Multi-agent collaborative confrontation case trial method and system

Abstract

The invention provides a multi-agent collaborative confrontation case trial method and system, and relates to the technical field of artificial intelligence, and the method comprises the steps: setting different system cues for different agents, obtaining a role agent set, and constructing a court trial environment based on the role agent set and an integrated knowledge base; an initial information set is formed through a judge agent according to case statement information of a to-be-examined case, and a core fact evidence set is constructed according to complaint and answer document data and evidence information; calling a quality experience knowledge base through an original and informed lawyer agent, and determining dispute information based on the initial information set and the core fact evidence set; calling a primary and defensive lawyer agent to perform confrontation debate for the dispute information according to a debate experience knowledge base in the integrated knowledge base so as to obtain first discussion data and second discussion data; and generating a formatted judgment document through the judge agent according to the core fact evidence set, the first discussion data, the second discussion data and the integrated knowledge base.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method and system for handling cases involving multi-agent collaborative adversarial proceedings. Background Technology

[0002] With the breakthrough development of artificial intelligence technology, especially large language models, the intelligent transformation of the legal field is deepening. Traditional legal simulation and training mostly rely on static analysis of pre-set cases or human role-playing, which suffers from problems such as limited scenarios, limited interactivity, high costs, and difficulty in replicating the complex and dynamic processes of real court trials at scale. In recent years, researchers have begun to explore the introduction of artificial intelligence into legal practice training and decision support. For example, the "Meta-Law Big Model" developed by Peking University has significantly improved the accuracy and logical interpretability of legal citations by introducing a rule-based reasoning engine; Nanjing University's LaWGPT has enhanced the model's semantic understanding of professional texts through large-scale legal corpus pre-training and domain vocabulary expansion; in addition, some judicial institutions have also piloted the application of large legal models to assist in handling cases, demonstrating the potential to improve judicial efficiency.

[0003] However, existing technological solutions still face significant challenges in constructing highly realistic and dynamic mock court environments. First, most systems focus on one-way legal question-and-answer or document generation, lacking mechanisms for collaboration and adversarial interaction among multiple intelligent agents, and thus failing to simulate the complex interactions and strategic games between plaintiffs, defendants, judges, and other parties during trials. Second, these systems generally suffer from homogenized role behavior; the agents struggle to reflect the differences in experience, style, and strategy among real legal practitioners, resulting in a superficial simulation lacking immersion. More importantly, most existing systems operate in a closed, automated mode with weak human-computer interaction capabilities. Users cannot deeply participate in the trial process as core participants (such as lawyers), engaging in real-time rebuttals and strategic adjustments, thereby limiting their practical value in legal practice teaching and skills assessment. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a case adjudication method based on multi-agent collaborative adversarial approaches. This method addresses the problems of homogenized agent behavior, lack of dynamic adversarial interaction in real legal scenarios, weak human-computer interaction leading to a monotonous user experience, and difficulty in achieving a complete "practice-feedback-optimization" closed loop in existing technologies.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0006] Firstly, this application provides a method for adjudicating cases involving multi-agent cooperative confrontation, comprising the following steps: Different system prompts are set for different intelligent agents to obtain a set of role intelligent agents, which includes judge intelligent agent, plaintiff intelligent agent, plaintiff's lawyer intelligent agent, defendant intelligent agent, and defendant's lawyer intelligent agent. A court hearing environment is constructed based on the set of role intelligent agents and an integrated knowledge base based on a legal knowledge graph architecture. The integrated knowledge base includes a legal provisions knowledge base, an evidence examination experience knowledge base, a debate experience knowledge base, and a judge's trial experience knowledge base. In the courtroom setting, the judge's AI agent forms an initial information set based on the case statements of the plaintiff's AI agent and the defendant's AI agent. The judge's AI agent then constructs a core factual evidence set based on the plaintiff's AI agent's written response data, the plaintiff's AI agent's evidence information, the defendant's AI agent's written response data, and the defendant's AI agent's evidence information. The plaintiff's lawyer's AI agent and the defendant's lawyer's AI agent access a knowledge base of cross-examination experience to determine the disputed information based on the initial information set and the core factual evidence set. Based on the plaintiff's response document data, the first task information of the plaintiff's lawyer is determined, and based on the defendant's response document data, the second task information of the defendant's lawyer is determined. The plaintiff's lawyer and the defendant's lawyer are then invoked to conduct adversarial debate on the disputed information according to the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information. The judge's intelligent agent performs legal reasoning and generates formatted judgment documents based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base.

