Automatic contract auditing system and method based on large language model

Abstract

The invention discloses an automatic contract auditing system and method based on a large language model, and relates to the technical field of automatic contract auditing, the system comprises a document preprocessing module, an analysis scheduling module, an expert agent cluster module, a large model knowledge reserve module and an auditing report module; the document preprocessing module is used for performing multi-modal analysis processing on the uploaded contract file and extracting contract text data; the analysis scheduling module is used for carrying out semantic segmentation on the contract text data, generating auditing tasks and distributing the auditing tasks; the expert agent cluster module is used for scheduling the large model knowledge reserve module to perform contract auditing based on the auditing task; the large model knowledge reserve module is used for providing a plurality of large language models and a preset knowledge base; and the auditing report module is used for integrating the contract auditing results of the expert agent cluster module and generating a contract auditing report. Through cooperative work of all the modules, automation and intelligentization of contract auditing are achieved.

Description

Technical Field

[0001] This application relates to the field of automated contract review technology, and more specifically, to an automated contract review system and method based on a large language model. Background Technology

[0002] Contract review is a crucial step in business activities, involving multi-dimensional analysis including legal compliance, risk control, financial assessment, and technical feasibility. As business collaborations become increasingly complex and specialized, contract terms are becoming more refined, especially in technology contracts (such as R&D, transfer, and service contracts), which often contain numerous technical terms and complex clauses, placing higher demands on the depth and efficiency of review. Traditional manual review methods are no longer sufficient to meet the real-time, accurate, and consistent requirements of modern business contract review, necessitating the introduction of intelligent technologies for optimization.

[0003] Currently, contract review mainly relies on the following technical means: purely manual review: legal or business experts review the contract content line by line, and the quality and efficiency of the review depend entirely on the reviewer's personal experience, professional level and work status; rule-based automated tools: identify contract differences through keyword matching or template comparison; general document analysis platforms: use OCR and basic NLP technologies to extract text information and discover differences by comparing the contract to be reviewed with a preset standard template.

[0004] However, existing technologies have the following core problems: Insufficient semantic understanding: Existing tools cannot analyze the true meaning of clauses in conjunction with context, leading to misjudgment of risks; Lack of cross-domain collaboration: The review processes of legal, financial, and technical departments are isolated from each other, making it impossible to adapt to the special requirements of contracts in different industries. Simple and complex tasks are handled using a single processing mode, making it difficult to comprehensively assess complex risks. Summary of the Invention

[0005] To address at least one deficiency or improvement need in the existing technology, this invention provides an automated contract review system and method based on a large language model. This system solves the problems in the existing technology, such as the inability to analyze the true meaning of clauses in context, leading to misjudgment of risks, fragmented cross-domain collaborative review processes, difficulty in comprehensively assessing complex risks, and inability to adapt to the special requirements of contracts in different industries.

[0006] To achieve the above objectives, according to a first aspect of the present invention, an automated contract review system based on a large language model is provided, comprising: a document preprocessing module, a parsing and scheduling module, an expert agent cluster module, a large model knowledge reserve module, and a review report module; the document preprocessing module, the parsing and scheduling module, the expert agent cluster module, the large model knowledge reserve module, and the review report module are connected sequentially. The document preprocessing module is used to perform multimodal parsing on the uploaded contract files and extract the contract text data. The parsing and scheduling module is used to perform semantic segmentation on contract text data, generate review tasks, and assign review tasks. The expert intelligent agent cluster module is used for contract review based on the knowledge reserve module of the large model for review task scheduling; The large model knowledge reserve module is used to provide several large language models and preset knowledge bases; The review report module is used to integrate the contract review results from the expert intelligent agent cluster module and generate a contract review report.

[0007] In one possible implementation, the document preprocessing module also includes: The text recognition unit is used to extract contract text content using a document recognition model; The layout analysis unit is used to parse the layout structure of contract documents using graph neural networks; The text cleaning unit is used to perform standardized processing of contract text content based on the layout structure of the contract document to obtain contract text data.

[0008] In one possible implementation, the parsing and scheduling module also includes: The semantic domain segmenter uses a multi-label classification model to perform semantic segmentation on contract text data and outputs the domain labels and subtype labels of the preset clauses. The task scheduler generates review tasks based on domain labels and subtype labels, and dynamically optimizes the task allocation strategy based on reinforcement learning algorithms.

[0009] In one possible implementation, the expert agent cluster module also includes: The clause analysis unit is used to review legal clauses in contract text data based on review tasks. The risk assessment unit is used to review contract text data for compliance risks based on the review task. The compliance inspection unit is used to conduct textual compliance reviews of contract text data based on review tasks. The financial audit unit is used to conduct cost-benefit audits on contract text data based on audit tasks.

[0010] In one possible implementation, the large model knowledge reserve module also includes: Heterogeneous large model unit, containing multiple large language models of different sizes; The resource management unit allocates review requests to the corresponding large language models based on the review tasks. Domain knowledge base, which establishes a multimodal vector database based on preset laws and regulations.

