Intelligent review method and system for contract compliance review

Abstract

This invention relates to the field of document review technology and proposes an intelligent review method and system for contract compliance review. The steps include: preloading compliance regulations into a local node machine to build an initial knowledge base; receiving contract documents to be reviewed and review instructions uploaded by the user; performing fine-grained structural deconstruction on the contract documents to be reviewed, constructing a nested contract knowledge graph, mapping the subjects, clauses, constraints, and exceptions in the contract to entities and relationships; converting the review instructions into a graph query language, performing graph pattern matching retrieval in the nested contract knowledge graph, and extracting subgraphs related to the review instructions; inputting the extracted subgraphs and the review instructions into a large language model, guiding the large language model to perform step-by-step reasoning along the logical paths defined in the subgraphs, and generating review opinions containing compliance judgments and supporting evidence. This invention significantly improves the interpretability and accuracy of review results and effectively suppresses the illusion phenomenon of large models.

Description

Technical Field

[0001] This invention relates to the field of document review technology, and more specifically, to an intelligent review method and system for contract compliance review. Background Technology

[0002] Contract compliance review is a core component of corporate risk management and legal work, and its accuracy directly impacts the company's operational security and profit protection. Statistics show that in 2023, the average value of disputes in the global construction industry reached US$42.8 million, with an average resolution cycle of 13 months. Misunderstandings in contract terms, breaches of performance obligations, and errors or omissions in documentation were listed as major contributing factors. Traditional contract review heavily relies on manual experience, requiring reviewers to meticulously compare clauses in complex contract texts against laws, regulations, industry standards, and internal compliance guidelines. This process is not only time-consuming and labor-intensive but also prone to overlooking risks due to insufficient business acumen or incomplete framework thinking on the part of the reviewer.

[0003] With the rapid development of natural language processing technology, automated contract analysis methods based on rule engines and machine learning models are gradually being applied in practice. Empirical studies show that NLP-assisted review can improve the accuracy of contract classification to 83% while reducing review time by 76%. However, existing technologies still face three fundamental limitations: First, rule-based and model-based methods are mostly limited to text classification or fragmented label prediction, making it difficult to capture the complex nested logic and deep semantic relationships between contract clauses. For example, expressions containing multiple layers of constraints, such as "If the contractor fails to submit a report within X days, payment will be postponed for Y days, except in cases of force majeure," can only be analyzed superficially by traditional models. Second, while large language models demonstrate strong semantic understanding and generation capabilities in contract review applications, their "black box" nature and lack of domain knowledge constraints in their generation mechanisms lead to frequent factual illusions. Studies have shown that DeepSeek-R1's model illusion rate is as high as 14.3%, and the model may cite incorrect legal provisions or fabricate non-existent case numbers, seriously threatening the reliability of review results. Third, existing solutions lack effective knowledge representation and logical verification mechanisms, failing to transform explicit knowledge such as accumulated corporate compliance regulations and historical review experience into a structured knowledge source that can be reasoned about. Review conclusions are often difficult to trace back to their judgment basis.

[0004] To address the aforementioned issues, the industry has begun exploring the integration of knowledge graphs and large-scale models. Related research has confirmed that the Graph Retrieval Enhanced Generation (GraphRAG) method based on nested contract knowledge graphs (NCKG) can significantly improve the accuracy and interpretability of review. Meanwhile, at the national level, the development of standards such as the "General Principles for Artificial Intelligence-Assisted Drafting and Review of Contract Services" is being promoted, requiring strengthened knowledge base construction and risk classification and alert mechanisms. Therefore, how to construct an intelligent review method that can deeply understand the complex semantic structure of contracts and achieve logical traceability has become a key technical challenge urgently needing to be solved in this field. Summary of the Invention

[0005] This invention proposes an intelligent review method for contract compliance review, comprising: Preload compliance documents into the local node machine to build an initial knowledge base; Receive contract documents and review instructions uploaded by users; The contract documents to be reviewed are deconstructed in a fine-grained structure to construct a nested contract knowledge graph. The nested contract knowledge graph adopts the RDF-star data model to map the subjects, clauses, constraints and exceptions in the contract into entities and relationships. The review instruction is converted into a graph query language, and graph pattern matching retrieval is performed in the nested contract knowledge graph to extract subgraphs related to the review instruction. The extracted subgraph and the review instructions are input into the large language model, which is then guided to reason step by step along the logical path defined in the subgraph to generate review opinions that include compliance judgments and supporting evidence.

