts as the smallest matching unit to perform bidirectional alignment compliance rule matching on the multi-dimensional legal compliance knowledge graph to obtain the target compliance rule set corresponding to the atomic clause; The compliance review module 204 is used to construct review prompt words based on the target compliance rule set, input the review prompt words into a preset large language model, and obtain the clause compliance review results output by the large language model; The report output module 205 is used to perform structured parsing and confidence assessment on the clause compliance review results output by the large language model, generate a structured clause review report, and construct the clause review report into a hierarchical overall contract compliance review report based on the hierarchical structure in the contract text to be reviewed.

[0062] The intelligent review device for legally compliant contract terms provided in this embodiment of the invention can execute the method described in any embodiment, and the principle and beneficial effects are similar, so they will not be described again.

[0063] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0064] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, electronic devices, and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0065] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0066] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0067] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0068] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0069] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for intelligent review of legally compliant contract terms, characterized in that, include: Obtain the contract text to be reviewed, and perform structured parsing on the contract text to extract the hierarchical structure and atomic clauses in the contract text to be reviewed. Then, perform semantic vectorization processing on each atomic clause to obtain the clause semantic vector corresponding to each atomic clause. Construct a multi-dimensional legal compliance knowledge graph; the multi-dimensional legal compliance knowledge graph includes legal and regulatory entities, compliance rule entities, and the relationships between these entities; each compliance rule is represented in a structured form as a rule condition tree, where the internal nodes of the rule condition tree are logical operators, and the leaf nodes are atomic conditions; The atomic clauses are broken down into atomic facts, and the atomic facts are used as the smallest matching units to perform bidirectional alignment compliance rule matching on the multi-dimensional legal compliance knowledge graph to obtain the target compliance rule set corresponding to the atomic clauses. Based on the target compliance rule set, review prompt words are constructed, and the review prompt words are input into a preset large language model to obtain the clause compliance review results output by the large language model; The clause compliance review results output by the large language model are subjected to structured parsing and confidence assessment to generate a structured clause review report. Based on the hierarchical structure in the contract text to be reviewed, the clause review report is constructed into a hierarchical overall contract compliance review report.

2. The intelligent review method for legally compliant contract terms according to claim 1, characterized in that, Obtain the contract text to be reviewed and perform structured parsing on the contract text to extract the hierarchical structure and atomic clauses, including: Obtain the contract text to be reviewed, and perform a first preprocessing on the contract text to be reviewed to obtain the contract text to be reviewed after the first preprocessing. Using the number of indentation characters, numbering pattern, relative font size, and type of the first line starting character of each paragraph in the contract text to be reviewed after the first preprocessing as input features, a Transformer-based sequence labeling clause structure recognition model is used to obtain the hierarchical labels of each paragraph. A hierarchical relationship tree of clauses is constructed based on the hierarchical tags, and the text content under each level in the hierarchical relationship tree is delineated to determine the atomic clauses; the atomic clause refers to the lowest level complete semantic unit in the hierarchical relationship tree that does not contain any child clause numbers.

3. The intelligent review method for legally compliant contract terms according to claim 1, characterized in that, Each atomic clause is semantically vectorized to obtain a clause semantic vector corresponding to each atomic clause, including: The atomic clauses are subjected to a second preprocessing to obtain atomic clauses after the second preprocessing. A pre-trained legal domain semantic encoding model is used to encode the atomic clauses after the second preprocessing to generate a fixed-dimensional semantic vector, thereby obtaining the clause semantic vector corresponding to the atomic clause. Iterate through each atomic clause to obtain the clause semantic vector corresponding to each atomic clause.

4. The intelligent review method for legally compliant contract terms according to claim 1, characterized in that, Constructing a multi-dimensional legal compliance knowledge graph, including: Collect legal and regulatory documents related to the contract, and perform entity identification and relationship extraction on the legal and regulatory documents to extract legal and regulatory clauses and compliance rule entities, as well as the relationships between the entities; each compliance rule is represented in a structured form as a rule condition tree, where the internal nodes of the rule condition tree are logical operators and the leaf nodes are atomic conditions; The extracted entities and relationships are stored in the Neo4j graph database as triples to construct a multi-dimensional legal compliance knowledge graph.