[0011] According to a second aspect of the present invention, a method for automated contract review based on a large language model is also provided, comprising a contract automated review system based on a large language model as described in any of the preceding claims, including: The uploaded contract file is processed using a document preprocessing module to perform multimodal parsing and extract the contract text data. The contract text data is semantically segmented by the parsing and scheduling module to generate and assign review tasks. Contract review is performed using the expert intelligent agent cluster module based on the knowledge reserve module of the large model for review task scheduling. The large model knowledge reserve module provides several large language models and preset knowledge bases to support the review process; The contract review report is generated by integrating the contract review results from the expert intelligent agent cluster module into the review report module.

[0012] In one possible implementation, the uploaded contract file is subjected to multimodal parsing processing by a document preprocessing module to extract the contract text data, and the process also includes: The uploaded contract file is processed using a document recognition model with a pre-defined architecture to extract the contract text data. A graph neural network is used to parse the layout structure of contract documents and identify the hierarchical relationships between headings, paragraphs, tables, and lists. Based on the layout structure of the contract document, the contract text content is standardized by removing irrelevant formatting, correcting identification errors, and standardizing text encoding to obtain structured contract text data.

[0013] In one possible implementation, the contract text data is semantically segmented through a parsing and scheduling module to generate and assign review tasks. This also includes: A multi-label classification model is used to perform semantic segmentation on contract text data, and output the domain label and subtype label of the preset clauses; Review tasks are generated based on domain labels and subtype labels, and the task allocation strategy is dynamically optimized based on reinforcement learning algorithms.

[0014] In one possible implementation, contract review is performed by an expert intelligent agent cluster module based on the knowledge reserve module of the large model for review task scheduling, and also includes: Based on the review task, legal clause review is performed on contract text data, including semantic role labeling, clause type determination and logical relationship analysis; Based on the review task, conduct a violation risk review of contract text data, including risk level classification, risk transmission analysis, and risk hedging suggestions; Based on the review task, the contract text data is reviewed for textual standardization, including legal matching, industry standard verification, and internal policy review; Cost-benefit reviews of contract text data are conducted based on review tasks, including payment terms analysis, cost-benefit assessment, and tax impact analysis.

[0015] In one possible implementation, the review process is supported by providing several large language models and pre-defined knowledge bases through a large model knowledge reserve module, and also includes: Models are selected from heterogeneous large model units containing multiple large language models of different sizes, and model resources are dynamically allocated according to the review task. A multimodal vector database is established based on pre-defined laws and regulations; The review requests are assigned to the corresponding large language models based on the review task.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: This invention provides an automated contract review system based on a large language model. By performing multimodal parsing on uploaded contract files, it comprehensively acquires various types of information within the contract, facilitating a complete understanding of the contextual environment of contract clauses. This allows for a more accurate analysis of the true meaning of the clauses and effectively reduces risk misjudgments caused by missing information. Semantic segmentation of the contract text data divides the contract content into different parts according to semantic logic, aiding in understanding the relationships and hierarchical structure between contract clauses. This generates and rationally allocates review tasks, making the review process more targeted and systematic. It enables better in-depth analysis of clauses within the context, further reducing the possibility of risk misjudgments. Through semantic segmentation and task allocation, content related to different domains within the contract can be assigned to corresponding review units, enabling a comprehensive assessment of potential complex risks and improving the comprehensiveness and accuracy of the review. Contract review is based on a large model knowledge reserve module for review task scheduling. The expert intelligent agent cluster can acquire relevant knowledge and model support from the large model knowledge reserve module according to the review needs of different domains. This allows for a comprehensive assessment of complex risks, considering factors from various domains within the contract, and providing more comprehensive and accurate review opinions. It provides several large language models and preset knowledge bases, covering knowledge and information from different fields, which can meet the needs of cross-domain collaborative review. It can deeply understand the semantics and logical relationships of contract texts, and combined with the rich knowledge in the preset knowledge base, it can more accurately analyze the true meaning of the clauses, providing strong support for review. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the structure of an embodiment of the automated contract review system based on a large language model provided by the present invention. Figure 2 This is a flowchart illustrating an embodiment of the automated contract review method based on a large language model provided by the present invention. Figure 3 Provided by the present invention Figure 1 A flowchart illustrating an embodiment of step S201. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] The terms "first," "second," "third," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0021] This invention provides an automated contract review system and method based on a large language model, which will be described below.

[0022] Please see Figure 1 , Figure 1This is a schematic diagram of an embodiment of the automated contract review system based on a large language model provided by the present invention. In a specific embodiment of the present invention, an automated contract review system 100 based on a large language model is disclosed, including: a document preprocessing module 110, a parsing and scheduling module 120, an expert intelligent agent cluster module 130, a large model knowledge reserve module 140, and a review report module 150; the document preprocessing module 110, the parsing and scheduling module 120, the expert intelligent agent cluster module 130, the large model knowledge reserve module 140, and the review report module 150 are connected in sequence; The document preprocessing module 110 is used to perform multimodal parsing on the uploaded contract file and extract the contract text data. The parsing and scheduling module 120 is used to perform semantic segmentation on contract text data, generate review tasks, and assign review tasks. The expert intelligent agent cluster module 130 is used to perform contract review based on the knowledge reserve module 140 of the large model for review task scheduling; The large model knowledge reserve module 140 is used to provide several large language models and preset knowledge bases; The audit report module 150 is used to integrate the contract audit results of the expert intelligent agent cluster module 130 and generate a contract audit report.