[0007] Optionally, the step of setting different system prompts for different agents to obtain a set of role agents includes: Different system prompts are set for different intelligent agents to obtain a set of role-based intelligent agents. The system prompts include basic system prompts and feature-based system prompts. The basic system prompts for the judge intelligent agent include a specific set of instructions for process status progression, summarizing points of contention, and generating the final judgment. The basic system prompts for the plaintiff's lawyer intelligent agent and the defendant's lawyer intelligent agent include differentiated content on the allocation of the burden of proof, selection of cross-examination strategies, and norms for debate rhetoric. The basic system prompts for the plaintiff's intelligent agent and the defendant's intelligent agent include format norms and information scope limitations for expressing case claims. The system prompts for the judge's AI agent include gender, years of practice, and the region and field of practice; the system prompts for the plaintiff's lawyer AI agent and the defendant's lawyer AI agent include gender, years of practice, historical win rate, and the region and field of practice; the system prompts for the plaintiff's AI agent and the defendant's AI agent are empty.

[0008] Optionally, the legal text knowledge base vectorizes legal texts and establishes a legal text association network to provide the intelligent agent with accurate legal basis retrieval; the cross-examination experience knowledge base stores and encodes cross-examination patterns and rebuttal strategies for preset evidence; the debate experience knowledge base contains rhetoric templates and logical frameworks for various debate scenarios; and the judge trial experience knowledge base learns the key points of historical cases and sentencing rules through machine learning models.

[0009] Optionally, the step of invoking the plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain first argument data corresponding to the first task information and second argument data corresponding to the second task information, includes: Obtain the query statements of the plaintiff's lawyer AI agent and the defendant's lawyer AI agent regarding the disputed information; The query statement is rewritten into an optimized query statement containing the knowledge base domain identifier using a domain adaptive model; Based on the optimized query statement, dense vector retrieval and sparse retrieval are performed in parallel, and preliminary results are obtained from the legal text knowledge base and the debate experience knowledge base through the similarity function; The preliminary results are sorted using an attention-based re-ranking model to generate the final retrieval set; Based on the optimized query statement and the final retrieval set, prompt context data is constructed, and the plaintiff's lawyer agent and the defendant's lawyer agent are invoked to generate first argument data and second argument data according to the prompt context data.

[0010] Optionally, the step of invoking the plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain first argument data corresponding to the first task information and second argument data corresponding to the second task information, includes: Calculate the real-time information state changes and criterion alignment based on the first and second argument data; When the real-time information status change meets the first preset condition and the criterion alignment meets the second preset condition, the step of calling the plaintiff lawyer intelligent agent and the defendant lawyer intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and debate experience knowledge base in the integrated knowledge base to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information is stopped. Calculate the first semantic similarity of the first argument data output by the plaintiff's lawyer intelligent agent in the adversarial debate process. When the first semantic similarity meets the preset condition, stop calling the plaintiff's lawyer intelligent agent. Calculate the second semantic similarity of the second argument data output by the defendant's lawyer agent during the adversarial debate process. When the second semantic similarity meets a preset condition, stop calling the defendant's lawyer agent.

[0011] Optionally, the step of performing legal reasoning and generating a formatted judgment document by the judge's intelligent agent based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base includes: The judge's intelligent agent uses a legal reasoning model to reason based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base, and obtains the reasoning result; Based on the reasoning results and the judgment document generation module, a formatted judgment document is generated.

[0012] Optionally, the method further includes: Obtain user request information and determine the target intelligent agent based on the user request information; When the target intelligent agent is a judge intelligent agent, the case statement information of the plaintiff intelligent agent and the defendant intelligent agent regarding the case to be tried, the litigation document data of the plaintiff intelligent agent, the evidence information of the plaintiff intelligent agent, the litigation document data of the defendant intelligent agent, and the evidence information of the defendant intelligent agent are displayed as display information to obtain the initial information set, factual evidence, and disputed information determined by the plaintiff lawyer intelligent agent and the defendant lawyer intelligent agent according to the cross-examination stage based on the displayed information; Based on the plaintiff's response document data, the first task information of the plaintiff's lawyer's intelligent agent is determined. Based on the defendant's response document data, the second task information of the defendant's lawyer's intelligent agent is determined. The plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent are invoked to conduct adversarial debate on the disputed information according to the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information. Using a legal reasoning model, legal reasoning is performed based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and the judge's trial experience knowledge base in the integrated knowledge base to generate a formatted judgment document. The first argument data and the second argument data regarding the disputed information are displayed to obtain feedback data on the first argument data and the second argument data. The formatted judgment document and the feedback data are matched to obtain a score for the feedback data.

