CN-121998597-A - Contract examination-oriented LLM multi-agent self-adaptive dynamic collaborative design method

CN121998597ACN 121998597 ACN121998597 ACN 121998597ACN-121998597-A

Abstract

The LLM multi-agent self-adaptive dynamic collaborative design method for contract examination comprises the steps of firstly constructing a multi-type specific agent library containing general foundation, special contract and the like, covering multi-field professional examination capability, then establishing a semantic ontology dynamic communication protocol among agents, guaranteeing accurate and efficient information transmission, constructing a workflow self-adaptive engine through contract feature vectors, dynamically generating an adjustment examination path, preprocessing a contract to be examined, selecting an adaptation agent to construct a collaboration group and triggering a workflow, performing examination by the agents according to the flow, processing cross-field clause conflict by adopting a multi-element collaborative decision mechanism, generating a staged structured result, finally fusing each node result, and outputting a standardized examination report. The invention solves the problem that the prior art is difficult to meet the core requirements of complex contract examination on 'specialized division of labor', 'cross-field collaboration', 'whole process controllability'.

Inventors

DUAN GUOHUA
WU LIN
LIU JIANYU
Hu Renbing
WANG SIYU
JIANG TAO
FENG WEI

Assignees

武汉数众科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251229

Claims (10)

1. A contract examination-oriented LLM multi-agent self-adaptive dynamic collaborative design method is characterized by comprising the following steps: S100, constructing a multi-type contract examination specialized intelligent library, wherein the intelligent library adopts a field fine-tuning LLM model, and comprises a general basic intelligent group, a special contract intelligent group, a collaborative decision intelligent group and a flow management intelligent group; s200, establishing a dynamic communication protocol based on semantic ontology among the intelligent agents, defining a core field of a communication message, configuring a mixed communication structure and a phased communication strategy scheduling mechanism, and guaranteeing the accuracy and the high efficiency of information transmission among the intelligent agents; S300, designing a workflow self-adaptive engine based on contract feature vectors, generating feature vectors by extracting contract key features, constructing a multi-stage workflow adaptation mechanism, and dynamically generating and adjusting an inspection path according to contract types, complexity and risk levels; S400, constructing a contract preprocessing and agent cooperation group, receiving a text of a contract to be checked, completing format conversion, redundant information rejection, structuring processing, hierarchical recognition and type recognition on the matched text, selecting an adaptive agent from an agent library to form the cooperation group based on a recognition result, and triggering a corresponding workflow instance; S500, carrying out dynamic examination and cross-domain collaborative processing on the contract, enabling an agent to execute examination tasks according to a dynamic workflow, and aiming at cross-domain clause conflict, adopting a collaborative decision mechanism processing combining confidence weighting, a voting mechanism and collaborative decision agent decision to generate a staged structural result comprising a clause analysis abstract, a risk annotation, a modification suggestion and a compliance basis chain; s600, collecting results after the task nodes are completed, carrying out multidimensional fusion and format output, and generating a standard compound co-examination report.
2. The collaborative design method according to claim 1, wherein in S100, the general basic agent group includes a term resolution agent, a term compliance agent, and a risk quantification agent, wherein: The term analysis intelligent agent uses a Legal-BERT as a basic large model, and identifies term boundaries and associated term definitions through a term analysis special term prompt word engineering guide model; The clause compliance agent performs matching and suitability judgment of clauses and laws through a compliance verification prompt word guide model based on LLaMA series models finely tuned in the legal field; And the risk quantification intelligent body takes the general large model as a base, and applies a risk matrix to analyze risk items through a risk assessment special prompt word guide model.
3. The collaborative design method according to claim 1, wherein in S100, the special contract agent group includes a purchase contract special agent, a technical development contract special agent, a labor contract special agent, and a cross-border contract special agent, wherein: The purchasing contract special intelligent agent is based on a general large model, and the purchasing contract special intelligent agent focuses on the description, acceptance standard, quality assurance period and payment proportion core clauses of a purchasing target through a purchasing clause examination prompting word guide model; the technical development contract special intelligent agent analyzes intellectual property attribution, development result delivery standard and technical index acceptance key terms through a technical contract core term examination prompt word guide model based on a large model sensitive to technical terms; The labor contract special intelligent agent is based on a legal big model, and checks the competition limit, social security payment, working hour system and pay payment terms through a labor contract special term examination prompting word guide model; The cross-border contract special intelligent agent processes the multi-language conversion and foreign rule adaptation problem through a cross-border contract comprehensive examination prompt word guide model based on the multi-language large model.
4. The collaborative design method as claimed in claim 1, wherein in S200, establishing a dynamic communication protocol based on semantic ontology between the agents comprises the following specific steps: S201, adapting a communication structure, wherein a core layer adopts a centralized communication and control agent fusion mode, and a WorkflowManagerAgent coordination task and CommunicationProxyAgent management communication decoupling are adopted; s202, designing a communication protocol content format, defining a JSON-LD expansion format unified semantic communication data structure, automatically generating all inter-agent communication messages by ProtocolEncoderAgent to contain fields, and checking by CommunicationProxyAgent; S203, scheduling the communication strategy, and respectively starting full communication, semantic weight driving communication and global control communication strategies according to three stages of examination initialization, task metaphase and risk conflict to realize efficient and controllable agent cooperation.