[0023] In the above embodiments, the document preprocessing module 110 is the front-end processing unit of the system, responsible for multimodal parsing of uploaded contract documents. It can process not only traditional text files (such as Word and PDF) but also contract documents containing complex elements such as images and charts. Through optical character recognition (OCR) technology and layout analysis algorithms, it extracts text data from the contract and performs cleaning and standardization processing to ensure the accuracy and efficiency of subsequent review. Multimodal parsing ensures the comprehensive acquisition of all key information in the contract, avoiding review biases caused by missing information. The automated preprocessing process significantly reduces the time spent on manual data entry and organization of contract information, improving overall review efficiency.

[0024] The parsing and scheduling module 120 is responsible for performing deep semantic segmentation on the contract text data. This module utilizes NLP (Natural Language Processing) technology to segment the contract text into independent semantic information based on legal or commercial semantics, such as legal clauses, technical specifications, and payment terms. Based on the type and complexity of the semantic modules, it generates corresponding review tasks and assigns them to the corresponding agents in the expert agent cluster module 130 for processing. Through semantic segmentation, the true meaning and contextual logic of the contract clauses are accurately understood, reducing the risk of misjudgment. The automated task allocation mechanism ensures that review tasks are quickly and accurately delivered to the most suitable expert agents, improving review efficiency.

[0025] The expert agent cluster module 130 comprises multiple parallel agents, each with its own specific function, such as a clause analysis unit, risk assessment unit, compliance check unit, financial audit unit, and technical specification unit. These agents, based on the audit tasks assigned by the parsing and scheduling module 120, utilize resources and knowledge from the large model knowledge reserve module 140 to conduct in-depth audits and analyses of contracts. The collaborative work of expert agents from different domains enables comprehensive consideration of various aspects of the contract, facilitating a holistic assessment of complex risks. The agent cluster module can automatically adapt to the contract audit needs of different structures and industries, improving the system's versatility and flexibility.

[0026] The Large Model Knowledge Reserve Module 140 provides several large language models and preset knowledge bases. The large language models include general LLMs and fine-tuned LLMs, providing natural language processing capabilities tailored to different review tasks. The preset knowledge bases cover laws and regulations, industry standards, company standard clauses, and industry-specific (e.g., meteorology) professional knowledge bases, providing real-time and accurate knowledge support for expert agents. The large language models can deeply understand the semantics and logical relationships of contract texts, and combined with knowledge from the preset knowledge bases, more accurately analyze the true meaning of clauses. The industry-specific professional knowledge bases provide the system with the ability to effectively review professional clauses, meeting the specific requirements of contracts in different industries.

[0027] The audit report module 150 integrates the contract audit results from the expert intelligent agent cluster module 130 to generate a structured contract audit report. This module summarizes, deduplicates, and performs logical checks on the audit results, generating final audit opinions based on preset rules or a lightweight decision-making model. These opinions are presented in a standardized format for easy review, archiving, and analysis by contract management personnel. The structured audit report comprehensively reflects the risks and problems present in the contract. The long-term accumulation and analysis of these reports helps enterprises summarize lessons learned, optimize contract management processes, strengthen risk control capabilities, and improve operational efficiency and competitiveness.

[0028] Compared with existing technologies, this embodiment provides a contract automated review system 100 based on a large language model. By performing multimodal parsing on uploaded contract documents, it comprehensively acquires various types of information within the contract, facilitating a complete understanding of the context of contract clauses. This allows for a more accurate analysis of the true meaning of the clauses and effectively reduces risk misjudgments caused by missing information. Semantic segmentation of the contract text data divides the contract content into different parts according to semantic logic, aiding in understanding the relationships and hierarchical structure between contract clauses. This generates and rationally allocates review tasks, making the review process more targeted and systematic. It allows for better in-depth analysis of clauses within the context, further reducing the possibility of risk misjudgments. Through semantic segmentation and task allocation, content related to different domains within the contract can be assigned to corresponding review units, enabling a comprehensive assessment of potential complex risks and improving the comprehensiveness and accuracy of the review. Contract review is conducted based on the large model knowledge reserve module 140 for review task scheduling. The expert intelligent agent cluster can obtain corresponding knowledge and model support from the large model knowledge reserve module 140 according to the review needs of different domains. This allows for a comprehensive assessment of complex risks, considering factors from various domains within the contract, and providing more comprehensive and accurate review opinions. It provides several large language models and preset knowledge bases, covering knowledge and information from different fields, which can meet the needs of cross-domain collaborative review. It can deeply understand the semantics and logical relationships of contract texts, and combined with the rich knowledge in the preset knowledge base, it can more accurately analyze the true meaning of the clauses, providing strong support for review.

