An interactive three-dimensional simulated court trial litigation training system and method

Abstract

1. TECHNICAL FIELD AND PROBLEM: The present application is an interactive three-dimensional simulated court trial litigation training system and method, belonging to the field of legal artificial intelligence and three-dimensional judicial simulation. Traditional artificial court trial training is high in cost, low in frequency and low in simulation degree; ordinary simulation system template is solidified and cannot adapt to real cases, lacking dynamic confrontation and real-time error correction function.2. CORE SCHEME: The system combines legal multi-modal large model, multi-agent architecture and three-dimensional rendering engine, relies on ten million judicial data to build vector base and fine-tune the model. Three agents analyze the case, dynamically generate a three-dimensional court confrontation scene, real-time correct answers, cross-examine, process errors, support segmented and unlimited times of training, and automatically generate a review report.3. BENEFICIAL EFFECTS: The three-dimensional scene highly restores the real courtroom, realizes individual case customization training, reduces the cost of tutoring, improves the user's court trial response, cross-examination and on-the-spot response ability, and adapts to civil and commercial, administrative dispute court trial practical teaching.

Description

Technical Field

[0001] This invention belongs to the fields of legal artificial intelligence, intelligent three-dimensional simulated court hearings, judicial human-machine confrontation exercises, and multimodal big data simulation technology. Specifically, it relates to an interactive simulated court hearing litigation exercise system and method based on a legal vertical multimodal big model, a multi-agent collaborative architecture, and judicial vector retrieval technology. It is mainly applicable to the simulation of court hearing processes, interactive confrontation exercises, court hearing defense and error correction, and review and improvement of litigation capabilities in various types of litigation cases such as civil, commercial, administrative, and labor disputes. It can provide full-process intelligent court hearing practical training for ordinary litigants, legal personnel, and trainee lawyers. Background Technology

[0002] Courtroom proceedings are the core of litigation, and the standardization of the trial process, the rigor of the examination of evidence, the logic of the defense, and the ability to adapt to changing circumstances directly determine the outcome of the case. Currently, ordinary litigants generally lack professional legal knowledge and practical experience in court proceedings. They are completely unfamiliar with courtroom procedures, judges' questioning rules, the norms of evidence examination, the logic of courtroom debate, and courtroom etiquette, and have weak ability to respond independently.

[0003] Traditional methods of court trial preparation and rehearsal have significant technical shortcomings and industry deficiencies. Traditional court trial rehearsals rely heavily on one-on-one human guidance from practicing lawyers. This human guidance model is costly, has limited training sessions, fixed timeframes, and cannot facilitate frequent retraining. Furthermore, the simulated scenarios are simplified, lack adversarial elements, and have low fidelity to real court trials. They cannot simulate unexpected questioning, cross-examination, and debate in real court trials, resulting in poor training effectiveness.

[0004] Meanwhile, most existing mock court systems on the market are fixed template-based process demonstrations, supporting only standardized process playback and static teaching displays. They lack personalized case adaptation, dynamic human-computer interaction, intelligent attack and defense interaction, and error identification and correction mechanisms, making it impossible to conduct targeted and customized drills based on the user's actual litigation claims, evidence materials, points of contention, and case facts.

[0005] Meanwhile, most existing mock court systems on the market are fixed template-based process demonstrations, supporting only standardized process playback and static teaching displays. They lack personalized case adaptation, dynamic human-computer interaction, intelligent attack and defense interaction, and error identification and correction mechanisms, making it impossible to conduct targeted and customized drills based on the user's actual litigation claims, evidence materials, points of contention, and case facts.

[0006] Existing technologies cannot capture user defense loopholes, cross-examination errors, logical deviations, and procedural mistakes in real time, and cannot complete intelligent error correction and targeted optimization. This leads to parties being nervous in court, responding incoherently, omitting key points in cross-examination, making logical errors in their defense, and violating procedural rules, which can easily cause adverse consequences in court and seriously affect the quality of case trials.

