Multi-role adaptive court trial intelligent auxiliary system and method thereof

Abstract

The application relates to the fields of intelligent court, artificial intelligence and legal information technology, and discloses a court trial intelligent auxiliary system and method with multi-role adaptation. The system comprises an identity recognition module, a core processing engine module and a role adaptation module. The core processing engine module is integrated with a domestic large language model and an autonomous intelligent agent. The system provides differentiated views and functions for judges, lawyers and parties, has core capabilities such as lie detection and question and answer scheme generation, lawyer practice risk early warning, party old-age barrier-free assistance and the like, adopts localized deployment to ensure data security, all outputs are auxiliary suggestions, and the final decision-making right belongs to the judge. The application solves the problems of backward technology, single service, poor practicality of functions, data security risks and the like of the existing trial auxiliary system, realizes full-process intelligent assistance, takes into account judicial efficiency and fairness, and meets the requirements of intelligent court and justice construction.

Description

Technical Field

[0001] This invention relates to a multi-role adaptive intelligent auxiliary system and method for court trials, belonging to the fields of smart courts, artificial intelligence and legal information technology. Background Technology

[0003] Existing courtroom assistance technologies mostly employ the BERT+Bi-LSTM keyword matching approach. While some solutions have achieved voice keyword capture and basic lie detection, they still suffer from several core shortcomings: First, insufficient technical capabilities, lacking large-scale deep semantic understanding and autonomous intelligent agent task planning and execution capabilities, limiting them to fixed-process retrieval and matching. Second, limited service coverage, primarily targeting judges and court clerks, leaving lawyers and litigants without equivalent intelligent assistance tools, leading to information asymmetry among litigation participants. Third, insufficient functional practicality, as some litigants repeatedly lie in court, and false statements necessitate post-trial verification or even reconvening of the trial. The existing lie detection systems waste valuable and scarce judicial resources. They only output results and cannot generate actionable questioning sequences and rebuttal strategies. Furthermore, they fail to achieve seamless data integration across the entire process from case filing to pre-trial and trial. Fourth, they lack adaptability. They do not design exclusive interactive modes for vulnerable groups such as the elderly and those with language impairments, nor do they provide lawyers with evidence reading and verification functions or professional risk warnings, which can easily lead to trial delays and wasted scheduling. Fifth, their security and boundaries are unclear. Most solutions do not adopt domestic large-scale models and localized deployment, posing data security risks. Moreover, the systems tend to output conclusive judgments and do not clearly define their "auxiliary" role, thus infringing on the judge's final decision-making power. Meanwhile, the contradiction of "too many cases and too few judges" is prominent in judicial practice. The average number of cases handled per judge at the grassroots level remains high, and they face difficulties such as cumbersome evidence review, difficulty in identifying false statements, and lack of follow-up questioning strategies. Lawyers often change cases at the last minute due to scheduling conflicts and insufficient pre-trial preparation, which affects the quality of the trial. In order to ascertain the facts and ensure the quality of the judgment, the court may have to hold a second trial. At the same time, vulnerable groups have difficulty participating in litigation on an equal footing due to lack of legal knowledge and physical limitations. All of the above problems urgently need to be systematically solved through intelligent means. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by providing a multi-role adaptive intelligent auxiliary system and method for court trials, meeting practical application requirements.

[0005] To solve the above problems, the technical solution adopted by the present invention is as follows: A multi-role adaptive intelligent court trial assistance system includes an identity recognition module, a core processing engine, a role adaptation module, a multi-terminal function module, a party-side selection control module, a standardized access module for litigation materials, an oral narration assistance module, a trial response assistance module, a lie detection and question-and-answer scheme generation module, and a local deployment and data security module. The output of the identity recognition module is connected to the input of the role adaptation module, and is used to identify the current user's role type as judge, lawyer, or party. The core processing engine is bidirectionally connected to the role adaptation module and various functional modules, providing the entire system with semantic analysis, task planning and data processing capabilities. The output of the role adaptation module is connected to the multi-terminal function module and is used to load differentiated function views according to the identified role type. The output of the standardized litigation materials access module is connected to the core processing engine, which is used to connect to the official litigation platform and parse standardized electronic materials. The oral narration assistance module, court hearing response assistance module, and lie detection and question-and-answer scheme generation module are all bidirectionally connected to the core processing engine, respectively realizing barrier-free interaction, real-time court hearing response assistance, contradiction identification and strategy generation functions; The party-side selection control module is connected to the party-side terminal of the multi-terminal functional module and is used to control the activation and deactivation of the intelligent assistance function on the party-side terminal. The local deployment and data security module is connected to all modules of the system to achieve localized data storage and compatibility with domestically produced software and hardware. All system outputs are marked as auxiliary suggestions, and the final decision-making authority belongs to the judges.

[0006] As an improvement to the above technical solution: the core processing engine includes a domestically developed large language model module and an autonomous intelligent agent module; The output of the domestic large language model module is connected to the input of the autonomous intelligent agent module, and is used to complete semantic understanding of case materials, evidence association analysis, text generation and contradiction semantic comparison. The autonomous intelligent agent module includes a task planning submodule, a multi-step execution submodule, and a failure retry submodule. The task planning submodule receives user natural language instructions and breaks them down into multi-level subtasks. The multi-step execution submodule calls the corresponding tools to complete the subtasks. The failure retry submodule is used to execute retry and rollback strategies when the subtask fails.

[0007] As an improvement to the above technical solution: the multi-terminal functional module includes a judge terminal module, a lawyer terminal module, and a party terminal module; The judge terminal module is used to realize automatic prompting of disputed points, verification of the integrity of the chain of evidence, accurate push of similar cases, and generation of draft judgments. The lawyer terminal module is used to generate cross-examination opinions, retrieve rebuttal evidence, and generate draft legal statements. It also integrates functions such as evidence sufficiency analysis, statement consistency verification, and false statement risk warning. The party terminal module is used to visualize and guide the litigation process, provide simplified explanations of evidence, and analyze the expected outcomes of the case.

[0008] As an improvement to the above technical solution: the lie detection and question-and-answer solution generation module includes a data source verification submodule, a contradiction detection submodule, a question-and-answer solution generation submodule, and an evidence linkage submodule; The output of the data source verification submodule is connected to the contradiction detection submodule to verify that the input data is only court hearing audio and video recordings, pre-trial transcripts, evidence materials and official case information. The contradiction detection submodule is used to compare the court statements, pre-trial transcripts and evidence materials to identify factual contradictions, quantity and amount contradictions and timeline contradictions and calculate the confidence level. The question-and-answer scheme generation submodule is connected to the contradiction detection submodule and is used to generate a multi-level follow-up question sequence and court rebuttal strategy based on the contradiction points. The evidence linkage submodule is connected to the question-and-answer scheme generation submodule and the judge terminal module, respectively, and is used to automatically highlight the corresponding evidence documents and key paragraphs when displaying follow-up questions and strategies.

[0009] As an improvement to the above technical solution: the party-side selection control module includes a user profile submodule, an access control submodule, and a switch control submodule; The user profiling submodule is used to collect information on case type, age of the parties involved, physical condition, representation status, and litigation integrity records. The permission management submodule is connected to the user profiling submodule and is used to determine whether the person has permission to use auxiliary functions according to preset rules. The switch control submodule is connected to the permission management submodule and the party terminal module, and is used for the court or judges to manually / automatically turn on or off the intelligent auxiliary function of the party terminal.