[0013] On the other hand, this application provides a case adjudication system for multi-agent cooperative confrontation, the system comprising: The module is used to set different system prompts for different intelligent agents to obtain a set of role intelligent agents, which includes a judge intelligent agent, a plaintiff intelligent agent, a plaintiff's lawyer intelligent agent, a defendant intelligent agent, and a defendant's lawyer intelligent agent. Based on the set of role intelligent agents and an integrated knowledge base based on a legal knowledge graph architecture, a court hearing environment is constructed. The integrated knowledge base includes a legal provisions knowledge base, an evidence examination experience knowledge base, a debate experience knowledge base, and a judge's trial experience knowledge base. The first calculation module is used to form an initial information set based on the case statements of the plaintiff and defendant in the trial environment by the judge's intelligent agent. The judge's intelligent agent constructs a core factual evidence set based on the plaintiff's response documents, the plaintiff's evidence information, the defendant's response documents, and the defendant's evidence information. The judge's intelligent agent then calls the plaintiff's lawyer and the defendant's lawyer's intelligent agents to determine the disputed information based on the core factual evidence set according to the cross-examination experience knowledge base. The second calculation module is used to determine the first task information of the plaintiff's lawyer intelligent agent based on the plaintiff's intelligent agent's response document data, and to determine the second task information of the defendant's lawyer intelligent agent based on the defendant's intelligent agent's response document data. The module then calls upon the plaintiff's lawyer intelligent agent and the defendant's lawyer intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information. The output module is used to perform legal reasoning and generate formatted judgment documents by the judge's intelligent agent based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base.

[0014] Thirdly, this application also provides an electronic device, including a memory, a processor, and a first computer program stored in the memory and executable on the processor, wherein the processor executes the first computer program to implement the case trial method of multi-agent cooperative confrontation as described above.

[0015] Fourthly, this application also provides a computer-readable storage medium storing a second computer program, which, when executed by a processor, implements the multi-agent cooperative adversarial case handling method described above.

[0016] Beneficial effects: This application constructs a standardized four-stage trial process, defining a collaborative adversarial framework for multiple intelligent agents, including judges, court clerks, plaintiffs, defendants, and their lawyers. It shapes differentiated behavioral patterns by injecting characteristic data from real legal practitioners; provides professional support to the agents through the integration of four knowledge bases: legal provisions, evidence examination experience, debate experience, and judge's trial experience; achieves intelligent control of the debate phase through a dual convergence mechanism based on information state monitoring and semantic similarity analysis; automatically generates standardized judgment documents through legal reasoning based on multi-source information fusion; and supports users to participate in the trial as lawyers and obtain evaluation feedback through an immersive human-computer interaction interface. Ultimately, it realizes a legal simulation system integrating full-process simulation, personalized roles, professional knowledge, intelligent debate, automated judgment, and closed-loop training. The aim is to solve problems in existing mock court systems such as homogenized agent roles, lack of dynamic adversarial interaction in real legal scenarios, weak human-computer interaction functions, and difficulty in achieving a complete "practice-feedback-optimization" closed loop. Attached Figure Description

[0017] Figure 1 This is an architecture diagram of a multi-agent collaborative adversarial case trial system in this invention; Figure 2 This is a technical roadmap of the dynamic convergence mechanism for debate in this invention; Figure 3 This is a framework diagram of the multi-source heterogeneous knowledge base system in this invention; Figure 4 This is a flowchart of the multi-structure knowledge retrieval mechanism in this invention; Figure 5 This is a diagram of the two-way human-computer interaction and evaluation feedback framework in this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0019] The embodiments described in this application are merely some, not all, embodiments of the present invention. Based on the spirit of the present invention, all other embodiments obtained by those skilled in the art without inventive effort are within the protection scope of the present invention.

[0020] See Figure 1 As shown, the present invention provides a method for handling cases involving multi-agent cooperative adversarial proceedings, comprising the following steps: Different system prompts are set for different intelligent agents to obtain a set of role intelligent agents, which includes judge intelligent agent, plaintiff intelligent agent, plaintiff's lawyer intelligent agent, defendant intelligent agent, and defendant's lawyer intelligent agent. A court hearing environment is constructed based on the set of role intelligent agents and an integrated knowledge base based on a legal knowledge graph architecture. The integrated knowledge base includes a legal provisions knowledge base, an evidence examination experience knowledge base, a debate experience knowledge base, and a judge's trial experience knowledge base. In the courtroom setting, the judge's AI agent forms an initial information set based on the case statements of the plaintiff's AI agent and the defendant's AI agent regarding the case to be heard. The judge's AI agent then constructs a core factual evidence set based on the plaintiff's AI agent's pleadings, the plaintiff's AI agent's evidence, the defendant's AI agent's pleadings, and the defendant's AI agent's evidence. The judge's AI agent then invokes the plaintiff's lawyer's AI agent and the defendant's lawyer's AI agent to determine the disputed information based on the core factual evidence set using a knowledge base of cross-examination experience. Based on the plaintiff's response document data, the first task information of the plaintiff's lawyer is determined, and based on the defendant's response document data, the second task information of the defendant's lawyer is determined. The plaintiff's lawyer and the defendant's lawyer are then invoked to conduct adversarial debate on the disputed information according to the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information. The judge's intelligent agent performs legal reasoning and generates formatted judgment documents based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base.