[0029] In some embodiments of the present invention, the document preprocessing module 110 further includes: The text recognition unit is used to extract contract text content using a document recognition model; The layout analysis unit is used to parse the layout structure of contract documents using graph neural networks; The text cleaning unit is used to perform standardized processing of contract text content based on the layout structure of the contract document to obtain contract text data.

[0030] In the above embodiments, the text recognition unit utilizes document recognition models (such as OCR technology combined with deep learning algorithms) to efficiently and accurately extract text content from contract documents. It can recognize text in various fonts, sizes, and layouts, ensuring the extracted text content is complete and error-free. Automated text recognition avoids errors and omissions that may occur with manual input, improving the accuracy and efficiency of text extraction.

[0031] The layout analysis unit uses graph neural network (GNN) technology to parse the layout structure of contract documents. It can handle complex graphical structure data and effectively identify elements such as paragraphs, headings, tables, and charts in the contract, as well as their interrelationships. Through layout analysis, the system can more accurately understand the hierarchical structure and information distribution of contract documents, providing precise layout information for text cleaning and semantic segmentation, thus helping to improve the accuracy and efficiency of review.

[0032] The text cleaning unit, based on the layout structure information provided by the layout analysis unit, standardizes the contract text content, removing noise data (such as redundant spaces and line breaks), unifying text format (such as font and font size), and segmenting the text into sentences and paragraphs to obtain structured contract text data. The text cleaning unit ensures the standardization and consistency of the contract text data, providing a high-quality data foundation for subsequent processing by the parsing and scheduling module 120 and the expert agent cluster module 130, thus contributing to improved accuracy and stability of the review process.

[0033] In some embodiments of the present invention, the parsing and scheduling module 120 further includes: The semantic domain segmenter uses a multi-label classification model to perform semantic segmentation on contract text data and outputs the domain labels and subtype labels of the preset clauses. The task scheduler generates review tasks based on domain labels and subtype labels, and dynamically optimizes the task allocation strategy based on reinforcement learning algorithms.

[0034] In the above embodiments, the semantic domain segmenter uses a multi-label classification model to perform refined semantic segmentation of contract text data. This requires not only that the model can accurately identify the various clauses in the contract, but also that these clauses be labeled with both domain and subtype labels in order to achieve a comprehensive analysis of the contract content.

[0035] The multi-label classification model is built on deep learning technology. It learns the complex mapping relationships between clauses and domain labels and subtype labels by training on a large amount of labeled contract text data. The model employs neural network structures such as variants of convolutional neural networks (CNNs) or recurrent neural networks (RNNs) to capture semantic features and contextual information in the text. During the inference phase, the model automatically analyzes the input contract text clause by clause, outputting the domain label and subtype label corresponding to each clause.

[0036] The semantic domain segmenter achieves accurate semantic segmentation of contract text through a multi-label classification model. It can not only identify key clauses in the contract, but also assign the most appropriate domain and subtype labels to them according to the specific content of the clauses. This simplifies the design and allocation of subsequent review tasks, enabling reviewers to quickly locate the types of clauses that need attention and improve review efficiency.

[0037] The task scheduler automatically generates corresponding review tasks based on the domain labels and subtype labels output by the semantic domain segmenter, and dynamically optimizes the task allocation strategy through reinforcement learning algorithms to achieve the optimal allocation of review resources.

[0038] The task scheduler first receives the tag information output by the semantic domain divider, and then generates corresponding review tasks for each clause according to the preset review rules and processes. These tasks include, but are not limited to, clause legality checks, clause integrity verification, and clause consistency comparisons. This ensures that each review task clearly specifies the review object, review standards, and expected results, providing clear guidance for subsequent review work.

[0039] To further improve review efficiency and quality, the task scheduler incorporates reinforcement learning algorithms to dynamically optimize task allocation strategies. In task scheduling scenarios, the scheduler continuously adjusts based on the current task queue status and task priorities, gradually learning the optimal task allocation strategy and improving overall review efficiency.

[0040] In some embodiments of the present invention, the expert agent cluster module 130 further includes: The clause analysis unit is used to review legal clauses in contract text data based on review tasks. The risk assessment unit is used to review contract text data for compliance risks based on the review task. The compliance inspection unit is used to conduct textual compliance reviews of contract text data based on review tasks. The financial audit unit is used to conduct cost-benefit audits on contract text data based on audit tasks.

[0041] In the above embodiments, the clause analysis unit utilizes Natural Language Processing (NLP) technology and legal knowledge graphs to meticulously dissect and compare each clause in the contract text. First, it identifies the legal clauses in the contract, then matches and analyzes them against a pre-set legal provisions database and historical case database to assess the legality, validity, and potential legal risks of the clauses, generating a clause compliance report that clearly points out any potential legal loopholes, ambiguities, or non-compliance in the clauses.