5. The collaborative design method as in claim 1, wherein in S300, designing a workflow adaptation engine based on contract feature vectors comprises the steps of: s301, constructing a contract feature vector, extracting text and structural features by using a TF-IDF+BERT vector fusion algorithm after contract preprocessing, and generating a feature vector with a preset length; S302, automatically judging the flow type, and outputting a flow label by using a lightweight classification model, wherein the flow label comprises a simple flow, a standard flow and a complex flow; s303, generating a task map, and defining the task map as a directed map G=V, wherein V is a task node set comprising term analysis, special examination and manual review, E is a task dependency edge which represents execution sequence and data circulation, and automatically calling a map template library according to a flow type and performing personalized adjustment; s304, node scheduling and resource allocation are carried out, and if the time consumption of each node is expected to be larger than a preset time consumption threshold value, parallel processing or granularity division is carried out on the node so as to improve the examination efficiency; S305, defining a process health function, and monitoring and dynamically adjusting a map structure or a node strategy in real time.
6. The collaborative design method as in claim 1, wherein in S400, the specific steps for constructing the contract pre-process and agent collaboration group include: S401, performing standardization processing on the combined receiving and format, extracting texts through an OCR module and performing semantic layer error correction by combining an edit distance algorithm for image contracts, and uniformly converting all the texts into a standard intermediate structure so as to be compatible with downstream structure identification and feature extraction; S402, identifying a contract structure and extracting meta information, and carrying out format cleaning, paragraph reconstruction and clause segmentation on the contract; extracting core meta information which at least comprises contract name, target, main body information and signing time and is organized into contract structure description vectors in a key value pair form; S403, identifying and classifying the joint types, embedding the joint types by using a multi-mode contract fused with TF-IDF and BERT vectors, inputting the joint types into a lightweight classification model, and outputting a contract type prediction result under a three-level label system; S404, automatically constructing an agent cooperation group, dynamically selecting a matched agent set from an agent library according to contract types and structural features, comprehensively considering capability vectors, execution cost and field adaptation degree multidimensional indexes of an agent matching function, finally generating a task cooperation group, and initializing corresponding workflow examples by WorkflowManagerAgent.
7. The collaborative design method as in claim 1, wherein in S500, the dynamic review and cross-domain collaborative processing of the syndication comprises the steps of: s501, executing a checking task, executing a serial-parallel checking flow according to a task map topological structure, processing a term analysis and compliance checking task by an agent division, and driving a node execution strategy by an accuracy rate, cost and time delay weighting objective function; s502, carrying out multi-agent cooperative processing on cross-domain clauses, adopting three fusion strategies of confidence weighting, voting and coordination agent arbitration aiming at the cross-domain composite clauses, and integrating multi-agent judgment results; s503, carrying out staged structured output, outputting structured contents containing clause abstracts, risk levels, modification suggestions and legal basis at key nodes, organizing according to a standard JSONSchema, and supporting subsequent aggregation and report generation.
8. The collaborative design method as set forth in claim 1, wherein in S600, the results of the completion of each task node are collected, multidimensional fusion and format output are performed, and standard compound co-examination reports are generated, the specific steps include: S601, result aggregation and consistency processing are carried out, all sub-agent outputs are summarized by ResultAggregatorAgent, redundancy conflict is eliminated through semantic merging and standard correction, and JSON-LD format unified result representation is adopted; S602, integrating risk assessment and advice, calculating comprehensive risk through weighting of risk occurrence probability and influence degree, generating optimization advice by combining a rule base and a case base, and enhancing credibility according to rules or cases; S603, generating a structured examination report, namely generating the structured report containing contract information, examination conclusion, clause abstract and annex, wherein the examination report supports PDF/HTML/Markdown format export and has automatic catalog and intelligent index skip functions.
9. The collaborative design method according to claim 1, further comprising S700. Optimizing the multi-agent adaptive dynamic collaborative method by continuously optimizing agent collaborative policies and workflow adaptation rules through reinforcement learning algorithms based on censored historical data and user feedback, the specific steps comprising: s701, feedback acquisition and rewarding modeling, namely automatically recording user modification advice, adoption condition and path behavior correction after examination is completed, converting the advice, the adoption condition and the path behavior correction into rewarding signals, and constructing reinforcement learning input triples; s702, strengthening learning strategy optimization, namely adopting a preset optimization algorithm to iteratively update an agent scheduling strategy with the aim of maximizing long-term jackpot rewards; S703, system updating and strategy deployment, namely automatic online deployment after strategy stability and income are improved to reach standards, and synchronous updating of prompt words, rules and task map modules to realize system evolution and form a self-driven closed-loop intelligent body system.
10. An electronic device, comprising: One or more processors; a memory for storing one or more programs; When executed by the one or more processors, causes the one or more processors to implement the co-design method of any one of claims 1 to 9.