[0021] In one possible implementation, the step of setting different system prompts for different agents to obtain a set of role agents includes: Different system prompts are set for different intelligent agents to obtain a set of role-based intelligent agents. The system prompts include basic system prompts and feature-based system prompts. The basic system prompts for the judge intelligent agent include a specific set of instructions for process status progression, summarizing points of contention, and generating the final judgment. The basic system prompts for the plaintiff's lawyer intelligent agent and the defendant's lawyer intelligent agent include differentiated content on the allocation of the burden of proof, selection of cross-examination strategies, and norms for debate rhetoric. The basic system prompts for the plaintiff's intelligent agent and the defendant's intelligent agent include format norms and information scope limitations for expressing case claims. The system prompts for the judge's AI agent include gender, years of practice, and the region and field of practice; the system prompts for the plaintiff's lawyer AI agent and the defendant's lawyer AI agent include gender, years of practice, historical win rate, and the region and field of practice; the system prompts for the plaintiff's AI agent and the defendant's AI agent are empty.

[0022] In one possible implementation, the legal text knowledge base vectorizes legal texts and establishes a legal text association network to provide the intelligent agent with accurate legal basis retrieval; the cross-examination experience knowledge base stores and encodes cross-examination patterns and rebuttal strategies for preset evidence; the debate experience knowledge base contains rhetoric templates and logical frameworks for various debate scenarios; and the judge trial experience knowledge base learns the key points of historical cases and sentencing rules through machine learning models.

[0023] In one possible implementation, the step of invoking the plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base, the cross-examination experience knowledge base, and the debate experience knowledge base in the integrated knowledge base, in order to obtain first argument data corresponding to the first task information and second argument data corresponding to the second task information, includes: Obtain the query statements of the plaintiff's lawyer AI agent and the defendant's lawyer AI agent regarding the disputed information; The query statement is rewritten into an optimized query statement containing the knowledge base domain identifier using a domain adaptive model; Based on the optimized query statement, dense vector retrieval and sparse retrieval are performed in parallel, and preliminary results are obtained from the legal text knowledge base and the debate experience knowledge base through the similarity function; The preliminary results are sorted using an attention-based re-ranking model to generate the final retrieval set; Based on the optimized query statement and the final retrieval set, prompt context data is constructed, and the plaintiff's lawyer agent and the defendant's lawyer agent are invoked to generate first argument data and second argument data according to the prompt context data.

[0024] In one possible implementation, the step of invoking the plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, in order to obtain first argument data corresponding to the first task information and second argument data corresponding to the second task information, includes: Calculate the real-time information state changes and criterion alignment based on the first and second argument data; When the real-time information status change meets the first preset condition and the criterion alignment meets the second preset condition, the step of calling the plaintiff lawyer intelligent agent and the defendant lawyer intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and debate experience knowledge base in the integrated knowledge base to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information is stopped. Calculate the first semantic similarity of the first argument data output by the plaintiff's lawyer intelligent agent in the adversarial debate process. When the first semantic similarity meets the preset condition, stop calling the plaintiff's lawyer intelligent agent. Calculate the second semantic similarity of the second argument data output by the defendant's lawyer agent during the adversarial debate process. When the second semantic similarity meets a preset condition, stop calling the defendant's lawyer agent.

[0025] In one possible implementation, the step of performing legal reasoning and generating a formatted judgment document by the judge's intelligent agent based on the core set of factual evidence, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base includes: The judge's intelligent agent uses a legal reasoning model to reason based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base, and obtains the reasoning result; Based on the reasoning results and the judgment document generation module, a formatted judgment document is generated.

[0026] In one possible implementation, the method further includes: Obtain user request information and determine the target intelligent agent based on the user request information; When the target intelligent agent is a judge intelligent agent, the case statement information of the plaintiff intelligent agent and the defendant intelligent agent regarding the case to be tried, the litigation response data of the plaintiff intelligent agent, the evidence information of the plaintiff intelligent agent, the litigation response data of the defendant intelligent agent, and the evidence information of the defendant intelligent agent are displayed as display information to obtain the initial information set, factual evidence, and disputed information for the displayed information; Based on the plaintiff's response document data, the first task information of the plaintiff's lawyer's intelligent agent is determined. Based on the defendant's response document data, the second task information of the defendant's lawyer's intelligent agent is determined. The plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent are invoked to conduct adversarial debate on the disputed information according to the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information. Using a legal reasoning model, legal reasoning is performed based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and the judge's trial experience knowledge base in the integrated knowledge base to generate a formatted judgment document. The first argument data and the second argument data regarding the disputed information are displayed to obtain feedback data on the first argument data and the second argument data. The formatted judgment document and the feedback data are matched to obtain a score for the feedback data.