[0006] Furthermore, before constructing the nested contract knowledge graph, the following steps are also included: The layout of the contract document to be reviewed is analyzed and the table is recognized by OCR. The table, signature area and text are separated, the table structure and cell content are extracted, and the extracted data is mapped to the corresponding node of the knowledge graph.

[0007] Furthermore, the graph pattern matching retrieval specifically includes: Automatically retrieve subgraphs in the pending contract that have a path related to compliance rule nodes in the knowledge base, and extract complete reasoning subgraphs including triggering conditions, defaulting party, compensation calculation method, and disclaimer clauses.

[0008] Furthermore, the review opinions are marked with different colors in the contract text to indicate the severity of compliance issues, and the specific clauses on which the opinions are based and their corresponding path are simultaneously output.

[0009] Furthermore, it also includes the steps of building a knowledge graph visualization interface: Perform multimodal parsing on user-uploaded knowledge base documents to extract text, layout, and image features; Entities and relations are jointly extracted based on a cross-modal contrastive learning model, and triples are output. Import the triples into a graph database and use a relational graph neural network to complete the implicit relations; Generate a relationship graph based on force-directed layout, and divide the topic clusters using a community detection algorithm, identifying them with different colors; Generate a word cloud view, where the font size of keywords in the word cloud is proportional to the connectivity of the entities in the graph; It provides interactive operations, allowing users to expand the neighbor subgraph or locate the corresponding node in the word cloud by clicking on an entity.

[0010] Furthermore, the multimodal parsing is based on the LayoutLMv3 architecture, which simultaneously processes text, image, and layout information, outputs a fused representation of each token, and identifies document structure elements.

[0011] Furthermore, in the pre-training stage, the cross-modal contrastive learning model constructs positive samples using entity pairs related to the same document layout, and constructs negative samples using entity pairs randomly across documents or across unrelated regions, and optimizes the encoder through a contrastive loss function.

[0012] Furthermore, it also includes multiple rounds of interaction and feedback processing steps: Receive user feedback text regarding review comments, and parse their intent type and focus entity; Retrieve dynamically expanded contextual subgraphs from the knowledge graph, centered on the focus entity; The user feedback, the original text fragments of the focus entity, and the context subgraph are concatenated into prompt words, which are then input into a large language model to generate multiple candidate explanations or suggested texts. The logical consistency of each candidate answer is evaluated using a reverse reasoning framework. After screening and correction, the final answer is output.

[0013] Furthermore, the evaluation of the logical consistency of candidate answers using the reverse reasoning framework specifically includes: The candidate answer text is transformed into a first-order logical proposition, where the predicate corresponds to the relation type in the knowledge graph and the constant corresponds to the entity ID. A Prolog-based inference engine is used to check whether the implicit preconditions of candidate answers can be derived from the knowledge graph. For premises that cannot be directly supported, the reverse reasoning module is activated to find missing intermediate facts and complete the reasoning chain.

[0014] On the other hand, to achieve the above objectives, this invention proposes an intelligent review system for contract compliance review, comprising: The knowledge base building module is used to preload compliance regulations and supports users to upload custom files; The contract parsing module is used to deconstruct the contract documents to be reviewed and build a nested contract knowledge graph. The instruction parsing and retrieval module is used to convert user review instructions into graph query language and perform graph pattern matching retrieval in the knowledge graph to extract relevant subgraphs; The reasoning generation module is used to input the extracted subgraphs and instructions into the large language model to generate review comments; The visualization module is used to perform multimodal parsing, entity relation extraction, and knowledge graph construction on knowledge base documents, and to generate interactive relational graphs and word clouds; The feedback processing module is used to receive user feedback and generate and filter explanatory or suggestive text through a reverse reasoning framework.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention constructs a deep semantic understanding and logical verification framework based on nested contract knowledge graphs, cross-modal comparative learning, and reverse reasoning. By employing the RDF-star data model for fine-grained structural deconstruction of contract texts and integrating layout analysis and cross-modal comparative learning using the LayoutLMv3 architecture, the system accurately extracts entities and their complex nested relationships from the document, transforming unstructured contracts into a knowledge network rich in logical paths. Based on this, graph pattern matching retrieval and structure-aware reasoning from a large model generate preliminary review opinions. Then, a logical verification mechanism based on reverse reasoning is introduced, transforming candidate answers generated by the large model into logical propositions and comparing their consistency with the knowledge graph. Hypothetical reasoning fills in missing intermediate facts, selecting the optimal explanation with the lowest cost. This invention significantly improves the interpretability and accuracy of review results, effectively suppresses the illusion phenomenon of large models, and enables users to intuitively understand knowledge relationships and dynamically improve review strategies through visualized knowledge graphs and multi-round feedback optimization loops. Ultimately, it achieves an intelligent leap from keyword matching to logic-driven and traceable contract compliance review. Attached Figure Description