[0010] As an improvement to the above technical solution: the standardized access module for litigation materials includes a platform interface submodule, a material parsing submodule, and a structured storage submodule; The platform interface submodule is bidirectionally connected to the People's Court Online Service Platform for automatic synchronization of electronic litigation materials; The material analysis submodule is connected to the platform interface submodule and is used to identify materials with a unified naming format and extract case elements and evidence information. The structured storage submodule is connected to the material analysis submodule and the core processing engine, and is used to store the analyzed materials in a structured form for the core processing engine to quickly retrieve and call.

[0011] As an improvement to the above technical solution: the oral assistance module includes a voice interaction submodule, a semantic transcription submodule, a document generation submodule, and an age-friendly submodule; The voice interaction submodule is used to receive user verbal instructions and case statements, and also supports the output content of the voice broadcast system. The semantic transcription submodule is connected to the voice interaction submodule and is used to convert voice information into text and perform semantic error correction. The document generation submodule is connected to the semantic transcription submodule and the core processing engine, and is used to generate a structured complaint, a list of evidence, and a draft of the answer based on the oral content. The age-friendly submodule is connected to the system's front-end interface and is used to provide large fonts, high-contrast display, and simplified operation processes.

[0012] Specifically: A method for assisting a multi-role-adaptive intelligent court trial system includes the following steps: Step 1: User identity authentication and role recognition. Users log in to the system through the unified court authentication platform. The identity recognition module verifies the user's identity and determines their role as a judge, lawyer, or party. Step 2: Standardized access and preprocessing of litigation materials. The standardized access module for litigation materials synchronizes with the electronic materials on the People's Court Online Service Platform, and completes parsing, element extraction and structured storage. Step 3: Role-adapted view loading and function initialization. The role-adaptation module loads the corresponding function interface according to the user role, and the core processing engine preloads case-related data and model resources. Step 4: Pre-trial intelligent assistance and risk warning. The system provides judges with evidence chain verification and dispute focus sorting, lawyers with evidence reading reminders and statement consistency verification, and eligible parties with oral statement assistance and process guidance. Step 5: Real-time semantic analysis and contradiction detection in court proceedings. The system transcribes court proceedings audio in real time. The lie-debunking and question-and-answer generation module verifies the legality of the data source, compares the statements made in court with the evidence on file, identifies contradictions, and calculates confidence levels. Step 6: Differentiated intelligent assistance output. The system generates a sequence of follow-up questions and rebuttal strategies for judges, real-time evidence examination suggestions for lawyers, and response suggestions for parties who have activated the assistance function. All outputs are marked "System suggestion, for reference". Step 7: Local data storage and audit log retention. All court proceedings data and system operation records are stored on the court's local server. Lawyer risk warnings and system auxiliary operations generate tamper-proof audit logs.

[0013] Compared with the prior art, the implementation effects of the present invention are as follows: 1. Achieve a leapfrog upgrade in technical capabilities: By adopting a domestically developed large language model combined with an independent intelligent agent architecture, it breaks through the technical limitations of traditional keyword matching, possesses deep semantic understanding and autonomous task execution capabilities, and can automatically complete complex multi-step tasks such as evidence chain verification and case element extraction. The response speed and analysis accuracy are greatly improved, solving the problem of low intelligence level in existing systems.

[0014] 2. Constructing a fair empowerment system for all roles: For the first time, it achieves differentiated intelligent assistance covering judges, lawyers, and litigants, breaking down the technical barriers of only serving judges and eliminating information asymmetry among litigation participants; at the same time, through the litigant's right of choice control mechanism, it provides assistance to vulnerable groups and restricts litigants with false litigation records, thus achieving precise judicial assistance.

[0015] 3. Enhance the substantive nature and efficiency of court hearings: The innovative lie-debunking and question-and-answer generation functions can not only identify false statements, but also output a logically rigorous multi-level questioning sequence, rebuttal strategies, and corresponding evidence to help judges expose lies in court and reduce the need for second hearings and appeals; the evidence sufficiency detection and false statement risk warning functions designed for lawyers effectively avoid court hearings that are not completed due to insufficient preparation by lawyers or last-minute changes of lawyers, saving the court's tight scheduling resources.

[0016] 4. Implement judicial humanistic care: Through functions such as oral assistance, large font high contrast interface, and process visualization guidance, solve the difficulties of litigation participation for vulnerable groups such as the elderly, people with language impairments, and people with mobility difficulties, lower the threshold for litigation, improve the accessibility of justice, and make judicial services more humane.

[0017] 5. Safeguarding judicial accountability and data security: Clearly define the system's "auxiliary" role, with all outputs being advisory in nature, and the final decision-making power rests entirely with the judge, in accordance with the requirements of the judicial accountability system; adopt a fully localized deployment solution, adapt to domestic hardware and domestic large-scale models, ensuring that case data does not leave the court, meeting the requirements of information technology innovation compliance, and fundamentally eliminating the risk of data leakage.

[0018] 6. Achieve seamless integration and efficient expedited trials: Streamline the data flow across the entire process from case filing to pre-trial proceedings and trial, avoiding redundant data entry and information silos; enable intelligent assistance for simplified procedure cases, which can significantly shorten trial time, improve the efficiency of expedited trial cases, and effectively alleviate the contradiction of "too many cases and too few judges" in courts. Attached Figure Description

[0019] Figure 1 : The system architecture diagram described in this invention; Figure 2 : Flowchart of the multi-role adaptation process described in this invention; Figure 3 : A schematic diagram of the judge's interface described in this invention; Figure 4 : A schematic diagram of the lawyer's client interface described in this invention; Figure 5 : A schematic diagram of the functional interface of the party-side device described in this invention; Figure 6 : Flowchart of the party-side selection right control as described in this invention; Figure 7 : Workflow diagram of the oral communication assistance module described in this invention; Figure 8 The flowchart of the lie detection and question-and-answer solution generation module described in this invention. Detailed Implementation

[0020] The present invention will now be described in conjunction with specific embodiments.

[0021] A detailed implementation of a multi-role adaptive intelligent auxiliary system for court trials and its method; The present invention will be further described in detail below with reference to Figures 1-8 and specific embodiments, but the scope of protection of the present invention is not limited thereto. Equivalent substitutions, parameter adjustments, framework changes, etc., made by those skilled in the art based on the technical concept of the present invention all fall within the scope of protection of the present invention.

[0022] Example 1: System Overall Architecture and Domestic Multi-Terminal Deployment This embodiment provides the basic deployment architecture and operating environment of the intelligent court hearing assistance system described in this invention, referring to... Figure 1 The system architecture diagram shown indicates that the entire system is deployed on a physical server cluster within the court, adopting a fully localized offline deployment solution. All case data, model parameters, and operation logs are stored locally within the court, strictly prohibiting data leakage and meeting the compliance requirements for information technology innovation.

[0023] 1.1 Hardware and Software Infrastructure Hardware layer: Prioritize compatibility with domestically produced servers, such as Huawei Ascend 910 / 310 inference servers and Kunpeng 920 general-purpose servers, with ≥32 CPU cores, ≥128GB memory, and ≥2TB high-speed solid-state drives for model and data storage; for high-concurrency court hearing scenarios, distributed cluster deployment can be adopted, supporting horizontal scaling.

[0024] Model layer: The core reasoning engine adopts domestic large language models, such as DeepSeek-14B / 32B INT8 quantized version and Wenxin Yiyan local deployment version. After quantization, the model reasoning speed is improved by 3-5 times, and the response time for single evidence association analysis is ≤2 seconds. The vector database uses Milvus or Chroma to store the semantic vectors of legal provisions, cases, and case elements, and supports millisecond-level similarity retrieval.