[0027] For example, a complete mock court system was constructed based on the collaborative work of several key technologies, including multi-agent collaborative adversarial architecture, personalized role modeling, multi-source knowledge base integration, dynamic convergence control, automatic judgment generation, and human-computer interaction feedback. This system constructs a standardized four-stage trial process, defining a collaborative adversarial framework for multiple intelligent agents, including judges, court clerks, plaintiffs, defendants, and their lawyers. It shapes differentiated behavioral patterns by injecting characteristic data from real legal practitioners; provides professional support to the agents through the integration of four knowledge bases: legal provisions, evidence examination experience, debate experience, and judges' trial experience; achieves intelligent control of the debate stage through a dual convergence mechanism based on information state monitoring and semantic similarity analysis; automatically generates standardized judgment documents through legal reasoning based on multi-source information fusion; and allows users to participate in trials as lawyers and obtain evaluation feedback through an immersive human-computer interaction interface. Ultimately, it realizes a legal simulation system integrating full-process simulation, personalized roles, professional knowledge, intelligent debate, automated judgment, and closed-loop training. Its aim is to solve problems in existing mock court systems such as homogenized agent roles, lack of dynamic adversarial interaction in real legal scenarios, weak human-computer interaction functions, and difficulty in achieving a complete "practice-feedback-optimization" closed loop. See also Figure 1 As shown, the present invention includes the following steps: In one possible implementation, a four-stage trial framework driven by state sequence and with information aggregated step by step is established, which structurally decomposes the complete litigation process into four core stages: trial preparation, court investigation, court debate, and final statement and trial.

[0028] During the trial preparation stage, the judge's intelligent agent initializes the trial, and the plaintiff and defendant's intelligent agents present the case in parallel. The initial case information set I_initial is constructed through computation.

[0029] The exchange of complaint documents was completed during the court investigation phase. Answer) and Evidence Presentation and Examination (Evidence_plaintiff) Evidence_defendant), through (Complaint) Answer) (Evidence_plaintiff Evidence_defendant constructs the core set of factual evidence, E_facts.

[0030] During the court debate phase, the judge's intelligent agent summarizes the points of contention based on I_initial and E_facts. Guide both lawyers to conduct Confrontational debate.

[0031] During the final statement and sentencing phase, integrate all process information T_total, through... The legal reasoning function automatically generates formatted judgment documents, completing a closed-loop process.

[0032] In one possible implementation, a multi-agent role system is constructed, including judges, court clerks, plaintiffs, plaintiffs' lawyers, defendants, and defendants' lawyers, to establish a complete judicial simulation subject with six types of roles.

[0033] Configure differentiated system prompts for each role, and define the judge's process control authority, the court clerk's recording responsibilities, the plaintiff and defendant's obligation to state facts, and the lawyer's authority to argue and represent.

[0034] The system prompts precisely define the roles of judges ("ensuring the fair implementation of the law"), court clerks ("ensuring procedural compliance and complete record-keeping"), plaintiffs ("the initiator of litigation, the defendants ("the respondent of litigation, the defenders"), plaintiffs' lawyers ("planning litigation strategies"), and defendants' lawyers ("developing defense strategies").

[0035] In one possible implementation, such as Figure 2 As shown, an adaptive decision-making mechanism based on information status monitoring is established, which monitors changes in information status in real time. Alignment with criteria A decision function is constructed by combining the court hearing history H_t and the role / task T_role. .

[0036] Construct intelligent termination judgment logic, when the information gain of multiple consecutive rounds of debate... ( (Information gain threshold) and viewpoint alignment ( When the alignment threshold is reached, the system automatically determines that the core argument has been fully elaborated and terminates the speech.

[0037] Establish a semantic evolution analysis mechanism using the cosine similarity function. Continuously calculate the semantic similarity of adjacent statements to detect semantic repetition patterns.

[0038] Set strict cyclical prevention conditions, when When established, it is determined that the system is trapped in a semantic repetition loop and its speaking rights are automatically terminated.

[0039] In one possible implementation, a database of legal practitioners' characteristics is established, extracting multi-dimensional characteristics from authoritative platforms, such as lawyers' gender, years of practice, success rate, practice location, and areas of expertise, as well as judges' gender, years of trial experience, plaintiff's success rate, court level, and cause of action.

[0040] Build for lawyers Feature vectors and judge construction Feature vectors are used to set professional parameter standards such as years of practice, quantification of win rate, and division of court levels.

[0041] By using a dual-channel attention mechanism to deeply couple feature vectors with basic role prompts, personalized enhanced prompts are generated, ensuring that features have a targeted impact on the decision-making mechanism.

[0042] Based on the fusion results, differentiated behavior is shaped, enabling lawyers with high win rates to demonstrate innovative viewpoints and senior judges to exhibit personalized characteristics such as a preference for procedural justice.

[0043] In one possible implementation, such as Figure 3 As shown, an integrated knowledge architecture based on a legal knowledge graph is constructed, which includes four core knowledge base modules: legal provisions, cross-examination experience, debate experience, and judge's trial experience.