[0016] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. In the drawings: Figure 1 This is a schematic diagram of the overall process of the intelligent review method for contract compliance review of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] This embodiment proposes an intelligent review method for contract compliance review, such as... Figure 1 As shown, it includes: Basic compliance documents, including national laws and regulations, industry standards, and internally common compliance management documents, are pre-loaded on the local node machine. These documents serve as the foundation of the initial knowledge base for subsequent reviews. In practice, users can upload custom regulations, policy documents, project-specific compliance requirements, or special terms and specifications provided by clients through the system interface, depending on the specific contract type or review task.

[0019] The system displays a progress bar in real time during the upload process and updates the knowledge base status after the file upload is complete. Users can view the file names in the knowledge base and the learning progress of the large model on the file content at any time, ensuring the accuracy and timeliness of the knowledge base.

[0020] Once the contract or other text documents to be reviewed are uploaded, the user enters a review instruction in the dialog box, requesting a compliance check of the contract terms in accordance with the uploaded regulations. After receiving the instruction, the system combines the document content in the knowledge base with the understanding capabilities of the large model to compare and analyze the contract text line by line.

[0021] Specifically, in the large-scale model analysis process, this embodiment employs a GraphRAG (Graph Retrieval Enhanced Generation) reasoning mechanism based on a nested contract knowledge graph (NCKG). This mechanism achieves accurate and traceable intelligent review by modeling and retrieving deep semantic relationships between contract clauses. Its specific implementation includes the following three levels: Layer 1: Contract semantic modeling based on nested knowledge graphs Upon receiving the contract text to be reviewed, the system first activates its parsing engine to perform fine-grained structural deconstruction of the contract. Unlike simple text segmentation, the system employs a Nested Contract Knowledge Graph (NCKG) construction mechanism to address complex contract logic. This mechanism utilizes the RDF-star data model to map each subject, clause, constraint, and exception in the contract to entities and relationships in the graph. It excels at handling complex logic containing multiple layers of nested conditions, such as "If the contractor fails to submit a report within X days, payment will be postponed by Y days, except in cases of force majeure." The system parses the contract in real time and aligns it with the pre-loaded compliance knowledge graph, transforming the linear contract text into a semantically rich knowledge network composed of thousands of nodes and edges, laying the data foundation for subsequent deep reasoning.

[0022] Second layer: Instruction-driven graph pattern matching and subgraph retrieval User-issued review instructions are translated into graph query language within the system, and graph pattern matching is performed on the constructed knowledge graph. The system accurately retrieves subgraph structures in the graph that have specific relationship paths with the "liability for breach of contract" entity node. Specifically, the system automatically retrieves the association path between the "liability penalty ratio" node in the contract under review and the "risk control guidance" node in the knowledge base, extracting a complete reasoning subgraph that includes triggering conditions, the breaching party, compensation calculation methods, and disclaimers. This retrieval method ensures that the contextual information obtained by the large model contains a knowledge structure with clear logical dependencies, fundamentally solving the problem of information misuse caused by semantic similarity but logical misalignment in traditional RAGs.