[0025] Intelligent Agent Layer: Based on open-source intelligent agent frameworks (such as OpenClaw), autonomous intelligent agent modules are built, integrating three core capabilities: task planning, tool invocation, and failure retry. They can automatically decompose and execute complex multi-step tasks according to natural language instructions.

[0026] Front-end layer: Supports three access methods: web browser, Windows client, and domestic operating system (Tongxin UOS, Kylin OS) client, and is compatible with the court's existing office terminals; the mobile terminal only supports case progress inquiry and notification reception, and does not open core court hearing auxiliary functions.

[0027] 1.2 System Initialization and Authentication Process Users can log in to the system through the court's unified identity authentication platform, which supports three authentication methods: account password, digital certificate, and facial recognition.

[0028] The identity recognition module calls the court case management system interface to verify the user's identity and match their role in the current case (judge / lawyer / party), and generates a role permission token.

[0029] The role adaptation module loads the corresponding functional views and permission configurations based on the token: the judge's side loads the full trial management functions, the lawyer's side loads the agency-related functions, and the party's side loads the corresponding auxiliary functions based on the permission configuration.

[0030] The core processing engine preloads all electronic litigation materials for the current case, completes the extraction of case elements, the structured storage of evidence, and the construction of semantic vectors, in preparation for subsequent pre-trial and trial support.

[0031] Example 2: Analysis of the sufficiency of evidence reading and early warning of the risk of false statements from the lawyer's perspective This embodiment addresses the problem of insufficient pre-trial preparation by lawyers and the resulting trial delays due to last-minute changes in lawyers. It details the complete implementation process of evidence reading, detection, and risk warning on the lawyer's end, referring to the schematic diagram of the lawyer's end interface shown in Figure 4. This embodiment uses a construction contract dispute case as an example: Due to a scheduling conflict with the original lawyer, Attorney Zhang, the defendant's attorney, temporarily took over the case one day before the trial. The plaintiff submitted a total of 20 pieces of electronic evidence.

[0032] 2.1 Evidence Reading Status Detection Rules The system uses front-end tracking technology to collect multi-dimensional data on every evidence review activity of lawyers. These data collection dimensions include: File open status: Record the first opening time and the last closing time of the evidence file to determine whether there is any evidence that was never opened.

[0033] Effective reading time: Excluding time spent minimizing files and switching to other pages, this counts the actual time lawyers spend on the evidence page.

[0034] Content scrolling depth: Records the ratio of the maximum scroll position of the scroll bar to the total height of the file to determine whether there is shallow reading behavior that only browses the beginning.

[0035] Key paragraph coverage: The system automatically identifies key paragraphs in the evidence (such as core contract clauses, acceptance conclusions, and signature and seal pages) and determines whether the paragraph has been scrolled into the user's field of view.

[0036] Interaction behavior log: Records whether lawyers add annotations, highlight, or favorite evidence, serving as an auxiliary basis for in-depth reading and judgment.

[0037] The system employs multi-level judgment rules, and a system is deemed insufficiently read if any of the following conditions are met: Rule A (Completely Unread): Evidence documents show no record of being opened; Rule B (Too Short Reading Time): Effective reading time is less than the preset threshold (5 seconds for simple evidence, 30 seconds for complex contracts / expert opinions; thresholds can be configured by the court administrator according to case type). Rule C (Shallow Browsing): Scroll depth less than 25% (configurable); Rule D (Key Paragraph Omission): At least one key paragraph in the evidence has not been viewed.

[0038] 2.2 Automatic identification mechanism for key evidence The system automatically identifies key evidence in a case through multi-rule fusion. The priority of the identification rules, from highest to lowest, is as follows: Judge's marking rules: Evidence that the judge manually marks as "requiring special attention" during the pre-trial review of the case file; The opposing party's marking rules: Evidence marked as "key evidence" in the system by the opposing party or their lawyer; Document type rules: High-priority document types include contracts, agreements, IOUs, receipts, signed receipts, expert opinions, notarized documents, and effective judgments; Element density rule: The number of key elements extracted from the evidence, such as amount, date, signature, location, and liability for breach of contract, exceeds the preset threshold (default 5).

[0039] In this case, the system automatically identified E-015 (project acceptance form) and E-018 (WeChat chat records) submitted by the plaintiff as key evidence.

[0040] 2.3 Risk Warning Process Pre-trial reading reminder: After logging into the system, Attorney Zhang's lawyer's homepage automatically displays the "Evidence Reading Status Dashboard," showing "12 / 20 documents read, 8 documents unread (including 2 key pieces of evidence)," and a pop-up reminder at the top: "You have 8 pieces of evidence yet to be read. Among them, evidence E-015 (project acceptance form) and E-018 (WeChat chat records) have been marked as key evidence by the plaintiff. It is recommended to review them first before the trial. Failure to fully read key evidence may lead to unfavorable cross-examination during the trial."

[0041] Real-time consistency verification of statements: When Attorney Zhang entered "We have completed all the projects as agreed in the contract and there are no quality issues" in the statement editing area, the system performed a semantic comparison of the statement with all the evidence in real time.

[0042] High-risk warning triggered: The system detected a direct factual contradiction between this statement and the content in evidence E-015, which states "Three quality issues were found during acceptance: ① wall cracks; ② floor sanding; ③ drainage pipe blockage," and immediately displayed a red risk warning box. High-risk warning (confidence level: 0.92) Your proposed statement: We have completed all the work as stipulated in the contract, and there are no quality issues. Conflicting evidence: Project Acceptance Form E-015 (Page 2, Lines 3-5) Type of contradiction: Direct factual contradiction Recommendation: Please verify the evidence before making any statements. False statements may result in legal liability for obstructing civil litigation.

[0043] Audit log retention: The system automatically records the timestamp of this warning, user ID, case ID, contradictory evidence number, confidence score, and subsequent user actions (such as modifying statements / ignoring warnings). The log data is stored using blockchain technology, which is tamper-proof and cannot be deleted. Only the court's discipline inspection and supervision department can access it.

[0044] 2.4 Contradiction Detection and Confidence Calculation The system supports the identification of four types of contradictions and calculates the confidence level of contradictions using a weighted algorithm: Direct factual contradiction: The statement is completely contrary to the facts recorded in the evidence (e.g., "the goods were not received" versus "there is a signed receipt"). Indirect factual contradiction: There is a logical conflict between the statement and the facts recorded in the evidence (such as "being present throughout the entire process" and "the flight ticket record shows that the person was in another city on that day"). Quantity / Amount Discrepancy: The stated quantity and amount are inconsistent with the evidence (e.g., "owed 100,000 yuan" versus "IOU recorded 150,000 yuan"). Timeline contradictions: The order of events stated does not match the chain of evidence (e.g., "payment before delivery" versus "delivery note date is earlier than payment date").

[0045] The confidence level is calculated using the following formula: Confidence level = w1 × Evidence authority score + w2 × Direct contradiction score + w3 × Entity matching score; Among them, w1, w2, and w3 are configurable weight coefficients (each defaults to 1 / 3); the evidence authority score is determined according to the evidence type (notarized document 1.0, expert opinion 0.9, documentary evidence 0.8, electronic data 0.7, witness testimony 0.5); the directness of contradiction score (direct contradiction 1.0, indirect contradiction 0.6); and the entity matching score (complete entity match 1.0, partial match 0.5). The system presets a high-risk threshold of 0.7 and a medium-risk threshold of 0.4, corresponding to red and yellow warnings, respectively.

[0046] 2.5 Privacy and Compliance Protection Lawyers' reading behavior data, draft legal statements, and risk warning records are only displayed to the lawyers themselves and are not proactively pushed to the judges. Judges may only request access to a lawyer's file review status and warning records with the lawyer's written consent or as otherwise provided by law. All lawyer-related data will be retained for 3 years after the case is concluded, and will be automatically destroyed upon expiration. It may not be used for any purpose other than the case hearing.