[0044] By establishing a legal provision association network through legal concept entity recognition and relation extraction technologies, and using the BGE-m3 model to vectorize the content of the provisions, a multi-layered index retrieval system for the legal provision knowledge base is constructed.

[0045] Based on the key points of evidence examination extracted from the court hearing transcripts, examination strategies were stored according to the cause of action; based on the argumentation framework of debate theory and case summary; and based on the trial model mining of judgment documents, three experience knowledge bases were constructed respectively.

[0046] By integrating various knowledge bases through a unified retrieval enhancement generation framework, it supports semantic retrieval, similarity retrieval, and relationship retrieval, and establishes an incremental learning mechanism based on case feedback to achieve continuous system evolution.

[0047] In one possible implementation, such as Figure 4 As shown, the original query q is rewritten into an optimized query through a domain-adaptive model in the query understanding phase. It supplements professional terms based on legal dictionaries and semantic rules, and identifies intents and routes them to the corresponding knowledge base through the query classification module.

[0048] Dense vector retrieval is performed in parallel during the hybrid retrieval phase. Sparse Search Based on the principle that legal provision queries emphasize sparse retrieval and case queries emphasize vector retrieval, an adaptive weighted fusion is performed to form a preliminary result set. .

[0049] A multi-level attention re-ranking model based on word-level attention (key term matching), paragraph-level attention (logical integrity assessment), and document-level attention (comprehensive consideration) is adopted. Perform fine sorting to generate the final search set. .

[0050] A dynamic caching mechanism is constructed to create a cache index for high-frequency queries and fine-ranking results. Based on the feedback from court hearings, the query understanding model and re-ranking model are continuously optimized through a reinforcement learning mechanism, updating the dictionary, adjusting weights, and optimizing attention configuration.

[0051] In one possible implementation, a complete dataset of case information and knowledge information is constructed. It integrates evidence sets, debate summaries, case law provisions, trial experience, and historical information throughout the entire process.

[0052] The analysis includes fact-finding (constructing an evidence association graph model to assess credibility), legal application (matching legal provisions with algorithms to map constituent elements), liability determination (calculating the degree of fault and causal relationship), and sentencing assessment (conducting a balanced test by referring to case standards).

[0053] Based on structured templates that conform to document format standards, the reasoning results are transformed into professional legal texts, automatically generating standardized judgment documents that include information about the parties, litigation claims, factual findings, points of contention, applicable laws, and the main text of the judgment, ensuring authority and persuasiveness.

[0054] In one possible implementation, such as Figure 5 As shown, it provides a case self-setting interface, supporting users to customize core elements such as case background, points of contention, evidence materials, and litigation claims, and extracts key information through intelligent analysis technology to build a personalized case library.

[0055] It enables role selection and pre-trial preparation, automatically generates complaints / answers, lists of evidence, and preliminary legal opinions based on case information set by the user, and provides case analysis tools to help formulate litigation strategies.

[0056] The system supports users to participate in court hearings in real time as lawyers through a natural language interaction interface. Users can make statements, present evidence, cross-examine witnesses, and argue based on self-defined cases. The system dynamically generates adversarial responses that conform to the logic of the case.

[0057] Establish a multi-dimensional performance evaluation system to monitor users' performance in dimensions such as the accuracy of their application of case facts, compliance with legal procedures, logical organization of evidence, persuasiveness of legal arguments, and effectiveness of litigation strategies, and generate personalized evaluation reports.

[0058] A complete learning loop is formed by recording user performance data, comparing the results of multiple training sessions, visually displaying the progress trajectory, and adjusting strategies based on evaluation results, thereby achieving a continuous improvement cycle of "case setting - practical training - evaluation feedback - strategy optimization".

[0059] On the other hand, this application provides a case adjudication system for multi-agent cooperative confrontation, the system comprising: The module is used to set different system prompts for different intelligent agents to obtain a set of role intelligent agents, which includes a judge intelligent agent, a plaintiff intelligent agent, a plaintiff's lawyer intelligent agent, a defendant intelligent agent, and a defendant's lawyer intelligent agent. Based on the set of role intelligent agents and an integrated knowledge base based on a legal knowledge graph architecture, a court hearing environment is constructed. The integrated knowledge base includes a legal provisions knowledge base, an evidence examination experience knowledge base, a debate experience knowledge base, and a judge's trial experience knowledge base. The first calculation module is used to form an initial information set based on the case statements of the plaintiff and defendant in the trial environment by the judge's intelligent agent. The judge's intelligent agent constructs a core factual evidence set based on the plaintiff's response documents, the plaintiff's evidence information, the defendant's response documents, and the defendant's evidence information. The judge's intelligent agent then calls the plaintiff's lawyer and the defendant's lawyer's intelligent agents to determine the disputed information based on the core factual evidence set according to the cross-examination experience knowledge base. The second calculation module is used to determine the first task information of the plaintiff's lawyer intelligent agent based on the plaintiff's intelligent agent's response document data, and to determine the second task information of the defendant's lawyer intelligent agent based on the defendant's intelligent agent's response document data. The module then calls upon the plaintiff's lawyer intelligent agent and the defendant's lawyer intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information. The output module is used to perform legal reasoning and generate formatted judgment documents by the judge's intelligent agent based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base.