[0023] Layer 3: Generative Reasoning Enhanced with Structure Awareness After acquiring a precise structured subgraph, the system inputs the subgraph along with the user's original instructions into the core of the large model's reasoning. At this stage, the large model no longer deals with lengthy original contract texts but interprets them based on a "knowledge graph context." The system guides the model to reason step-by-step along the paths defined in the graph. Specifically, when reviewing the "price adjustment clause," the model strictly adheres to the logical chain of "price adjustment mechanism, market price index, and data source" in the graph for verification. Simultaneously, thanks to the anchoring relationship between graph nodes and the original contract text, when generating a review opinion that "the price benchmark date definition in this clause is ambiguous, potentially leading to settlement risks," the model can simultaneously output the specific clause location and graph path upon which this opinion is based. This process ensures that the large model's output is not only controllable and logically rigorous, but also that each conclusion has a traceable chain of evidence, effectively suppressing illusions during the generation process.

[0024] In addition, to support the accuracy of the above-mentioned graph construction and retrieval, the system integrates high-precision layout analysis and table OCR technology in the preprocessing stage to accurately separate tables, signature areas, headers and footers from the main text in PDF or scanned contracts. This ensures that data extracted from complex tables such as "performance deadlines" or "price lists" can be correctly mapped to the corresponding nodes in the knowledge graph, providing reliable data support for the GraphRAG process.

[0025] After the review is completed, the system generates a detailed review report and uses different colors to indicate the severity of various compliance issues in the contract text. For example, red indicates significant compliance risks, yellow indicates matters requiring attention, and green indicates compliance is correct. Users can choose to download the reviewed contract documents or view the annotation results directly in the interface.

[0026] In another implementation, the system includes a built-in knowledge base management module that supports categorized management of knowledge bases across different business areas. For contract compliance review scenarios, the system allows users to select one or more knowledge base files as the basis for review. Specifically, when reviewing procurement contracts, users can simultaneously select both the "Procurement Compliance Library" and the "Data Security Library," and the system will perform a comprehensive analysis combining regulations from multiple knowledge bases.

[0027] Furthermore, to improve the accuracy of retrieval and analysis, the system employs graph database technology to model the relationships within the knowledge base content, constructing a knowledge graph. After files are uploaded and the database is built, the system generates a keyword relationship graph or word cloud. Users can intuitively browse the relationships between file content through a visual interface, facilitating understanding of the knowledge structure and optimization of review strategies. The specific implementation of this process includes the following steps: Step 1: Document Multimodal Analysis and Structured Preprocessing The system first performs multimodal parsing on user-uploaded PDF, Word, or scanned documents. The parsing engine is based on the LayoutLMv3 architecture, which can simultaneously process text, images, and layout information. For each page of the document, the system extracts the text content, character-level bounding box coordinates, and segments the document image into visual regions. The model concatenates text tokens with their corresponding 2D position embeddings (x0, y0, x1, y1) and image region features, outputting a fused representation of each token. The parsing results reconstruct the document's reading order and identify structured elements such as titles, paragraphs, lists, and table cells. For tables, the system further applies a table structure recognition algorithm to extract cell content and their row and column relationships as structured data, providing high-quality input for subsequent relationship extraction.

[0028] Step 2: Joint entity and relation extraction through cross-modal contrastive learning After obtaining the multimodal fusion representation, the system performs a joint task of entity recognition and relation extraction. Specifically, a pre-trained cross-modal contrastive learning model is used, which utilizes the principle that related entity pairs within the same document should be close to each other in the embedding space, while unrelated entity pairs should be far apart. During the pre-training phase, the model constructs positive sample pairs from the same page unit and random negative sample pairs across documents or unrelated regions. The encoder is optimized through a contrastive loss function, enabling the model to not only understand textual semantics but also perceive the relationships implied by the page structure.

[0029] During the fine-tuning phase, the model receives labeled entity and relation data, adds a classification layer on top of the fused representation, and outputs entity boundaries, entity types, and relation types between entities. Specifically, when the document contains the entities "penalty rate: 20%" and "overdue days: 30 days," the model leverages their position in the same row in a table to more easily extract the "triggering condition" relationship between "penalty rate" and "overdue days." The final output is a series of triples (entity 1, relation, entity 2), along with the location information of each entity in the original text.

[0030] Step 3: Knowledge Graph Construction and Relationship Completion The extracted triples are imported in batches into the graph database Neo4j, where each entity becomes a node and relations become directed edges. To ensure the integrity of the knowledge base, the system further employs a Relational Graph Neural Network (R-GCN) for implicit relation completion. The R-GCN model, based on the existing graph structure, updates node representations by aggregating neighbor node information and predicts potentially missing edges. The completion process infers implicit associations using existing explicit relations in the documents. The completed graph is richer, providing more comprehensive association information for subsequent retrieval.