[0047] Example 3: Lie Detection and Question-and-Answer Solution Generation Module This embodiment addresses the challenges of identifying false statements and lacking effective questioning strategies in court proceedings. It details the technical implementation of lie detection and question-and-answer generation, referring to the module workflow diagram shown in Figure 8. This embodiment uses a sales contract dispute trial as an example: the defendant stated in court, "I have never received the plaintiff's goods."

[0048] 3.1 Data source validity verification This module strictly limits the scope of data sources, using only the following data legally obtained by the court, and resolutely rejects any audio or video materials privately recorded by the parties involved: The entire court hearing was audio and video recorded (in accordance with Article 1 of the "Provisions of the Supreme People's Court on Audio and Video Recording of Court Hearings"). Pre-trial mediation record, interrogation record, and evidence exchange record; All evidentiary materials submitted by the parties (confirmed by the court after examination). Historical case information and litigant integrity records in the court case management system.

[0049] 3.2 Real-time Conflict Detection Process Real-time transcription of court proceedings: The system uses speech recognition technology to transcribe court proceedings into text in real time, marking the speaker's identity and timestamp.

[0050] Semantic vector encoding: The defendant's in-court statement is encoded into a semantic vector, and all evidence materials and pre-trial transcripts are encoded into vectors and stored in a vector database.

[0051] Similarity comparison and entity extraction: Calculate the cosine similarity between the statement vector and the evidence vector, and extract the core entities in the statement (time: March 2024, location: defendant's warehouse, behavior: receiving goods).

[0052] Conflict Detection: When the semantic similarity is below 0.3 and the entity overlap is below 0.2, the system determines that there is a contradiction. In this case, the system comparison found: Evidence E-023 (logistics receipt) shows that the defendant signed for the goods at the warehouse on March 15, 2024; The pre-trial mediation record, page 3, lines 12-15, states that the defendant admitted to "receiving some of the goods".

[0053] Confidence Calculation: Based on the authority of the evidence (signed receipt 0.8, mediation record 0.9), the directness of the conflict (1.0), and the matching degree of the entities (1.0), the system calculates the confidence level of this conflict to be 0.94, and marks it as "highly suspicious".

[0054] 3.3 Generation of Multi-Level Question Answering Schemes The system automatically generates a logically progressive sequence of follow-up questions and courtroom rebuttal strategies based on points of contradiction. The follow-up questions are divided into four levels to ensure that lies are gradually exposed: Level 1: Confirmation of basic facts: Start with undisputed basic facts to avoid direct confrontation, such as "Was a shipment delivered to your warehouse on March 15, 2024?"

[0055] Level 2: Evidence Details Verification: Confirm key information in the evidence to solidify its validity, such as "Is the signature on this logistics receipt numbered WL-20240315 yours?"

[0056] Level 3: Questioning on contradictions: Directly point out the contradictions between the party's statement in court and other evidence, and require the party to provide a reasonable explanation. For example, "You clearly stated in the pre-trial mediation on April 10, 2024, that you received 3 boxes of goods, which is inconsistent with your statement today. Please explain why."

[0057] Level 4: Legal Consequences Warning: Inform the party concerned of the legal responsibility for making false statements, such as "If you insist on denying receiving the goods, the court will initiate a handwriting identification procedure in accordance with the law. The identification fee shall be prepaid by the applicant. If the identification conclusion proves that the signature is your own, you will bear the identification fee and legal responsibility for obstructing civil litigation."

[0058] Simultaneously, generate rebuttal strategy suggestions: It is recommended that the judge present the original E-023 (logistics receipt) in court and require the defendant to verify and sign it; It is suggested that the court clerk read aloud the contents of lines 12-15 on page 3 of the pre-trial mediation record for the defendant to confirm. If the defendant still denies it, inform him of the legal consequences of making a false statement in court and record it.

[0059] 3.4 Evidence Linkage and Output Standards When a judge clicks on any follow-up question, the system automatically highlights the corresponding evidence document and key paragraphs in the evidence preview area, eliminating the need for the judge to manually search for them. All outputs are marked "System suggestion, for reference only," and the final decision on whether to adopt them rests with the judge. The system outputs data in a standardized JSON format, including fields such as conflict point ID, party information, original statement, confidence level, questioning sequence, rebuttal strategy, and integrity assessment, which facilitates subsequent data statistics and analysis.

[0060] Example 4: Multi-step task execution of the autonomous intelligent agent module This embodiment illustrates how the autonomous intelligent agent module can autonomously plan and execute complex multi-step tasks based on user natural language instructions, solving the problem that traditional systems can only execute fixed processes.

[0061] 4.1 Core Architecture of Intelligent Agents The autonomous intelligent agent module comprises three sub-modules: Task planning submodule: Based on the semantic understanding capabilities of domestic large language models, it breaks down the user's natural language instructions into multi-level executable subtasks and generates a task execution flowchart; Multi-step execution submodule: According to the task flowchart, the corresponding tools (document extraction tool, evidence verification tool, graph construction tool, etc.) are called in sequence to complete the sub-tasks; The failure retry submodule: When a subtask fails, the retry strategy is automatically executed (up to 3 retries, with intervals of 1 second, 3 seconds, and 5 seconds respectively); if the retry still fails, the rollback strategy is executed, returning the completed part of the result and marking the incomplete items, prompting the user to provide additional information.

[0062] 4.2 Task Execution Example: Evidence Chain Integrity Verification The judge inputs the command into the system: "Please help me verify whether the plaintiff's chain of evidence in this case is complete." The intelligent agent automatically executes the following 7 sub-tasks: Extracting the plaintiff's claims: Using document information extraction tools, extract the core facts claimed by the plaintiff from the complaint (such as "the defendant received the goods on March 15, 2024", "the total price of the goods was RMB 150,000", "the defendant failed to pay for the goods") and the corresponding claims.

[0063] Extracting the defendant's response: Use document information extraction tools to extract the facts admitted and denied by the defendant and the reasons for their defense from the response.

[0064] Obtain the evidence list for this case: Use the evidence search tool to obtain a list of all submitted evidence and their types.

[0065] Evidence legality verification: Use evidence verification tools to check the legality of the form of each piece of evidence (such as whether there is an original, whether the time limit for presenting evidence has been exceeded, and whether it conforms to the legal form).

[0066] Constructing a Fact-Evidence Relationship Graph: By using a graph construction tool, each fact claimed by the plaintiff is associated with its corresponding evidence, generating a visual relationship graph.

[0067] Identify gaps in the chain of evidence: Use gap analysis tools to compare the facts claimed by the plaintiff with the relevant evidence and identify factual points that are not supported by evidence (such as "the plaintiff claims that the defendant's overdue payment resulted in a penalty of RMB 20,000, but failed to submit the basis for calculating the penalty").

[0068] Generate verification report: Call the report generation tool to generate the "Evidence Chain Integrity Verification Report", which includes an evidence list, legality verification results, fact-evidence relationship diagram, evidence chain gap analysis and supplementary evidence suggestions.

[0069] The entire execution process requires no manual intervention and takes approximately 15 seconds, which is much faster than manual verification.

[0070] Example 5: Supporting Courtroom Statements of Parties with Language Impairments This embodiment addresses the issue of special groups such as those with speech or hearing impairments having difficulty participating in court hearings independently, illustrating how the barrier-free assistance function on the parties' end is implemented.