[0042] The risk assessment unit identifies potential risks of non-compliance during contract execution, including but not limited to contract fraud, breach of contract, and compliance risks. By constructing a risk assessment model and combining big data analytics and machine learning algorithms, it quantifies the risks associated with key information in the contract text. Simultaneously, it considers the impact of external factors such as market changes and policy adjustments on contract risks, providing risk level classifications and detailed risk descriptions, offering risk warnings to both parties, and recommending appropriate risk mitigation measures.

[0043] The compliance check unit ensures that contract texts comply with industry standards, internal corporate policies, and international / domestic laws and regulations, avoiding compliance issues caused by non-standard text. Based on a pre-set compliance standard library, it performs a step-by-step check on the contract text, including but not limited to formatting, terminology, and clause completeness. Simultaneously, it utilizes OCR technology to recognize textual information in scanned documents, ensuring a comprehensive review, generating a compliance check report, listing non-compliance items and their specific locations, and providing compliance improvement suggestions to help users quickly revise contract texts.

[0044] The financial review unit is responsible for cost-benefit analysis, assessing the economic rationality and financial feasibility of contracts. By integrating financial analysis models with a budget management system, it conducts in-depth analysis of key financial information in contracts, such as price terms, payment methods, cost structures, and revenue forecasts. Simultaneously, considering the financial impact of factors such as the time value of money and inflation rates, it provides financial feasibility analysis reports, including cost-benefit analysis, return on investment forecasts, and cash flow analysis, providing a financial basis for contract decisions and ensuring the maximization of contract economic benefits.

[0045] It is understood that the expert intelligent agent cluster module 130 can also adjust the number and types of internal units according to the actual situation. In addition to the units listed above, other functional units can be added or deleted as needed. This invention will not list them one by one.

[0046] In some embodiments of the present invention, the large model knowledge reserve module 140 further includes: Heterogeneous large model unit, containing multiple large language models of different sizes; The resource management unit allocates review requests to the corresponding large language models based on the review tasks. Domain knowledge base, which establishes a multimodal vector database based on preset laws and regulations.

[0047] In the above embodiments, the heterogeneous large model unit integrates multiple large language models (LLMs) of different sizes and architectures, covering general-purpose large models (such as GPT-4, LLaMA, etc.), domain-specific fine-tuning models (such as the legal vertical model Legal-BERT), and lightweight models (such as TinyBERT). These different models complement each other in terms of parameter scale, training data, and inference speed. The general-purpose large model handles tasks such as open-ended text generation and complex logical reasoning; the domain-specific fine-tuning model focuses on vertical scenarios such as legal clause parsing and compliance judgment; and the lightweight model supports low-latency requirements such as real-time interaction and rapid response.

[0048] The resource management unit addresses resource contention issues during concurrent calls to multiple models, optimizing inference efficiency and cost. Based on the type of review task, data modality, and real-time load, and using review quality, response time, and computational cost as reward functions, it dynamically adjusts the model allocation strategy, allocating review tasks to the most suitable large model.

[0049] The domain knowledge base is built on pre-defined laws and regulations, industry standards and historical cases to construct a multimodal vector database, which supports efficient retrieval of structured and unstructured knowledge. It transforms legal provisions, contract templates, case documents and other documents into high-dimensional semantic vectors, and labels clause types, risk levels, compliance standards and other elements as knowledge graph nodes. It also generates node vectors through graph neural networks (GNN) and extracts feature vectors from non-text elements such as charts and signatures in contracts.

[0050] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the automated contract review method based on a large language model provided by the present invention. According to a second aspect of the present invention, an automated contract review method based on a large language model is also provided, comprising: [The method is described in the original text, but the provided text is incomplete and cannot be translated accurately.] S201. The uploaded contract file is processed using a multimodal parsing module to extract the contract text data. S202. Semantically segment the contract text data through the parsing and scheduling module, generate review tasks, and assign review tasks. S203. Contract review is conducted through the expert intelligent agent cluster module based on the knowledge reserve module of the large model for review task scheduling. S204. The large model knowledge reserve module provides several large language models and preset knowledge bases to support the review process; S205. Integrate the contract review results from the expert intelligent agent cluster module through the review report module to generate a contract review report.

[0051] In the above embodiments, the system receives contract files uploaded by users, supporting formats such as PDF, Word, images, and scanned documents. Text content is extracted using Optical Character Recognition (OCR) technology and layout analysis algorithms. Simultaneously, Non-Textual Logic (NLP) technology is used to identify non-text elements in the contract, such as tables, signatures, and seals. The text content is normalized, unifying Chinese and English punctuation and number formats, correcting OCR recognition errors, and breaking down the contract into independent clause units based on features such as paragraph indentation and serial number identification. Table data is associated with corresponding clause texts, establishing a mapping relationship between clauses and data. Structured contract text data is generated, including clause text, clause type, and associated attachments.

[0052] NLP technology is used to perform semantic segmentation on structured contract text data to obtain multiple clauses. Each clause undergoes domain label subtype label prediction to generate corresponding review tasks. Based on the current task queue length, the load of each model, and the historical review accuracy, a priority-marked review task queue is generated and distributed to the expert agent cluster.