Description

Contract examination-oriented LLM multi-agent self-adaptive dynamic collaborative design method Technical Field The invention relates to the technical field of natural language processing and intelligent decision making, in particular to a contract examination-oriented LLM multi-agent self-adaptive dynamic collaborative design method. Background Contract review is a core risk control link in business activities, and the quality of the contract review directly relates to legal compliance of enterprises and business benefit guarantee. With the advanced development of market economy, contract types are diversified day by day, and multiple professional fields of cross-border purchase, data compliance, intellectual property rights permission, medicine GSP adaptation and the like are covered, the complexity of terms is exponentially increased, the number of large commercial contract terms often breaks through 500, and multi-field cross contents such as law, finance, technology and the like are involved. The traditional manual examination mode has the remarkable limitation that an average 100-page complex contract requires 3-5 working days to complete examination, the examination quantity of the enterprise contract is increased by more than 30%, manual team is difficult to cope with high-frequency and high-complexity examination demands, and meanwhile, a single examination person cannot cover multi-field expertise, so that a professional blind area exists in examination. The current industry mainstream contract review solutions are mainly divided into three categories: The rule engine is driven, wherein examination is realized through a preset keyword matching rule, the response speed is high, but the rule maintenance cost is high, and fuzzy expression and complex semantic scenes cannot be processed; The single LLM application type is based on a general large language model to develop an inspection function, semantic understanding capability is superior to that of a rule engine, but 'failure in consideration of the failure' is easy to occur in complex division tasks, the calculation error rate in the professional field is high, and the accuracy rate is obviously reduced when the industrial corpus is lacking; And the human-computer collaborative auxiliary mode is adopted, wherein AI is only responsible for clause element extraction, closed loop intelligent examination capability is not formed, examination quality depends on manual experience, and stability is insufficient. The complex contract examination essence is a multi-task cooperation process, and links such as key information extraction, compliance verification, risk assessment and the like are required to be completed in sequence, but the existing scheme has the problems of task splitting and cooperation loss, and the single model end-to-end output mode causes insufficient granularity and poor interpretability of examination results. Therefore, the prior art is difficult to meet the core requirements of complex contract checking on 'specialized division of labor', 'cross-field collaboration', 'full-process controllability', and a novel intelligent checking frame needs to be constructed. Disclosure of Invention The invention aims to solve at least one technical problem existing in the prior art, and provides a contract examination-oriented LLM multi-agent self-adaptive dynamic collaborative design method. In a first aspect, an embodiment of the present invention provides a contract-oriented LLM multi-agent adaptive dynamic collaborative design method, including: S100, constructing a multi-type contract examination specialized intelligent library, wherein the intelligent library adopts a field fine-tuning LLM model, and comprises a general basic intelligent group, a special contract intelligent group, a collaborative decision intelligent group and a flow management intelligent group; s200, establishing a dynamic communication protocol based on semantic ontology among the intelligent agents, defining a core field of a communication message, configuring a mixed communication structure and a phased communication strategy scheduling mechanism, and guaranteeing the accuracy and the high efficiency of information transmission among the intelligent agents; S300, designing a workflow self-adaptive engine based on contract feature vectors, generating feature vectors by extracting contract key features, constructing a multi-stage workflow adaptation mechanism, and dynamically generating and adjusting an inspection path according to contract types, complexity and risk levels; S400, constructing a contract preprocessing and agent cooperation group, receiving a text of a contract to be checked, completing format conversion, redundant information rejection, structuring processing, hierarchical recognition and type recognition on the matched text, selecting an adaptive agent from an agent library to form the cooperation group based on a recognition result, and triggering a