[0031] Step 4: Generation of Visual Relationship Diagram and Word Cloud The system provides users with an interactive visual interface, employing a graph rendering engine based on force-directed layout. When generating the relationship graph, the system first runs the Louvain community discovery algorithm, dividing the graph into multiple topic clusters and marking them with different colors. Node size is determined by importance calculated using the PageRank algorithm, while edge thickness is dynamically mapped based on the strength of relationships between entities. Users can search for or click on a focused entity, and the system will expand the subgraph of that entity's first to third-degree neighbors in real time, avoiding an overly complex overall graph. The relationship graph supports zooming, dragging, and hovering for detailed viewing. Simultaneously, the system generates a word cloud view, where each keyword corresponds to an entity, and its font size is proportional to the entity's connectivity in the graph, helping users quickly grasp the core themes of the knowledge base. Clicking on a keyword in the word cloud automatically locates the corresponding node in the relationship graph.

[0032] Step 5: Optimize User Interaction and Feedback In the visual interface, users can manually adjust the relationship diagram, deleting incorrect associations and merging duplicate entities. User operation records are fed back to the backend for incremental updates to the knowledge base and triggering model fine-tuning, forming a continuous optimization loop. In this way, users can intuitively browse the knowledge structure and participate in the improvement of the knowledge base, significantly enhancing the accuracy and adaptability of the review strategy.

[0033] In another implementation, the system supports multi-round interaction and dynamic adjustment. After the initial review results are returned, users can raise further questions regarding specific marked areas. Based on user feedback, combined with knowledge base content and large model generation capabilities, the system provides further explanations or suggested text. Users can continuously optimize the contract content based on system feedback until compliance requirements are met. This process supports multiple iterations to ensure the contract text reaches an optimal compliance state before submission. The specific implementation of this process includes the following steps: Step 1: Fine-grained analysis and intent alignment of user feedback The system performs multi-dimensional parsing of user-input feedback text. The parsing module employs a lightweight yet high-precision intent recognition model based on a pre-trained RoBERTa architecture, fine-tuned on contract review data. The model outputs two key pieces of information: the intent type and the focus entity, i.e., the contract clause or specific node in the knowledge graph referred to by the user. The system aligns the user-mentioned clause description with the corresponding node in the knowledge graph using entity linking technology. If the user does not explicitly specify a focus, the system defaults to the clause the user hovered over or clicked on in the most recent review result within the current dialogue context.

[0034] Step 2: Context subgraph retrieval based on the focus entity The system retrieves relevant contextual subgraphs from the contract compliance knowledge graph, centered on the focal entity. The retrieval strategy employs a dynamic expansion mechanism: first, it extracts the first-degree neighbor nodes and relation edges of the focal entity to form a basic subgraph; then, it calculates the information gain index for each neighbor entity in the basic subgraph—this index comprehensively considers the neighbor's degree centrality, the strength of its relationship with the focal entity, and whether it involves compliance rule nodes. If the information gain exceeds a threshold, the second-degree relationship of that neighbor is also included in the subgraph. The final generated subgraph contains the focal entity, key neighbor entities, relevant compliance rule nodes, and the path relationships between them. This subgraph is temporarily stored as a list of RDF triples, while retaining the anchor point of each entity's position in the original document for traceability.

[0035] Step 3: Generate candidate explanations / suggested texts from the large model The system concatenates user feedback, original text fragments of the focus entity, and the knowledge graph subgraph retrieved in the previous step into structured prompts, which are then input into the large language model. The prompts are designed to guide the model in "graph-based logical reasoning generation," explicitly requiring the model to generate answers to user questions based on relationships within the graph. The large model generates multiple candidate answers in parallel, each containing a natural language explanation or suggestion, and annotating the knowledge graph node ID or relationship name after key arguments for subsequent verification. This step leverages the semantic understanding and generation capabilities of the large model to ensure the fluency and apparent plausibility of the answers.