[0060] In one possible implementation, this application provides a terminal device, including: a memory, a processor, and a first computer program stored in the memory and executable on the processor. When the processor executes the first computer program, it implements the steps of a case trial method for multi-agent collaborative confrontation.

[0061] In one possible implementation, this application provides a computer-readable storage medium storing a second computer program, which, when executed by a processor, implements the steps of a multi-agent cooperative adversarial case trial method.

[0062] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0063] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments 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. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0064] 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 medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0065] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0066] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0067] In the embodiments provided in this application, it should be understood that the disclosed apparatus / network devices and methods can be implemented in other ways. For example, the apparatus / network device embodiments described above are merely illustrative. For instance, the division of modules or 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 interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0068] 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.

[0069] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

[0070] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. A method for handling cases involving multi-agent cooperative adversarial proceedings, characterized in that, Includes the following steps: Different system prompts are set for different intelligent agents to obtain a set of role intelligent agents, which includes judge intelligent agent, plaintiff intelligent agent, plaintiff's lawyer intelligent agent, defendant intelligent agent, and defendant's lawyer intelligent agent. A court hearing environment is constructed based on the set of role intelligent agents and an integrated knowledge base based on a legal knowledge graph architecture. The integrated knowledge base includes a legal provisions knowledge base, an evidence examination experience knowledge base, a debate experience knowledge base, and a judge's trial experience knowledge base. In the courtroom setting, the judge's AI agent forms an initial information set based on the case statements of the plaintiff's AI agent and the defendant's AI agent. The judge's AI agent then constructs a core factual evidence set based on the plaintiff's AI agent's written response data, the plaintiff's AI agent's evidence information, the defendant's AI agent's written response data, and the defendant's AI agent's evidence information. The plaintiff's lawyer's AI agent and the defendant's lawyer's AI agent access a knowledge base of cross-examination experience to determine the disputed information based on the initial information set and the core factual evidence set. Based on the plaintiff's response document data, the first task information of the plaintiff's lawyer is determined, and based on the defendant's response document data, the second task information of the defendant's lawyer is determined. The plaintiff's lawyer and the defendant's lawyer are then invoked to conduct adversarial debate on the disputed information according to the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information. The judge's intelligent agent performs legal reasoning and generates formatted judgment documents based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base.

2. The case adjudication method for multi-agent cooperative confrontation according to claim 1, characterized in that, The step of setting different system prompts for different agents to obtain a set of role agents includes: Different system prompts are set for different intelligent agents to obtain a set of role-based intelligent agents. The system prompts include basic system prompts and feature-based system prompts. The basic system prompts for the judge intelligent agent include a specific set of instructions for process status progression, summarizing points of contention, and generating the final judgment. The basic system prompts for the plaintiff's lawyer intelligent agent and the defendant's lawyer intelligent agent include differentiated content on the allocation of the burden of proof, selection of cross-examination strategies, and norms for debate rhetoric. The basic system prompts for the plaintiff's intelligent agent and the defendant's intelligent agent include format norms and information scope limitations for expressing case claims. The system prompts for the judge's AI agent include gender, years of practice, and the region and field of practice; the system prompts for the plaintiff's lawyer AI agent and the defendant's lawyer AI agent include gender, years of practice, historical win rate, and the region and field of practice; the system prompts for the plaintiff's AI agent and the defendant's AI agent are empty.

3. The case adjudication method for multi-agent cooperative confrontation according to claim 1, characterized in that, The legal text knowledge base vectorizes legal texts and establishes a legal text association network to provide intelligent agents with accurate legal basis retrieval; the cross-examination experience knowledge base stores and encodes cross-examination patterns and rebuttal strategies for preset evidence; the debate experience knowledge base contains rhetoric templates and logical frameworks for various debate scenarios; and the judge trial experience knowledge base learns the key points of historical judgments and sentencing rules through machine learning models.

4. The case adjudication method for multi-agent cooperative confrontation according to claim 1, characterized in that, The step of invoking the plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information, includes: Obtain the query statements of the plaintiff's lawyer AI agent and the defendant's lawyer AI agent regarding the disputed information; The query statement is rewritten into an optimized query statement containing the knowledge base domain identifier using a domain adaptive model; Based on the optimized query statement, dense vector retrieval and sparse retrieval are performed in parallel, and preliminary results are obtained from the legal text knowledge base and the debate experience knowledge base through the similarity function; The preliminary results are sorted using an attention-based re-ranking model to generate the final retrieval set; Based on the optimized query statement and the final retrieval set, prompt context data is constructed, and the plaintiff's lawyer agent and the defendant's lawyer agent are invoked to generate first argument data and second argument data according to the prompt context data.