5. The intelligent review method for legally compliant contract terms according to claim 1, characterized in that, The atomic facts are represented as quadruples, specifically including: ； in, The fact type is determined by a standardized type selected from a predefined fact type dictionary. It represents the core legal subject or object involved in the atomic fact, and its value is the entity name after being mapped through legal terminology standardization. The core predicates representing atomic facts are taken from standard predicates selected from a predefined legal predicate dictionary; This represents the set of constraints on the atomic fact, where each constraint is a constraint attribute, a constraint operator, or a constraint value.

6. The intelligent review method for legally compliant contract terms according to claim 1, characterized in that, Using the atomic facts as the smallest matching unit, bidirectional alignment compliance rule matching is performed on the multi-dimensional legal compliance knowledge graph to obtain the target compliance rule set corresponding to the atomic clauses, including: For each compliance rule in the multi-dimensional legal compliance knowledge graph, extract all leaf nodes in the condition tree corresponding to the compliance rule, and in the atomic conditions corresponding to each leaf node, search for whether there is a matching item in all atomic facts of the atomic clause, obtain the matching result, and generate the rule fact matching degree based on the matching result; For each atomic fact of the atomic clause, a search is performed in all rule condition leaf nodes of the multi-dimensional legal compliance knowledge graph to obtain candidate rules, and atomic condition matching is performed on the candidate rules to generate fact rule matching degree; Based on the rule fact matching degree and the fact rule matching degree, a comprehensive matching score is calculated, and the target compliance rule set is determined based on the comprehensive matching score.

7. The intelligent review method for legally compliant contract terms according to claim 6, characterized in that, Based on the target compliance rule set, review prompts are constructed, and these prompts are input into a preset large language model to obtain the clause compliance review results output by the large language model, including: The original text of the atomic clause, the atomic facts, the target compliance rule set, the two-way alignment matching result summary, and the preset review instruction template are combined to generate review prompt words; Input the review prompts into a preset large language model and obtain the clause compliance review results output by the large language model.

8. The intelligent review method for legally compliant contract terms according to claim 7, characterized in that, The clause compliance review results output by the large language model are subjected to structured parsing and confidence assessment to generate a structured clause review report, including: The compliance review results of the terms output by the large language model are subjected to format verification and structured parsing to extract fields such as compliance judgment result, risk level, violation basis clause and / or modification suggestion, and obtain structured parsing results; Simultaneously, a multi-factor weighted model was used to obtain confidence assessment; Based on the structured analysis results and confidence assessment, a structured clause review report is generated.

9. The intelligent review method for legally compliant contract terms according to claim 8, characterized in that, Based on the hierarchical structure of the contract text to be reviewed, the clause review report is constructed into a hierarchical overall contract compliance review report, including: The structured clause review report corresponding to the atomic clause is organized according to the hierarchical structure in the contract text to be reviewed, generating a hierarchical overall contract compliance review report.

10. An intelligent review device for legally compliant contract terms, characterized in that, include: The contract parsing module is used to obtain the contract text to be reviewed, and to perform structured parsing on the contract text to be reviewed in order to extract the hierarchical structure and atomic clauses in the contract text to be reviewed, and to perform semantic vectorization processing on each atomic clause to obtain the clause semantic vector corresponding to each atomic clause. The graph construction module is used to construct a multi-dimensional legal compliance knowledge graph. The multi-dimensional legal compliance knowledge graph includes legal and regulatory entities, compliance rule entities, and the relationships between these entities. Each compliance rule is represented in a structured form as a rule condition tree, where the internal nodes of the rule condition tree are logical operators and the leaf nodes are atomic conditions. The rule matching module is used to break down the atomic clauses into atomic facts, and use the atomic facts as the smallest matching unit to perform bidirectional alignment compliance rule matching on the multi-dimensional legal compliance knowledge graph to obtain the target compliance rule set corresponding to the atomic clauses; The compliance review module is used to construct review prompts based on the target compliance rule set, input the review prompts into a preset large language model, and obtain the clause compliance review results output by the large language model. The report output module is used to perform structured parsing and confidence assessment on the clause compliance review results output by the large language model, generate a structured clause review report, and construct the clause review report into a hierarchical overall contract compliance review report based on the hierarchical structure in the contract text to be reviewed.

Images