[0007] In summary, the industry currently lacks an intelligent simulated court trial training technology solution that can adapt to real cases, provide highly realistic adversarial scenarios, interactive 3D scene drills, intelligent error correction, automatic review, and unlimited self-training. There is an urgent need for an intelligent, data-driven, and model-based interactive court trial training system to address the aforementioned technical pain points. Summary of the Invention

[0008] This invention aims to address the industry's technical bottlenecks in existing trial preparation models, such as high labor costs, low training frequency, poor scenario reproduction, lack of intelligent adversarial capabilities, lack of personalized adaptation, and lack of automatic error correction and review. It also solves the problems of traditional manual drills relying on professional lawyers, inability to conduct high-frequency training, and low simulation levels. Furthermore, it addresses the technical shortcomings of existing template-based simulated trial systems, such as fixed procedures, inability to adapt to real cases, lack of interactive attack and defense, inability to correct practical errors in trial proceedings, and inability to specifically improve users' litigation response capabilities.

[0009] It enables ordinary users to independently complete the entire process of personalized courtroom adversarial drills without the accompaniment of a lawyer. This is highly realistic, interactive, error-correctable, and reviewable, significantly lowering the threshold for litigation practice, avoiding problems such as errors in courtroom defense, omissions in cross-examination, logical errors, and procedural irregularities, and comprehensively improving users' courtroom response capabilities and on-the-spot adaptability.

[0010] This invention provides an interactive simulated court trial litigation training system and method. It relies on a legal vertical multimodal large model, a judicial multimodal data base of tens of millions of data points, vector retrieval technology and multi-agent collaborative architecture to build a full-process intelligent court trial simulation training system. The system includes five major architectures: underlying data base, model training layer, multi-agent interaction layer, court trial simulation training layer, and intelligent error correction and review layer.

[0011] The underlying data platform reuses massive judicial compliance data resources, including over 1.2 million currently effective legal provisions, over 60 million judicial judgment cases, a vast amount of real court transcripts, court question-and-answer data, evidence presentation standards data, and multimodal litigation evidence sample data. The original judicial data undergoes cleaning, deduplication, anonymization, standardization, and filtering of flawed data, unifying the multimodal data format. Legal terminology is segmented using BPE and WordPiece dual-mode word segmentation technology. A BERT embedding model is used to perform unified semantic embedding on text, image, and audio / video evidence multimodal data, generating high-dimensional, dense judicial semantic vectors, which are then batch-stored into a vector database. This constructs a dedicated judicial vector retrieval and logical reasoning platform for court hearings, supporting millisecond-level precise retrieval of similar court hearing scenarios, similar evidence presentation logic, and similar points of contention in court hearings.

[0012] The model training layer is based on the Transformer multimodal large model architecture and adopts a fine-tuning training paradigm specifically for legal trial scenarios. It relies on massive amounts of real trial sample data for iterative training, specifically learning court trial procedures, judges' questioning habits, trial evidence examination rules, evidence acceptance standards, the logic of plaintiff-defendant litigation, courtroom debate rhetoric, and the logic for handling unexpected trial scenarios. Through hundreds of thousands of real trial test samples, the model's accuracy is verified and parameters are optimized, ensuring the professionalism, standardization, and realism of the simulated trials, highly aligning with the logic of actual court trials.

[0013] The system is configured with three core legal intelligent agents to work together to complete interactive court trial simulations: a compliance identification intelligent agent, a multimodal evidence recognition intelligent agent, and a document defense simulation intelligent agent.

[0014] The compliance verification intelligence is used to pre-verify the legality, relevance, and authenticity of the litigation claims, case facts, and all submitted evidence in a user's case. It automatically identifies the core points of contention, procedural risks, and weaknesses in the litigation, providing underlying logical support for customized court trial drills.

[0015] The multimodal evidence recognition intelligent agent automatically analyzes all evidence materials uploaded by users, including text, images, audio recordings, and videos. It intelligently sorts out the complete evidence chain, key points of evidence examination, evidence flaws, and entry points for examination and defense, generating a case-specific examination exercise question bank and attack and defense logic.

[0016] The document defense simulation intelligent agent has multi-role simulation capabilities, which can dynamically simulate the roles of judge, plaintiff and defendant in court hearings. Based on the user's real case facts, it can dynamically generate court hearing questions, cross-examination, court debates and sudden interrogations in court hearings, realizing fully interactive, dynamic and non-template-based court hearing attack and defense drills.