[0036] Step 4: Candidate answer selection and logical correction based on reverse reasoning This step is the most innovative and core step in this embodiment. The system treats each candidate answer as an explanatory hypothesis about the user's observation and uses a reverse reasoning framework to evaluate its logical consistency. The specific implementation is as follows: The candidate answer text is transformed into a set of logical propositions. This is accomplished through a finely tuned text-to-logic transformer that breaks down natural language statements into atomic facts and inference rules, such as "if A and B then C". The transformed result is represented as a series of first-order logical clauses, where predicates correspond to relation types in the knowledge graph and constants correspond to entity IDs.

[0037] The transformed logical propositions are compared with facts in the knowledge graph. The knowledge graph, serving as a background knowledge base, provides a large number of confirmed fact triples. The system uses a Prolog-based inference engine to check whether all the implicit preconditions of the candidate answers can be derived from the knowledge graph. Specifically, if a candidate answer claims that "the penalty for breach of contract is too high because the contract lacks a maximum liability clause," the inference engine will search the graph for the existence of the fact "Contract C_lacking_ Maximum Liability Clause," and whether the causal chain between this fact and "the penalty for breach of contract is too high" is valid.

[0038] If some premises of a candidate answer cannot be directly supported in the knowledge graph, the system activates the reverse reasoning module. This module attempts to find missing intermediate facts to complete the entire reasoning chain. For example, the graph might contain "Contract C_refers to_Standard Contract Template S" and "Standard Contract Template S_includes_Liquidity Penalty Cap Clause," and reverse reasoning can deduce the intermediate fact that "Contract C_lacks_Liquidity Penalty Cap Clause." The system calculates the "explanation cost" for each candidate answer, which is determined by a weighted average of the number of new assumptions, the reasonableness of the assumptions, and the length of the logical chain. The candidate answer with the lowest cost is selected as the best explanation.

[0039] For the selected best candidate answer, the system fine-tunes the original text based on intermediate facts added during the reasoning process, making the implicit reasoning steps explicit, and ensuring that the final output explanation or suggestion is completely consistent with the knowledge graph. If the explanation cost of all candidate answers exceeds a threshold, the system returns to the large model generation stage, requiring regeneration or prompting the user to supplement information.

[0040] Step 5: Generate a traceable final answer The system returns the best answer to the user after filtering and correction through reverse reasoning. In the returned text, the system annotates the entities and relationships in the knowledge graph upon which each key argument is based, using hyperlinks or footnotes. Users can click to view the specific nodes in the graph and the location of the original document. Simultaneously, the system logs the feedback and corrections from this interaction for subsequent model fine-tuning or graph updates, achieving closed-loop optimization.

[0041] Through the above steps, the explanations and suggestions provided by the system not only possess the natural language expression capabilities of a large model, but also undergo logical verification through a knowledge graph, ensuring the accuracy and interpretability of the content and effectively supporting users' in-depth analysis and decision-making in contract compliance review scenarios.

[0042] In another implementation, the system also supports automated archiving and report generation of contract review results. Upon completion of the review, the system automatically generates a compliance review report containing basic contract information, review time, review basis, a list of issues, and suggested modifications. The report is then stored locally or on a designated server, linked to the original contract documents. Users can retrieve historical review records via the system interface by time, contract type, or reviewer, facilitating subsequent auditing and traceability.

[0043] In another implementation, the system also features access control and operation log functionality. Different user roles (such as legal specialists, compliance managers, and project leaders) have different access permissions, such as uploading files, viewing review results, and modifying knowledge base content. The system records each user's actions, including uploading files, initiating reviews, and downloading reports, forming a complete operation log to ensure the security and traceability of system usage.

[0044] In another implementation, the system supports interface integration with external business systems. For example, after contract review is completed, the system can automatically push the review results and contract documents to the enterprise's contract management system or OA system, achieving seamless integration of contract compliance review and contract management processes. Users do not need to manually export and import, improving work efficiency and reducing the risk of data loss or mistransmission.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An intelligent review method for contract compliance review, characterized in that, include: Preload compliance documents into the local node machine to build an initial knowledge base; Receive contract documents and review instructions uploaded by users; The contract documents to be reviewed are deconstructed in a fine-grained structure to construct a nested contract knowledge graph. The nested contract knowledge graph adopts the RDF-star data model to map the subjects, clauses, constraints and exceptions in the contract into entities and relationships. The review instruction is converted into a graph query language, and graph pattern matching retrieval is performed in the nested contract knowledge graph to extract subgraphs related to the review instruction. The extracted subgraph and the review instructions are input into the large language model, which is then guided to reason step by step along the logical path defined in the subgraph to generate review opinions that include compliance judgments and supporting evidence.