[0053] Please see Figure 3 , Figure 3 Provided by the present invention Figure 1 A flowchart illustrating an embodiment of step S201. In some embodiments of the present invention, the document preprocessing module 110 performs multimodal parsing processing on the uploaded contract file to extract contract text data, and further includes: S301. Use a document recognition model with a preset architecture to perform multimodal parsing on the uploaded contract file and extract the contract text data; S302. Use graph neural networks to parse the layout structure of contract documents and identify the hierarchical relationships of headings, paragraphs, tables and lists; S303. Based on the layout structure of the contract document, the contract text content is standardized by removing irrelevant formatting, correcting identification errors, and standardizing text encoding to obtain structured contract text data.

[0054] In the above embodiments, a multimodal deep learning architecture is used to comprehensively parse the contract document. This architecture integrates computer vision and natural language processing technologies, enabling it to simultaneously capture the document's visual features and textual semantic information. A cross-modal attention mechanism dynamically associates visual and textual features; for example, visual cues such as bold font and enlarged font size are used to locate the contract title, while contextual semantics are combined to confirm clause boundaries. For complex layouts, a block processing and global aggregation strategy can be used, dividing the document into local regions for independent parsing before merging the results, ensuring that local details are preserved without losing global logic.

[0055] Graph Neural Networks (GNNs) are introduced to structurally model contract layouts, transforming documents into node-edge relationship graphs. Nodes represent text blocks with attributes including position coordinates, font style, and line spacing; edges describe the spatial relationships between text blocks. Through graph attention mechanisms, dependencies between nodes can be dynamically learned. For table structures, edge weights are used to distinguish between table headers and bodies, and information on row and column merging within cells is parsed. Furthermore, GNNs support recursive reasoning, handling multi-level nested structures, ensuring the depth and accuracy of layout parsing, and providing a logical framework for subsequent text normalization, making clause location and relationship extraction more efficient.

[0056] By combining a rule engine with statistical models, the module identifies and removes non-core content such as headers, footers, watermarks, and revision marks. It utilizes layout features to locate noisy data and employs semantic analysis to avoid mistakenly deleting crucial information. Context-aware error correction technology, combined with language models and layout constraints, corrects OCR or model recognition errors. It standardizes basic elements such as character encoding, date formats, and currency symbols, and uses layout alignment algorithms to regulate paragraph indentation, line spacing, and other typesetting details. For tabular data, the module splits unstructured text into structured fields ("Subject Type" and "Subject Name") based on the logical relationship between the table header and body, facilitating subsequent clause analysis.

[0057] In some embodiments of the present invention, the process of semantically segmenting contract text data through the parsing and scheduling module 120 to generate and assign review tasks further includes: A multi-label classification model is used to perform semantic segmentation on contract text data, and output the domain label and subtype label of the preset clauses; Review tasks are generated based on domain labels and subtype labels, and the task allocation strategy is dynamically optimized based on reinforcement learning algorithms.

[0058] In the above embodiments, semantic segmentation of contract text is achieved through a multi-label deep learning framework. A hierarchical labeling system is adopted, simultaneously outputting domain labels and subtype labels for each clause. Through a multi-scale feature fusion mechanism, character-level, word-level, and sentence-level semantic information are combined to capture the implicit logical relationships in the clauses. Robustness to ambiguous expressions is enhanced through adversarial example training, and external knowledge graphs are used to supplement contextual information, improving the accuracy and consistency of semantic segmentation.

[0059] Based on the domain and subtype tags output by semantic segmentation, a rule-based inference engine combined with dynamic template matching is used to generate review tasks. A predefined task generation rule base is established to clarify the review points corresponding to different tag combinations and to link relevant legal and regulatory libraries and risk case libraries. For complex clauses, a logical decomposition engine is used to break them down into sub-tasks and to mark the dependencies between tasks.

[0060] A multi-objective reinforcement learning algorithm is introduced to dynamically optimize the task allocation strategy, with review efficiency, quality, and resource utilization as joint optimization objectives. Features such as reviewer skill tags, current workload, historical review accuracy, and task urgency are input into the policy network to allocate review tasks to the corresponding expert agent cluster module 130.

[0061] In some embodiments of the present invention, contract review is performed by the expert intelligent agent cluster module 130 based on the knowledge reserve module 140 of the large model for review task scheduling, and further includes: Based on the review task, legal clause review is performed on contract text data, including semantic role labeling, clause type determination and logical relationship analysis; Based on the review task, conduct a violation risk review of contract text data, including risk level classification, risk transmission analysis, and risk hedging suggestions; Based on the review task, the contract text data is reviewed for textual standardization, including legal matching, industry standard verification, and internal policy review; Cost-benefit reviews of contract text data are conducted based on review tasks, including payment terms analysis, cost-benefit assessment, and tax impact analysis.