[0017] The system is equipped with a real-time intelligent error correction mechanism. Throughout the exercise, the large model captures user defense loopholes, cross-examination errors, logical deviations, procedural operational mistakes, and non-standard courtroom rhetoric in real time. It provides immediate pop-up corrections and dynamic prompts for standard courtroom response norms and cross-examination strategies, correcting user errors in courtroom practice in real time.

[0018] The system is equipped with an interactive adaptation module, which allows users to freely switch between any trial stage, such as court investigation, court debate, and final statement. It supports unlimited repeated drills, single-stage intensive training, and full-process continuous drills, and adapts to the specific improvement needs of different users' weaknesses.

[0019] The system has a built-in fully automatic debriefing report generation module that records, documents, and performs structured analysis of the entire user's rehearsal process. After the rehearsal, it automatically generates a professional court hearing debriefing report, accurately marking weaknesses in the defense, omissions in the cross-examination, logical errors, procedural flaws, and wording issues. It also outputs standardized optimized wording, standardized cross-examination plans, court hearing process rectification suggestions, and special improvement plans.

[0020] Once deployed, the model can retrieve real court cases with similar causes of action and similar points of contention from the vector database in real time, dynamically iterate and optimize the adversarial logic and questioning methods in court, continuously improve the simulation fit, and realize personalized interactive court trial drills for each case.

[0021] It enables non-template-based, personalized, and highly realistic 3D interactive courtroom adversarial drills, completely breaking away from the traditional fixed-process demonstration mode. It can customize exclusive courtroom adversarial scenarios based on the user's real case facts, evidence, claims, and points of contention, highly restoring the atmosphere of a real courtroom trial, the judge's questioning logic, and the details of the offensive and defensive confrontation between the two sides, with extremely high simulation.

[0022] It enables low-cost, unlimited, autonomous 3D court trial simulations, eliminating reliance on one-on-one human guidance from practicing lawyers, significantly reducing the time and economic costs of court trial training. Users can conduct full-process training anytime and repeatedly, making it suitable for high-frequency capability enhancement.

[0023] It has real-time intelligent error correction and fully automatic review and optimization capabilities, which can accurately identify various errors in court defense, cross-examination, and process operation, correct them in real time and output standardized solutions, thereby avoiding problems such as on-the-spot mistakes, omissions in cross-examination, and non-standard responses in real court hearings from the root.

[0024] It comprehensively enhances users' courtroom response capabilities and on-the-spot adaptability. Through intelligent human-machine adversarial training, users become proficient in courtroom procedures, evidence examination rules, debate logic, response techniques, and emergency scenario handling skills, greatly improving their performance in real courtroom proceedings and the probability of winning cases.

[0025] It possesses strong value in popularizing legal knowledge and providing practical teaching, lowers the threshold for ordinary people to practice litigation, and enables users without legal background to quickly master professional court trial norms and litigation logic. It achieves intelligent, inclusive, and professional litigation practice training and has extremely high social promotion value. Detailed Implementation

[0026] Constructing a dedicated multimodal judicial data foundation for court hearings. This involves batch collecting massive amounts of laws and regulations, judicial cases, authentic court transcripts, court Q&A, evidence presentation guidelines, and multimodal litigation evidence data. The raw data is cleaned, deduplicated, anonymized, and invalid or flawed data is removed. A unified multimodal data storage format is implemented. Dual-mode word segmentation technology is used to process legal text data, and the BERT model is used to complete the semantic embedding of multimodal judicial data, generating standardized judicial semantic vectors, which are stored in a vector database. This establishes a court hearing judicial data foundation capable of real-time retrieval, inference, and matching.

[0027] Complete the fine-tuning training of a large-scale multimodal model specifically designed for legal trials. A basic large-scale model was built based on the Transformer architecture, employing a fine-tuning paradigm specific to trial scenarios. Iterative training was conducted using massive amounts of real trial samples, learning judges' discretionary logic, trial procedure norms, cross-examination logic, debating techniques, and strategies for handling unexpected trial situations. The model's accuracy was verified and parameters were optimized through real trial test samples, and the model was then deployed in the cloud and adapted locally.