[0071] This example uses an inheritance dispute case: The plaintiff, Li, suffers from a congenital speech disorder and is unable to express himself clearly orally, and has no legal representative. During the pre-trial review, the judge, based on Li's physical condition, activated the "Accessible Court Hearing Assistance" function on his client's end.

[0072] Pre-trial preparation: The system sends the litigation process, court rules, and the rights and obligations of the parties in the form of a combination of text and images to Mr. Li's client in advance, supporting text enlargement and voice broadcast.

[0073] Courtroom Statement Guidance: During the court hearing, when the judge asks, "Plaintiff, please state your claims and factual grounds," a structured selection interface automatically pops up on Li's client side, containing preset factual options (such as "Request to inherit the deceased's property," "The deceased passed away in January 2024," and "I am the deceased's child") as well as a custom input box.

[0074] Structured Statement Generation: Ms. Li completed her statement of facts by clicking options and handwriting input. The system automatically organized it into a standardized written statement: "Plaintiff Ms. Li requests to inherit a property located at No. XX, XX Road, XX District, XX City, registered under the name of the deceased Ms. Wang. Ms. Wang passed away on January 10, 2024. The plaintiff is Ms. Wang's biological daughter and has the legal right to inherit the property."

[0075] Judge's review and confirmation: The system sends the compiled statement text to the judge's end. After the judge confirms that it is correct, the court clerk records it in the court transcript. There is no need for the agent to relay it.

[0076] Real-time response assistance: During the court hearing, the system transcribes the questions asked by the judge or the defendant into text and displays it on Li's client's device in real time. Li can respond by selecting a preset answer or by handwriting input, and the system will automatically convert the response into voice and broadcast it to the court.

[0077] Through the above functions, individuals with speech impairments can independently complete their court statements and responses without relying on others, thus safeguarding their litigation rights.

[0078] Example 6: Control over the selection of intelligent assistance functions on the client's end This embodiment illustrates the mechanism for enabling and disabling the auxiliary functions on the party's end, achieving precise assistance to vulnerable groups and effective restrictions on malicious litigants, as shown in the selection control flowchart in Figure 6.

[0079] 6.1 Permission Determination Rules The authority to enable the intelligent assistance function on the client's end is determined by the court or the presiding judge. The system automatically determines and provides suggestions based on the following factors: Basic information of the parties involved: elderly people aged 65 or older, those with physical disabilities (difficulty in movement, speech / hearing impairment), those with low levels of education (illiterate or primary school education), and those without authorized litigation agents; the system recommends enabling full-function assistance. Litigation integrity record: If there are two or more records of false statements, a previous penalty for obstructing civil litigation, or a prior conviction for false litigation, the system recommends that the intelligent assistance function be turned off. Case types: It is recommended to enable simplified procedures and small claims procedures. For complex and difficult cases, and cases involving state secrets / personal privacy, some functions may be restricted.

[0080] 6.2 Example of starting a scene The plaintiff, Mr. Wang, is a 75-year-old elderly person living alone. He did not have a legal representative and filed a lawsuit against the defendant due to a private lending dispute. The system automatically identified Mr. Wang's age and the status of his legal representation, and prompted the judge to "enable full-function assistance on the party's end." After the judge confirmed the activation, Mr. Wang gained the following access to functions: Oral narration assists in generating the complaint and evidence list; The litigation process is visually guided, with real-time prompts for the next steps; The evidence is explained in a simplified way, converting legal terminology into easy-to-understand language; Courtroom response suggestions: Automatically generates suggested answers based on the judge's questions.

[0081] According to statistics, after enabling the auxiliary functions, the trial time in this case was shortened by about 30% compared with similar cases, and Wang successfully completed all litigation procedures.

[0082] 6.3 Example of closing the scene The defendant, Zhang, had two records of making false statements within the past year and was being sued in this case due to a sales contract dispute. The system automatically identified his litigation integrity record and prompted the judge to "disable the intelligent assistance functions on the party's end." After the judge confirmed the closure, Zhang's party's end only retained basic functions such as uploading materials, checking case progress, and receiving court notices, and no longer provided intelligent assistance functions such as response suggestions and evidence interpretation to prevent him from using the system to make false statements.

[0083] Example 7: Age-appropriate oral narration-assisted generation of litigation documents This embodiment addresses the issue of elderly people being unfamiliar with electronic devices and unable to type, explaining the workflow of the speech assistance module, referring to the speech assistance module flowchart shown in Figure 7.

[0084] This example uses Mr. Zhao, a 68-year-old plaintiff, as an example. Mr. Zhao wanted to sue his children for alimony disputes. He did not know how to type on a computer, so he used the oral assistance function through the self-service terminal in the court's litigation service hall.

[0085] Function activation: Zhao clicked the "Generate Complaint by Oral Statement" button on the self-service terminal, and the system automatically switched to the age-friendly interface (18-point large font, high-contrast black background with white text, large buttons).

[0086] Voice-guided questioning: The system guides Zhao to state case information step by step through voice, asking questions such as: "Please state your name and ID number," "Who is the person you are suing?" "How much alimony are you requesting?" and "What evidence do you have to prove that the other party has not paid alimony?"

[0087] Speech transcription and semantic error correction: The system uses offline speech recognition technology to convert Zhao's spoken content into text, and performs semantic error correction and completion through a large model. For example, "My son only gives me 200 yuan a month" is corrected to "The defendant pays the plaintiff 200 yuan alimony every month".

[0088] Document generation and modification: The system generates a structured draft complaint based on the transcribed content, including party information, claims, facts and reasons, and a list of evidence. Zhao can modify the content via voice commands, such as saying "Change the alimony to 1,000 yuan per month," and the system will automatically update the draft.

[0089] Document Confirmation and Submission: After Zhao confirmed that the complaint was correct, the system generated an electronic document, affixed an electronic signature, and automatically submitted it to the "People's Court Online Service" platform to complete the case filing application.

[0090] Example 8: Local Deployment and Performance Verification of Domestic Production This embodiment uses the actual deployment of a grassroots court as an example to verify the system's domestic adaptation capability and performance indicators.

[0091] 8.1 Deployment and Configuration Hardware: 2 Huawei Ascend 310 inference servers (each configured with an 8-core CPU, 32GB of memory, and 1 Ascend 310 AI chip), 1 Kunpeng 920 storage server (configured with a 4TB high-speed solid-state drive); Software: Tongxin UOS server operating system, DeepSeek-14B INT8 quantization model, Milvus vector database, OpenClaw intelligent agent framework; Network: The court's internal local area network, physically isolated from the internet.

[0092] 8.2 Performance Test Results The processing speed for single-case materials is as follows: ≤30 seconds for parsing and structuring 100 pages of electronic litigation materials; Evidence association retrieval response time: ≤2 seconds; Real-time speech-to-text accuracy in court proceedings: ≥98% (standard Mandarin scenario); Conflict detection response time: ≤1 second; Concurrency support capability: Simultaneously supports real-time assistance for 10 court hearings without lag or delay.

[0093] 8.3 Data Security Assurance All case data is stored using AES-256 encryption, and access requires dual authentication. The system operation log is retained throughout the entire process, recording all user login, query, modification, and export operations; Supports data backup and recovery, and automatically backs up data to offline storage media on a regular basis; The system has no external network connection interface, completely eliminating the risk of data leakage.

[0094] Example 9: Collaborative Operation of the Entire Process from Case Filing to Pre-Trial to Trial This embodiment integrates the functions of the above modules to illustrate the collaborative operation of the system throughout the entire case process, achieving integrated intelligent assistance from case filing to court hearing.

[0095] Case filing stage: Parties submit case filing applications through the "People's Court Online Service" platform, or generate a complaint through the court's self-service terminal using oral assistance. The standardized litigation materials access module automatically synchronizes electronic materials, completing parsing, element extraction, and structured storage.