[0062] In the above embodiments, Natural Language Processing (NLP) technology and legal knowledge graphs are used to meticulously dissect and compare each clause in the contract text. First, the legal clauses in the contract are identified, and then they are matched and analyzed against a pre-set legal provisions database and historical case database to assess the legality, validity, and potential legal risks of the clauses, generating a clause compliance report that clearly points out any legal loopholes, ambiguities, or non-compliance in the clauses.

[0063] Identify potential risks of non-compliance during contract execution, including but not limited to contract fraud, breach of contract, and compliance risks. By constructing a risk assessment model and combining big data analytics and machine learning algorithms, quantify the risks of key information in the contract text. Simultaneously, consider the impact of external factors such as market changes and policy adjustments on contract risks, providing risk level classifications and detailed risk descriptions, offering risk warnings to both parties, and recommending appropriate risk mitigation measures.

[0064] Ensure contract texts comply with industry standards, internal corporate policies, and international / domestic laws and regulations to avoid compliance issues arising from non-standard text. Based on a pre-defined compliance standard library, conduct a thorough check of the contract text, including but not limited to formatting, terminology, and clause completeness. Simultaneously, utilize OCR technology to recognize textual information in scanned documents, ensuring comprehensiveness of the review, generating a compliance check report, listing non-compliance items and their specific locations, and providing compliance improvement suggestions to help users quickly revise contract texts.

[0065] Responsible for cost-benefit analysis, assessing the economic rationality and financial feasibility of contracts. Through the integration of financial analysis models and budget management systems, conduct in-depth analysis of key financial information in contracts, such as price terms, payment methods, cost structures, and revenue forecasts. Simultaneously, consider the financial impact of factors such as the time value of money and inflation rates on the contract, providing financial feasibility analysis reports, including cost-benefit analysis, return on investment forecasts, and cash flow analysis, to provide a financial basis for contract decisions and ensure the maximization of contract economic benefits.

[0066] In some embodiments of the present invention, the large model knowledge reserve module 140 provides several large language models and a preset knowledge base to support the review process, and further includes: Models are selected from heterogeneous large model units containing multiple large language models of different sizes, and model resources are dynamically allocated according to the review task. A multimodal vector database is established based on pre-defined laws and regulations; The review requests are assigned to the corresponding large language models based on the review task.

[0067] In the above embodiments, the large model knowledge reserve module 140 integrates multiple large language models of different scales and architectures through a heterogeneous model resource pool, quantifying the performance of each model in dimensions such as legal terminology understanding, long text reasoning, and multilingual support. For specific review tasks, a multi-objective optimization algorithm is used to dynamically allocate model resources. First, the task requirements are analyzed, and then the optimal model combination is selected through a weighted scoring model, taking into account the real-time load of the models, historical performance data, and cost constraints.

[0068] To support accurate matching and reasoning of legal provisions, a multimodal legal vector database is constructed, uniformly encoding text, structured data, and semi-structured data into high-dimensional vectors. During queries, semantic similarity search and logical reasoning search are supported, and a vector fusion algorithm combines text, structured, and graph information to improve retrieval accuracy.

[0069] Based on the type of review task, data modality, and real-time load, and using review quality, response time, and computational cost as reward functions, the model allocation strategy is dynamically adjusted to dynamically allocate review tasks to the most suitable large model.

[0070] In summary, the contract automated review system 100 based on a large language model provided by this invention comprehensively acquires various types of information in the uploaded contract documents through multimodal parsing, which helps to fully understand the context of contract clauses and thus more accurately analyze the true meaning of the clauses, effectively reducing risk misjudgments caused by missing information. Semantic segmentation of the contract text data, dividing the contract content into different parts according to semantic logic, helps to understand the relationships and hierarchical structure between contract clauses, generates review tasks, and allocates them reasonably, making the review process more targeted and systematic. It can better combine context to conduct in-depth analysis of clauses, further reducing the possibility of risk misjudgments. Through semantic segmentation and task allocation, content involving different fields in the contract can be assigned to corresponding review units, enabling a comprehensive assessment of potential complex risks in the contract and improving the comprehensiveness and accuracy of the review. Based on the review task scheduling large model knowledge reserve module 140 for contract review, the expert intelligent agent cluster can obtain corresponding knowledge and model support from the large model knowledge reserve module 140 according to the review needs of different fields, fully considering factors from various fields in the contract, comprehensively assessing complex risks, and providing more comprehensive and accurate review opinions. It provides several large language models and preset knowledge bases, covering knowledge and information from different fields, which can meet the needs of cross-domain collaborative review. It can deeply understand the semantics and logical relationships of contract texts, and combined with the rich knowledge in the preset knowledge base, it can more accurately analyze the true meaning of the clauses, providing strong support for review.

[0071] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0072] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0073] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some service interface; the indirect coupling or communication connection between devices or units may be electrical or other forms.