[0028] Multi-agent initialization and collaborative startup. The system starts the compliance verification agent, the multimodal evidence recognition agent, and the document defense simulation agent, waiting for the user to import case materials and receiving the case facts, litigation claims, and all multimodal evidence materials uploaded by the user.

[0029] Pre-trial intelligent analysis and focus identification. The compliance verification agent automatically verifies the legality of case claims and evidence, identifies procedural risks and substantive disputes, and identifies the core points of contention in the case; the multimodal evidence identification agent analyzes all evidence content, sorts out the evidence chain structure, key points of cross-examination, and flaws in the evidence, and generates a courtroom offense and defense logic system tailored to the case.

[0030] Conduct interactive simulated courtroom adversarial drills. The simulated intelligent agent dynamically switches between the roles of judge and opposing party, randomly generating courtroom questions, cross-examination, debates, and unexpected inquiries based on the actual circumstances of each case. Users respond, cross-examine, and debate in real time, completing the entire process of an interactive courtroom adversarial trial. The model monitors user operations and responses in real time throughout the process, identifying errors and providing dynamic error correction prompts.

[0031] The system offers flexible adaptation to various training modes. Users can choose from full-process continuous drills, single-stage specialized training, or repetitive reinforcement training based on their weaknesses. The system dynamically adapts the training difficulty and intensity to specifically address the user's shortcomings.

[0032] The system automatically generates court trial debriefing reports and optimization plans. After the exercise, the system performs structured analysis on the data from the entire exercise, summarizing user errors in responses, omissions in cross-examination, logical deviations, procedural flaws, and non-standard wording issues. It generates a visual and professional debriefing report, providing standardized court trial response templates, cross-examination optimization plans, suggestions for court trial procedure standardization, and specific improvement strategies for each issue.

[0033] The model is continuously iterated and optimized. The system accumulates the data from this exercise in real time and continuously optimizes the model's adversarial logic and simulation accuracy by combining it with similar court cases in the vector database. This achieves the iterative effect of the system becoming more accurate with use and the exercise scenarios becoming more and more similar to real court trials.

[0034] Through the above implementation process, this invention stably realizes intelligent three-dimensional simulated court trial training functions such as personalized customization, high-fidelity adversarial training, real-time error correction, automatic review, and unlimited autonomous training, effectively solving various technical shortcomings of traditional court trial training models and comprehensively improving users' practical ability to respond to court trials.

Claims

1. An interactive three-dimensional simulated court trial litigation training system and method, characterized in that, include: The system comprises a foundational judicial data base, a model training layer, a multi-agent interaction layer, a 3D court trial simulation layer, and an intelligent error correction and review layer. The underlying judicial data platform is used to collect and organize judicial data, including legal provisions, judicial cases, court transcripts, court Q&A, evidence presentation guidelines, and multimodal litigation evidence data. The platform cleans, deduplicates, desensitizes, removes defective data from the original judicial data, and standardizes its format. It uses BPE and WordPiece dual-mode word segmentation technology to segment legal terms, and employs the BERT embedding model to semantically embed multimodal data, generating judicial semantic vectors which are stored in a vector database. This forms the foundation for court-specific judicial vector retrieval and logical reasoning. The model training layer is based on the Transformer multimodal large model architecture and adopts a court trial scenario-specific fine-tuning training paradigm. It relies on massive real court trial samples for iterative training to learn court trial procedure norms, judges' questioning habits, cross-examination rules, evidence acceptance standards, the adversarial logic of both parties in litigation, and strategies for dealing with unexpected events in court trials. The model accuracy is verified and the model parameters are optimized through real court trial samples. The multi-agent interaction layer includes a compliance verification agent, a multimodal evidence recognition agent, and a document defense simulation agent. The compliance verification agent is used to verify the legality, relevance, and authenticity of the case's claims, facts, and evidence, and to identify the focus of the dispute and litigation risks. The multimodal evidence recognition agent is used to analyze the text, images, audio, and video evidence uploaded by users, and to sort out the evidence chain, key points of cross-examination, evidence flaws, and points of attack and defense. The document defense simulation agent is used to dynamically simulate the roles of the judge, plaintiff, and defendant in court, and dynamically generate court questioning, cross-examination, court debate, and unexpected inquiries based on real cases. The 3D courtroom simulation training layer is used to construct an immersive three-dimensional courtroom scene. It dynamically generates personalized, non-template-based interactive courtroom attack and defense training scenarios based on the user's real cases, and supports users to freely switch between courtroom stages, single-stage specialized training, and unlimited repeated training. The intelligent error correction and review layer is used to identify user defense loopholes, cross-examination errors, logical deviations and process operation mistakes in real time and dynamically and intelligently correct them. It records the entire exercise data, automatically performs structured analysis after the exercise and generates a court trial review and optimization report, outputting standardized defense statements, cross-examination optimization plans and process rectification suggestions.