[0096] Pre-trial stage: Judge's side: The system automatically sorts out the key points of the case dispute, completes the verification of the integrity of the evidence chain, and generates a pre-trial case review report; On the lawyer's end: The system monitors the lawyer's evidence reading status, issues reminders for key evidence that has not been read, and verifies the consistency between the lawyer's opinions and the evidence in real time; For the parties involved: Eligible parties receive process guidance and evidence interpretation services, and prepare materials for their court hearing responses.

[0097] Trial stage: The system transcribes court proceedings audio in real time and simultaneously detects contradictions. When a false statement is detected, a sequence of follow-up questions and rebuttal strategies are sent to the judge. Push real-time evidence presentation suggestions and legal basis to lawyers; Send response suggestions to individuals who have enabled accessibility features.

[0098] Post-trial stage: The system automatically generates a draft judgment based on the trial record and evidence materials, which the judge can revise and improve; at the same time, it generates a trial debriefing report, summarizing the focus of the case dispute and the acceptance of evidence.

[0099] Regarding the working principle of this solution: This solution primarily utilizes seven core components: identity recognition, material preprocessing, role adaptation, pre-trial assistance, real-time courtroom analysis, differentiated output, and secure data storage. It leverages the collaborative operation of these modules to achieve intelligent assistance throughout the entire courtroom process. The specific operational flow is as follows: User identity authentication and role recognition: Users log in to the system through the unified court authentication platform. The identity recognition module 1 verifies the user's identity and matches the user's role type as a judge, lawyer, or party in the current case, generates a role permission token, and transmits it to the role adaptation module 3 to complete the initial verification of user identity and permissions.

[0100] Standardized access and preprocessing of litigation materials: 6. Standardized access module for litigation materials connects to the online service platform of the People's Court. The 6.1 platform connection sub-module automatically synchronizes electronic litigation materials; 6.2 material parsing sub-module identifies materials with a unified naming format and extracts case elements and evidence information; 6.3 structured storage sub-module stores the parsed content in a structured form and transmits it synchronously to the 2 core processing engine. The 2.1 domestic large language model module completes the semantic vector construction and preloading of case materials.

[0101] Role-based view loading and function initialization: 3. The role adaptation module receives the role permission token from the identity recognition module and issues differentiated view loading instructions to the multi-terminal function modules: 4.1 The judge terminal module loads full-featured court trial management functions such as dispute focus prompts, evidence chain verification, similar case push, and draft judgment generation; 4.2 The lawyer terminal module loads agency-related functions such as generating cross-examination opinions, rebuttal basis retrieval, and draft agency statements; 4.3 The party terminal module's permissions are managed by the party terminal selection control module: 5.1 The user profile submodule collects information such as the party's age, physical condition, litigation agency status, and litigation integrity record; 5.2 The permission management submodule determines permissions according to preset rules; 5.3 The switch control submodule performs the on / off operation of auxiliary functions. Simultaneously, the autonomous intelligent agent module initializes task planning, multi-step execution, and failure retry capabilities to prepare for subsequent task execution.

[0102] Pre-trial Intelligent Assistance and Risk Warning: 2. Core Processing Engines Drive Pre-trial Assistance Work in Each Module: For judges, the 2.2 Autonomous Intelligent Agent Module receives natural language instructions from the judge, breaks them down into multi-level executable sub-tasks, and calls corresponding tools to complete them, pushing the evidence chain verification report and the results of the dispute focus analysis to the 4.1 Judge Sub-Module; For lawyers, the 4.2 Lawyer Sub-Module collects data such as evidence opening status, reading time, and scrolling depth through front-end embedding points, identifies key evidence that has not been fully read, generates reading reminders, and transmits the lawyer's input agency opinions to the 2.1 Domestic Large Language Model Module for consistency comparison with the evidence materials. When contradictions are detected, a graded risk warning is triggered based on the confidence threshold, and the warning information is recorded in an unalterable audit log; For authorized parties, the 7 Oral Assistance Module receives oral content through the 7.1 Voice Interaction Sub-Module, the 7.2 Semantic Transcription Sub-Module completes speech-to-text conversion and semantic error correction, and the 7.3 Document Generation Sub-Module combines with 2.1 The domestic large language model module generates structured indictments, evidence catalogs, and other documents, while the 7.4 age-friendly sub-module provides support for large fonts and high contrast displays.

[0103] Real-time semantic analysis and contradiction detection during court proceedings: During the proceedings, the system transcribes the court proceedings audio in real time, annotating the speaker's identity and timestamp, and transmits it to module 9 (Liar Detection and Question-and-Answer Generation); module 9.1 (Data Source Verification) verifies that the input data is only official data such as court hearing audio and video recordings, pre-trial transcripts, and evidence materials. After verification, it is transmitted to module 9.2 (Contradiction Detection); module 9.2 (Contradiction Detection) calls module 2.1 (Domestic Large Language Model) to compare the statements made in court, pre-trial transcripts, and evidence materials, identifying direct factual contradictions, indirect factual contradictions, quantitative and monetary contradictions, and timeline contradictions, and calculating confidence levels. Simultaneously, module 8 (Court Response Assistance) receives court questions in real time and calls module 2 (Core Processing Engine) to retrieve case materials and generate response suggestions.

[0104] Differentiated Intelligent Assistance Output: Three role-adaptation modules push the analysis results from the two core processing engines in a differentiated manner based on the user's role: To the 4.1 Judge sub-module, the multi-level questioning sequence and rebuttal strategies generated by the 9.3 Question-and-Answer Scheme Generation sub-module are pushed; the 9.4 Evidence Linkage sub-module simultaneously highlights the corresponding evidence documents and key paragraphs in the evidence preview area; to the 4.2 Lawyer sub-module, real-time cross-examination suggestions and legal basis are pushed; to the 4.3 Parties sub-module with the assistance function enabled, response suggestions generated by the 8 Courtroom Response Assistance module are pushed. All output content is marked "System Suggestion, for Reference," and the final decision-making power belongs to the judges.

[0105] Local data storage and audit log retention: The 10 local deployment and data security modules manage the flow of data throughout the system. All case data, model parameters, and operation logs are stored on the court's local server. Lawyer risk warning records and system-assisted operations generate tamper-proof audit logs, which are encrypted and stored and access controlled by the 10 local deployment and data security modules. After the case is concluded, the relevant data is automatically destroyed within the prescribed period.

[0106] The core innovation of this solution lies in: This paper proposes a multi-role adaptive intelligent courtroom assistance architecture that fully covers judges, lawyers, and litigants, addressing the pain point of information asymmetry among litigation participants caused by existing technologies that only serve the judge's side. Through the collaboration of 1 identity recognition module, 3 role adaptation modules, and 4 multi-terminal function modules, differentiated functional views are customized for different roles. Simultaneously, a litigant-side selection control module is introduced for the first time. Through user profile collection, permission determination, and on / off control mechanisms, full-function assistance is enabled for vulnerable groups such as the elderly and those with language impairments, while assistance functions are restricted for litigants with false litigation records. This achieves precise judicial assistance that "helps the vulnerable and regulates malice," constructing a judicial technology system that fairly empowers all roles.

[0107] For the first time, a domestically developed large language model is deeply integrated with an autonomous intelligent agent as the core processing engine, solving the pain points of existing technologies that rely on BERT+Bi-LSTM keyword matching and lack deep semantic understanding and autonomous task execution capabilities. 2.1 The domestically developed large language model module realizes semantic understanding of case materials, evidence association analysis, and contradictory semantic comparison, breaking through the limitation of traditional technologies that can only identify keywords; 2.2 The autonomous intelligent agent module, through task planning, multi-step execution, and failure retry mechanisms, can automatically decompose and complete complex multi-step tasks such as evidence chain verification and case element extraction, achieving a technological leap from "passive retrieval and matching" to "active autonomous execution".