5. The case adjudication method for multi-agent cooperative confrontation according to claim 1, characterized in that, The step of invoking the plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information, includes: Calculate the real-time information state changes and criterion alignment based on the first and second argument data; When the real-time information status change meets the first preset condition and the criterion alignment meets the second preset condition, the step of calling the plaintiff lawyer intelligent agent and the defendant lawyer intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and debate experience knowledge base in the integrated knowledge base to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information is stopped. Calculate the first semantic similarity of the first argument data output by the plaintiff's lawyer intelligent agent in the adversarial debate process. When the first semantic similarity meets the preset condition, stop calling the plaintiff's lawyer intelligent agent. Calculate the second semantic similarity of the second argument data output by the defendant's lawyer agent during the adversarial debate process. When the second semantic similarity meets a preset condition, stop calling the defendant's lawyer agent.

6. A method for handling cases involving multi-agent cooperative adversarial proceedings according to claim 1, characterized in that, The step of generating a formatted judgment document by having the judge's intelligent agent perform legal reasoning based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base includes: The judge's intelligent agent uses a legal reasoning model to reason based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base, and obtains the reasoning result; Based on the reasoning results and the judgment document generation module, a formatted judgment document is generated.

7. A case adjudication method for multi-agent cooperative confrontation according to claim 1, characterized in that, The method further includes: Obtain user request information and determine the target intelligent agent based on the user request information; When the target intelligent agent is a judge intelligent agent, the case statement information of the plaintiff intelligent agent and the defendant intelligent agent regarding the case to be tried, the litigation response data of the plaintiff intelligent agent, the evidence information of the plaintiff intelligent agent, the litigation response data of the defendant intelligent agent, and the evidence information of the defendant intelligent agent are displayed as display information to obtain an initial information set and factual evidence for the displayed information. The cross-examination experience database is called according to the plaintiff lawyer intelligent agent and the defendant lawyer intelligent agent, based on the disputed information of factual evidence. Based on the plaintiff's response document data, the first task information of the plaintiff's lawyer's intelligent agent is determined. Based on the defendant's response document data, the second task information of the defendant's lawyer's intelligent agent is determined. The plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent are invoked to conduct adversarial debate on the disputed information according to the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information. Using a legal reasoning model, legal reasoning is performed based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and the judge's trial experience knowledge base in the integrated knowledge base to generate a formatted judgment document. The first argument data and the second argument data regarding the disputed information are displayed to obtain feedback data on the first argument data and the second argument data. The formatted judgment document and the feedback data are matched to obtain a score for the feedback data.

8. A case adjudication system based on multi-agent collaborative adversarial mechanisms, characterized in that: The system includes: The module is used to set different system prompts for different intelligent agents to obtain a set of role intelligent agents, which includes a judge intelligent agent, a plaintiff intelligent agent, a plaintiff's lawyer intelligent agent, a defendant intelligent agent, and a defendant's lawyer intelligent agent. Based on the set of role intelligent agents and an integrated knowledge base based on a legal knowledge graph architecture, a court hearing environment is constructed. The integrated knowledge base includes a legal provisions knowledge base, an evidence examination experience knowledge base, a debate experience knowledge base, and a judge's trial experience knowledge base. The first calculation module is used to form an initial information set based on the case statements of the plaintiff and defendant in the trial environment by the judge's intelligent agent, and to construct a core factual evidence set by the judge's intelligent agent based on the plaintiff's intelligent agent's written response data, the plaintiff's intelligent agent's evidence information, the defendant's intelligent agent's written response data, and the defendant's intelligent agent's evidence information. The plaintiff's lawyer's intelligent agent and the defendant's lawyer's intelligent agent call the cross-examination experience knowledge base to determine the disputed information based on the initial information set and the core factual evidence set. The second calculation module is used to determine the first task information of the plaintiff's lawyer intelligent agent based on the plaintiff's intelligent agent's response document data, and to determine the second task information of the defendant's lawyer intelligent agent based on the defendant's intelligent agent's response document data. The module then calls upon the plaintiff's lawyer intelligent agent and the defendant's lawyer intelligent agent to conduct adversarial debate on the disputed information based on the legal provisions knowledge base and the debate experience knowledge base in the integrated knowledge base, so as to obtain the first argument data corresponding to the first task information and the second argument data corresponding to the second task information. The output module is used to perform legal reasoning and generate formatted judgment documents by the judge's intelligent agent based on the core factual evidence set, the first argument data, the second argument data, and the legal provisions knowledge base and judge's trial experience knowledge base in the integrated knowledge base.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the multi-agent cooperative adversarial case handling method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the multi-agent cooperative adversarial case handling method as described in any one of claims 1 to 7.

Images