2. The intelligent review method for contract compliance review according to claim 1, characterized in that, Before constructing the nested contract knowledge graph, the following steps are also required: The layout of the contract document to be reviewed is analyzed and the table is recognized by OCR. The table, signature area and text are separated, the table structure and cell content are extracted, and the extracted data is mapped to the corresponding node of the knowledge graph.

3. The intelligent review method for contract compliance review according to claim 1, characterized in that, The graph pattern matching retrieval specifically includes: Automatically retrieve subgraphs in the pending contract that have a path related to compliance rule nodes in the knowledge base, and extract complete reasoning subgraphs including triggering conditions, defaulting party, compensation calculation method, and disclaimer clauses.

4. The intelligent review method for contract compliance review according to claim 1, characterized in that, The review comments are marked with different colors in the contract text to indicate the severity of compliance issues, and the specific clauses on which the comments are based and their corresponding path are also displayed.

5. The intelligent review method for contract compliance review according to claim 1, characterized in that, It also includes the steps of building a knowledge graph visualization interface: Perform multimodal parsing on user-uploaded knowledge base documents to extract text, layout, and image features; Entities and relations are jointly extracted based on a cross-modal contrastive learning model, and triples are output. Import the triples into a graph database and use a relational graph neural network to complete the implicit relations; Generate a relationship graph based on force-directed layout, and divide the topic clusters using a community detection algorithm, identifying them with different colors; Generate a word cloud view, where the font size of keywords in the word cloud is proportional to the connectivity of the entities in the graph; It provides interactive operations, allowing users to expand the neighbor subgraph or locate the corresponding node in the word cloud by clicking on an entity.

6. The intelligent review method for contract compliance review according to claim 5, characterized in that, The multimodal parsing is based on the LayoutLMv3 architecture, which processes text, images and layout information simultaneously, outputs a fused representation of each token and identifies document structure elements.

7. The intelligent review method for contract compliance review according to claim 5, characterized in that, The cross-modal contrastive learning model constructs positive samples using entity pairs related to the same document layout during the pre-training phase, and constructs negative samples using entity pairs randomly across documents or across unrelated regions. The encoder is then optimized using a contrastive loss function.

8. The intelligent review method for contract compliance review according to claim 1, characterized in that, It also includes multiple rounds of interaction and feedback processing steps: Receive user feedback text regarding review comments, and parse their intent type and focus entity; Retrieve dynamically expanded contextual subgraphs from the knowledge graph, centered on the focus entity; The user feedback, the original text fragments of the focus entity, and the context subgraph are concatenated into prompt words, which are then input into a large language model to generate multiple candidate explanations or suggested texts. The logical consistency of each candidate answer is evaluated using a reverse reasoning framework. After screening and correction, the final answer is output.

9. The intelligent review method for contract compliance review according to claim 8, characterized in that, The evaluation of the logical consistency of candidate answers using the reverse reasoning framework specifically includes: The candidate answer text is transformed into a first-order logical proposition, where the predicate corresponds to the relation type in the knowledge graph and the constant corresponds to the entity ID. A Prolog-based inference engine is used to check whether the implicit preconditions of candidate answers can be derived from the knowledge graph. For premises that cannot be directly supported, the reverse reasoning module is activated to find missing intermediate facts and complete the reasoning chain.

10. An intelligent review system for contract compliance review, characterized in that, include: The knowledge base building module is used to preload compliance regulations and supports users to upload custom files; The contract parsing module is used to deconstruct the contract documents to be reviewed and build a nested contract knowledge graph. The instruction parsing and retrieval module is used to convert user review instructions into graph query language and perform graph pattern matching retrieval in the knowledge graph to extract relevant subgraphs; The reasoning generation module is used to input the extracted subgraphs and instructions into the large language model to generate review comments; The visualization module is used to perform multimodal parsing, entity relation extraction, and knowledge graph construction on knowledge base documents, and to generate interactive relational graphs and word clouds; The feedback processing module is used to receive user feedback and generate and filter explanatory or suggestive text through a reverse reasoning framework.

Images