2. The interactive three-dimensional simulated court trial litigation training system according to claim 1, characterized in that, The underlying judicial data base stores more than 1.2 million currently effective laws and regulations, more than 60 million judicial judgment cases, and a massive amount of court trial practice sample data, supporting millisecond-level retrieval of similar court trial scenarios, similar cross-examination logic, and disputed points of similar cases.

3. The interactive three-dimensional simulated court trial litigation training system according to claim 1, characterized in that, The model training layer uses hundreds of thousands of real court trial test samples to verify the model's accuracy and optimize parameters, ensuring that the model's output logic conforms to the real court trial rules and judges' trial habits.

4. The interactive three-dimensional simulated court trial litigation training system according to claim 1, characterized in that, The compliance identification intelligence can automatically identify procedural risks, substantive disputes, and litigation weaknesses in cases, providing underlying case logic support for customized trial drills.

5. The interactive three-dimensional simulated court trial litigation training system according to claim 1, characterized in that, The multimodal evidence recognition intelligent agent can automatically generate a unique cross-examination question bank and courtroom attack and defense logic system for each user's case, adapting to the cross-examination training needs of different case types.

6. The interactive three-dimensional simulated court trial litigation training system according to claim 1, characterized in that, The document defense simulation intelligent agent can randomly generate scenarios such as sudden questions in court, in-court cross-examination and confrontation, and debate, realizing dynamic, unscripted, real-life courtroom confrontation drills.

7. The interactive three-dimensional simulated court trial litigation training system according to claim 1, characterized in that, The three-dimensional courtroom simulation training layer achieves a 1:1 three-dimensional courtroom scene restoration, providing an immersive and visual courtroom experience, completely breaking away from the traditional two-dimensional template-based demonstration mode.

8. The interactive three-dimensional simulated court trial litigation training system according to claim 1, characterized in that, The intelligent error correction and review layer provides real-time pop-up prompts throughout the entire exercise process, indicating the error type, error cause, and standard court trial operation procedures, and corrects deviations in the user's court trial practice in real time.

9. The interactive three-dimensional simulated court trial litigation training system according to claim 1, characterized in that, The system can accumulate data from previous drills and continuously iterate and optimize the model's adversarial logic and simulation accuracy by combining similar trial cases from a vector database, thus enabling the model to adaptively upgrade.

10. An interactive three-dimensional simulated court trial litigation exercise method, characterized in that, The system applied to any one of claims 1 to 9 includes the following steps: S1. Construct a multimodal judicial data foundation specifically for court hearings, and complete data cleaning, word segmentation, semantic embedding, and vector input. S2. Based on the Transformer architecture, complete the fine-tuning training and parameter optimization deployment of a large multimodal model specifically for court hearings; S3. Initiate the collaborative work of the three intelligent agents to receive case facts, litigation requests, and multimodal evidence materials uploaded by users; S4. Automatically analyze case details, verify evidence, identify points of contention, and construct a case-specific trial defense system through intelligent agents; S5. Load a 3D immersive courtroom scene to dynamically conduct interactive courtroom questioning, cross-examination, debate and confrontation drills and provide real-time intelligent error correction. S6. Conduct full-process drills or single-stage specialized reinforcement training according to user needs; S7. After the exercise, a structured debriefing report and a complete set of trial optimization plans will be automatically generated. S8. Accumulate and refine training data to achieve continuous iteration and optimization of the model, thereby improving the realism and professionalism of courtroom simulation.