[0108] The innovative design of the lie detection and question-and-answer strategy generation module addresses the pain point of existing lie detection technologies, which only output detection results and cannot generate implementable trial strategies. The lie detection and question-and-answer strategy generation module employs several key mechanisms: 9.1 Data Source Verification Submodule strictly limits the use of legitimate data sources to mitigate the legal risks associated with parties recording audio or video privately; 9.2 Contradiction Detection Submodule identifies contradictions from multiple dimensions and quantifies their confidence levels; 9.3 Question-and-Answer Strategy Generation Submodule generates a logically progressive multi-level questioning sequence and trial rebuttal strategies; and 9.4 Evidence Linkage Submodule enables real-time highlighting and redirection of follow-up questions and corresponding evidence, forming a complete closed loop of "contradiction identification - strategy generation - evidence support."

[0109] For the first time, the system integrates evidence sufficiency analysis and false statement risk warning functions for lawyers into the trial support system, addressing pain points such as insufficient pre-trial preparation by lawyers and trial delays and wasted scheduling caused by last-minute changes in lawyers. 4.2 The lawyer sub-module uses front-end embedded points to detect evidence reading status from multiple dimensions, automatically identifies key evidence and generates precise reading reminders by combining multiple rules; at the same time, it compares the consistency between the lawyer's proposed statements and the evidence on file in real time, calculates the confidence level of contradictions based on a weighted algorithm and triggers tiered warnings, and all warnings are only shown to the lawyer and recorded in the audit log, which protects the lawyer's professional privacy and reduces invalid trials from the source.

[0110] By clearly defining the system's "auxiliary" role and adopting a fully localized, domestically developed deployment solution, this approach addresses the pain points of existing systems where conclusive judgments infringe upon judges' final decision-making power, and the data security risks associated with the lack of domestic technology. This solution clarifies, from both institutional and technical perspectives, that all outputs are merely auxiliary suggestions, with the final decision-making power entirely resting with the judge, aligning with the core requirements of the judicial accountability system. Through 10 localized deployment and data security modules adapted to domestic hardware such as Huawei Ascend and large-scale domestic models like DeepSeek and Wenxin Yiyan, case data is ensured to remain within the court, meeting national compliance requirements for information technology innovation.

[0111] Technical effects of implementing this solution: Implementing this solution can achieve comprehensive improvements across multiple dimensions, including technological capabilities, judicial fairness, court efficiency, security and compliance, and humanistic care. Specific technical benefits are as follows: First, it achieves a leapfrog upgrade in court trial support technology capabilities, significantly alleviating the contradiction of "too many cases and too few judges" in courts. This solution adopts a core architecture combining a domestically developed large language model with an autonomous intelligent agent, replacing traditional keyword matching technology, and possesses deep semantic understanding and autonomous task execution capabilities. For example, after a judge inputs a natural language command to "verify whether the plaintiff's chain of evidence is complete," the 2.2 Autonomous Intelligent Agent module can automatically break it down into seven sub-tasks, including extracting the plaintiff's claims, extracting the defendant's defense, verifying the legality of evidence, and constructing a fact-evidence association graph, and execute them sequentially. The entire process requires no manual intervention and takes approximately 15 seconds, far faster than the efficiency of manual verification which takes several hours. Meanwhile, the 2.1 Domestic Large Language Model module has an evidence association retrieval response time of ≤2 seconds, a real-time speech-to-text accuracy rate of ≥98%, a contradiction detection response time of ≤1 second, and a time of ≤30 seconds for parsing and structuring 100 pages of electronic litigation materials for a single case. The system can simultaneously support real-time assistance for 10 court hearings without lag or delay, freeing judges from tedious administrative tasks and allowing them to focus on the determination of facts and the application of law.

[0112] Second, it constructs a comprehensive system of equal empowerment for all roles, balancing judicial efficiency and fairness. This solution is the first to achieve differentiated intelligent assistance covering judges, lawyers, and litigants, breaking down the technical barriers of serving only judges and eliminating information asymmetry among litigation participants. For lawyers, the evidence sufficiency detection and false statement risk warning functions help lawyers temporarily taking on cases quickly grasp the core facts, avoiding ineffective trials due to insufficient preparation. Actual calculations show that this function can reduce invalid trial scheduling by more than 40%, significantly saving scarce court resources. For litigants, barrier-free functions such as oral narration assistance, process visualization guidance, and simplified explanations of evidence address the litigation participation difficulties of vulnerable groups such as the elderly, those with language impairments, and those with mobility issues, lowering the litigation threshold and improving judicial accessibility. For example, a 75-year-old litigant without legal representation can generate a standardized complaint within 10 minutes through the oral narration assistance module. After enabling the trial assistance function, the trial time is shortened by approximately 30% compared to similar cases, truly implementing the principle of judicial equality.

[0113] Third, it significantly improves the substantive level of court hearings and reduces the waste of judicial resources. The innovative lie-debunking and question-and-answer generation functions of this solution can not only accurately identify false statements by parties, but also output logically rigorous multi-level questioning sequences and rebuttal strategies, and link corresponding evidence in real time. This helps judges expose lies in court, avoiding adjournments for supplementary review, appeals, and retrials due to false statements. For example, in a sales contract dispute hearing, after the system detected a contradiction between the defendant's statement that "they never received the plaintiff's goods" and the logistics receipt and pre-trial mediation record, it immediately generated a four-level questioning sequence and evidence linkage suggestions. Based on the suggestions, the judge verified the evidence in court, forcing the defendant to admit to receiving the goods. This shortened the hearing time by more than 50%, while avoiding potential subsequent handwriting identification and appeal procedures, saving a significant amount of judicial resources. In addition, the system connects the entire data link from case filing to pre-trial proceedings to trial, avoiding duplicate data entry and information silos. It provides intelligent assistance throughout the entire process for cases under simplified procedures and small claims procedures, which can shorten the trial time of a single case by an average of more than 30%, greatly improving the efficiency of handling expedited cases.

[0114] Fourth, this plan comprehensively safeguards the implementation of the judicial accountability system and data security compliance. It explicitly states that all system outputs are merely supplementary suggestions, with the final decision-making power resting entirely with the judge, eliminating the risk of the system replacing judicial judgment and aligning with the core requirements of the judicial accountability system. Simultaneously, it adopts a fully localized deployment solution, compatible with domestic hardware such as Huawei Ascend and large-scale domestic models like DeepSeek and Wenxin Yiyan. All case data and operation logs are stored on the court's local server, employing AES-256 encryption and a dual authentication mechanism. The system is physically isolated from the internet, completely eliminating the risk of data leakage. Furthermore, all operation records and risk warnings generated by the system are stored using blockchain technology, ensuring they are tamper-proof and undeletable, accessible only to the court's disciplinary and supervisory departments. This protects lawyers' professional privacy while providing data support for judicial supervision.

[0115] Fifth, this system achieves differentiated and precise judicial assistance, demonstrating judicial humanistic care. Through five party-side selection control modules, the system dynamically adjusts the permissions of auxiliary functions based on factors such as the party's age, physical condition, and litigation integrity record. It provides targeted assistance to vulnerable groups and effectively restricts malicious litigants. For example, for parties with congenital speech impairments who lack legal representation, the system provides a structured selection interface and text-to-speech functionality, enabling them to independently complete court statements and responses without relying on others. For parties with two or more records of false statements, the system automatically disables their intelligent assistance functions, retaining only basic functions such as uploading materials and checking case progress, preventing them from using the system to make false statements and achieving a harmonious balance between judicial efficiency and judicial fairness.