[0074] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0075] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0076] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0077] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0078] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0079] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0080] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A contract automated review system based on a large language model, characterized in that, include: The module includes a document preprocessing module, a parsing and scheduling module, an expert intelligent agent cluster module, a large model knowledge reserve module, and an audit report module. The document preprocessing module, the parsing and scheduling module, the expert intelligent agent cluster module, the large model knowledge reserve module, and the review report module are connected in sequence. The document preprocessing module is used to perform multimodal parsing on the uploaded contract file and extract the contract text data. The parsing and scheduling module is used to perform semantic segmentation on the contract text data, generate review tasks, and assign review tasks. The expert intelligent agent cluster module is used to schedule the large model knowledge reserve module to perform contract review based on the review task. The large model knowledge reserve module is used to provide several large language models and preset knowledge bases; The audit report module is used to integrate the contract audit results of the expert intelligent agent cluster module and generate a contract audit report.

2. The automated contract review system based on a large language model as described in claim 1, characterized in that, The document preprocessing module also includes: The text recognition unit is used to extract contract text content using a document recognition model; The layout analysis unit is used to parse the layout structure of contract documents using graph neural networks; The text cleaning unit is used to perform planning and processing of the contract text content based on the layout structure of the contract document to obtain contract text data.

3. The automated contract review system based on a large language model as described in claim 1, characterized in that, The parsing and scheduling module also includes: The semantic domain segmenter uses a multi-label classification model to perform semantic segmentation on the contract text data and outputs the domain labels and subtype labels of the preset clauses. The task scheduler generates review tasks based on the domain label and the subtype label, and dynamically optimizes the task allocation strategy based on a reinforcement learning algorithm.

4. The automated contract review system based on a large language model as described in claim 1, characterized in that, The expert intelligent agent cluster module also includes: The clause analysis unit is used to review the legal clauses of the contract text data based on the review task. A risk assessment unit is used to conduct a violation risk assessment on the contract text data based on the audit task. A compliance inspection unit is used to perform textual standardization review on the contract text data based on the review task. The financial audit unit is used to conduct a cost-benefit audit of the contract text data based on the audit task.

5. The automated contract review system based on a large language model as described in claim 1, characterized in that, The large model knowledge reserve module also includes: Heterogeneous large model unit, containing multiple large language models of different sizes; The resource management unit allocates the review requests to the corresponding large language models according to the review tasks. Domain knowledge base, which establishes a multimodal vector database based on preset laws and regulations.

6. A method for automated contract review based on a large language model, comprising the automated contract review system based on a large language model as described in any one of claims 1-5, characterized in that, include: The uploaded contract file is processed using a document preprocessing module to perform multimodal parsing and extract the contract text data. The contract text data is semantically segmented by the parsing and scheduling module to generate and assign review tasks. Contract review is performed by the expert intelligent agent cluster module based on the knowledge reserve module of the aforementioned review task scheduling big model; The large model knowledge reserve module provides several large language models and preset knowledge bases to support the review process; The contract review report is generated by integrating the contract review results from the expert intelligent agent cluster module through the review report module.

7. The automated contract review system based on a large language model as described in claim 6, characterized in that, The step of performing multimodal parsing on the uploaded contract file through the document preprocessing module to extract contract text data also includes: The uploaded contract file is processed using a document recognition model with a pre-defined architecture to extract the contract text data. A graph neural network is used to parse the layout structure of contract documents and identify the hierarchical relationships between headings, paragraphs, tables, and lists. Based on the layout structure of the contract document, the contract text content is standardized by removing irrelevant formatting, correcting identification errors, and standardizing text encoding to obtain structured contract text data.

8. The automated contract review system based on a large language model as described in claim 6, characterized in that, The step of performing semantic segmentation on the contract text data through the parsing and scheduling module, generating review tasks, and assigning review tasks also includes: A multi-label classification model is used to perform semantic segmentation on the contract text data, and the domain labels and subtype labels of the preset clauses are output. Review tasks are generated based on the domain labels and subtype labels, and the task allocation strategy is dynamically optimized based on reinforcement learning algorithms.

9. The automated contract review system based on a large language model as described in claim 6, characterized in that, The process of reviewing contracts through the expert intelligent agent cluster module based on the knowledge reserve module of the review task scheduling big model also includes: Based on the aforementioned review task, the contract text data is reviewed for legal clauses, including semantic role labeling, clause type determination, and logical relationship analysis. Based on the aforementioned review task, the contract text data is reviewed for violations and risks, including risk level classification, risk transmission analysis, and risk hedging suggestions. Based on the aforementioned review task, the contract text data is subjected to text standardization review, including regulatory matching, industry standard verification, and internal policy review. Based on the aforementioned audit task, a cost-benefit audit is performed on the contract text data, including payment terms analysis, cost-benefit assessment, and tax impact analysis.

10. The automated contract review system based on a large language model as described in claim 6, characterized in that, The provision of several large language models and preset knowledge bases through the large model knowledge reserve module to support the review process also includes: Models are selected from heterogeneous large model units containing multiple large language models of different sizes, and model resources are dynamically allocated according to the review task. A multimodal vector database is established based on pre-defined laws and regulations; The review request will be assigned to the corresponding large language model based on the review task.

Images