[0116] The above description, in conjunction with specific embodiments, provides a detailed explanation of the present invention, but it should not be construed as limiting the specific implementation of the invention to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A multi-role adaptive intelligent auxiliary system for court trials, characterized in that: It includes an identity recognition module, a core processing engine, a role adaptation module, a multi-terminal function module, a party-side selection control module, a standardized access module for litigation materials, an oral statement assistance module, a court hearing response assistance module, a lie detection and question-and-answer scheme generation module, and a local deployment and data security module. The output of the identity recognition module is connected to the input of the role adaptation module, and is used to identify the current user's role type as judge, lawyer, or party. The core processing engine is bidirectionally connected to the role adaptation module and various functional modules, providing the entire system with semantic analysis, task planning and data processing capabilities. The output of the role adaptation module is connected to the multi-terminal function module and is used to load differentiated function views according to the identified role type. The output of the standardized litigation materials access module is connected to the core processing engine, which is used to connect to the official litigation platform and parse standardized electronic materials. The oral narration assistance module, court hearing response assistance module, and lie detection and question-and-answer scheme generation module are all bidirectionally connected to the core processing engine, respectively realizing barrier-free interaction, real-time court hearing response assistance, contradiction identification and strategy generation functions; The party-side selection control module is connected to the party-side terminal of the multi-terminal functional module and is used to control the activation and deactivation of the intelligent assistance function on the party-side terminal. The local deployment and data security module is connected to all modules of the system to achieve localized data storage and compatibility with domestically produced software and hardware. All system outputs are marked as auxiliary suggestions, and the final decision-making authority belongs to the judges.

2. The court trial intelligent assistance system with multi-role adaptation as described in claim 1, characterized in that: The core processing engine includes a domestically developed large language model module and an autonomous intelligent agent module; The output of the domestic large language model module is connected to the input of the autonomous intelligent agent module, and is used to complete semantic understanding of case materials, evidence association analysis, text generation and contradiction semantic comparison. The autonomous intelligent agent module includes a task planning submodule, a multi-step execution submodule, and a failure retry submodule. The task planning submodule receives user natural language instructions and breaks them down into multi-level subtasks. The multi-step execution submodule calls the corresponding tools to complete the subtasks. The failure retry submodule is used to execute retry and rollback strategies when the subtask fails.

3. The court trial intelligent assistance system with multi-role adaptation according to claim 1, characterized in that: The multi-terminal functional module includes a judge terminal module, a lawyer terminal module, and a party terminal module; The judge terminal module is used to realize automatic prompting of disputed points, verification of the integrity of the chain of evidence, accurate push of similar cases, and generation of draft judgments. The lawyer terminal module is used to generate cross-examination opinions, retrieve rebuttal evidence, and generate draft legal statements. It also integrates functions such as evidence sufficiency analysis, statement consistency verification, and false statement risk warning. The party terminal module is used to provide visual guidance for the litigation process, popularize the explanation of evidence, and analyze the expected outcome of the case.

4. The court trial intelligent assistance system with multi-role adaptation according to claim 1, characterized in that: The lie detection and question-and-answer scheme generation module includes a data source verification submodule, a contradiction detection submodule, a question-and-answer scheme generation submodule, and an evidence linkage submodule. The output of the data source verification submodule is connected to the contradiction detection submodule to verify that the input data is only court hearing audio and video recordings, pre-trial transcripts, evidence materials and official case information. The contradiction detection submodule is used to compare the court statements, pre-trial transcripts and evidence materials to identify factual contradictions, quantity and amount contradictions and timeline contradictions and calculate the confidence level. The question-and-answer scheme generation submodule is connected to the contradiction detection submodule and is used to generate a multi-level follow-up question sequence and court rebuttal strategy based on the contradiction points. The evidence linkage submodule is connected to the question-and-answer scheme generation submodule and the judge terminal module, respectively, and is used to automatically highlight the corresponding evidence documents and key paragraphs when displaying follow-up questions and strategies.

5. The court trial intelligent assistance system with multi-role adaptation according to claim 1, characterized in that: The party-side selection control module includes a user profiling submodule, an access management submodule, and a switch control submodule. The user profiling submodule is used to collect information on case type, age of the parties involved, physical condition, representation status, and litigation integrity records. The permission management submodule is connected to the user profiling submodule and is used to determine whether the person has permission to use auxiliary functions according to preset rules. The switch control submodule is connected to the permission management submodule and the party terminal module, and is used for the court or judges to manually / automatically turn on or off the intelligent auxiliary function of the party terminal.

6. The court trial intelligent assistance system with multi-role adaptation according to claim 1, characterized in that: The standardized access module for litigation materials includes a platform interface submodule, a material parsing submodule, and a structured storage submodule; The platform interface submodule is bidirectionally connected to the People's Court Online Service Platform for automatic synchronization of electronic litigation materials; The material analysis submodule is connected to the platform interface submodule and is used to identify materials with a unified naming format and extract case elements and evidence information. The structured storage submodule is connected to the material analysis submodule and the core processing engine, and is used to store the analyzed materials in a structured form for the core processing engine to quickly retrieve and call.

7. The court trial intelligent assistance system with multi-role adaptation according to claim 1, characterized in that: The oral communication assistance module includes a voice interaction submodule, a semantic transcription submodule, a document generation submodule, and an age-friendly submodule. The voice interaction submodule is used to receive user verbal instructions and case statements, and also supports the output content of the voice broadcast system. The semantic transcription submodule is connected to the voice interaction submodule and is used to convert voice information into text and perform semantic error correction. The document generation submodule is connected to the semantic transcription submodule and the core processing engine, and is used to generate a structured complaint, a list of evidence, and a draft of the answer based on the oral content. The age-friendly submodule is connected to the system's front-end interface and is used to provide large fonts, high-contrast display, and simplified operation processes.

8. The auxiliary method of a multi-role adaptive intelligent court trial assistance system according to any one of claims 1-7, characterized in that: Includes the following steps: Step 1: User identity authentication and role recognition. Users log in to the system through the unified court authentication platform. The identity recognition module verifies the user's identity and determines their role as a judge, lawyer, or party. Step 2: Standardized access and preprocessing of litigation materials. The standardized access module for litigation materials synchronizes with the electronic materials on the People's Court Online Service Platform, and completes parsing, element extraction and structured storage. Step 3: Role-adapted view loading and function initialization. The role-adaptation module loads the corresponding function interface according to the user role, and the core processing engine preloads case-related data and model resources. Step 4: Pre-trial intelligent assistance and risk warning. The system provides judges with evidence chain verification and dispute focus sorting, lawyers with evidence reading reminders and statement consistency verification, and eligible parties with oral statement assistance and process guidance. Step 5: Real-time semantic analysis and contradiction detection in court proceedings. The system transcribes court proceedings audio in real time. The lie-debunking and question-and-answer generation module verifies the legality of the data source, compares the statements made in court with the evidence on file, identifies contradictions, and calculates confidence levels. Step 6: Differentiated intelligent assistance output. The system generates a sequence of follow-up questions and rebuttal strategies for judges, real-time evidence examination suggestions for lawyers, and response suggestions for parties who have activated the assistance function. All outputs are marked "System suggestion, for reference". Step 7: Local data storage and audit log retention. All court proceedings data and system operation records are stored on the court's local server. Lawyer risk warnings and system auxiliary operations generate tamper-proof audit logs.

Images