Case quality intelligent evaluation method and system based on large language model

Abstract

The invention provides a case quality intelligent evaluation method and system based on a large language model, and the method comprises the steps: building an evaluation standard library covering various cases based on a law normative file and judgment practice; designing a complete evaluation reasoning chain based on the evaluation standard library; constructing a multi-agent system based on the evaluation reasoning chain; based on the multi-agent system, realizing distributed task allocation for unplanned online tasks by applying an imprecise alternating direction multiplier method algorithm; based on the distributed task allocation result, combining planning and reinforcement learning technologies to solve a relational multi-agent case association problem; based on the multi-agent case association processing result, applying a forced zero method sparse graph technology to realize effective learning based on a case graph; and constructing a case quality assessment knowledge graph based on the effective learning result of the case graph. By adopting the technical scheme, the accuracy, comprehensiveness and efficiency of case quality evaluation are improved.

Description

Technical Field

[0001] This application relates to the technical field of legal artificial intelligence, and particularly to an intelligent case quality assessment method and system based on a large language model. Background Art

[0002] Case quality assessment is an important link in legal affairs and is of great significance for ensuring judicial fairness and improving the quality of legal services. With the continuous growth in the number of legal cases and the increasing complexity, traditional manual assessment methods are difficult to meet the needs of the modern legal system, and there is an urgent need for the support and assistance of intelligent technologies.

[0003] Currently, common case quality assessment technologies mainly include rule-based assessment systems and simple data statistical analysis systems. The rule-based assessment system checks cases through a preset rule library, such as procedural legality, document format, etc.; the data statistical analysis system evaluates by statistically analyzing quantitative indicators such as case processing time and appeal rate. Although these systems can assist in case quality assessment work to a certain extent, the assessment dimensions are relatively single and difficult to adapt to the needs of complex and changing legal cases.

[0004] More advanced existing technologies use machine learning methods to assess case quality. By analyzing historical case data, an assessment model is established to predict case quality. This technology uses natural language processing technology to parse legal documents and realizes a preliminary quality assessment of cases through feature extraction and classification algorithms, improving the objectivity and efficiency of assessment to a certain extent.

[0005] However, existing technologies have obvious deficiencies when facing complex and changing case types: First, there is a lack of a dynamically updated assessment standard library, making it difficult to adapt to changing regulatory requirements; second, the assessment logic chain is incomplete, making it difficult to comprehensively evaluate the procedural legality, entity handling appropriateness, and legal document standardization of cases; third, the system intelligence level is not high, and it is impossible to achieve precise management and intelligent assessment of case quality, especially with obvious shortcomings in dealing with unplanned online tasks and multi-agent case association problems. Summary of the Invention

[0006] In view of this, this application provides an intelligent case quality assessment method and system based on a large language model, which solves the problems of untimely update of assessment standards, incomplete assessment logic chain, and low system intelligence level in existing technologies. An embodiment of this application provides an intelligent case quality assessment method based on a large language model, including: Construct an assessment standard library covering various cases based on legal regulatory documents and trial practices; Based on the above-mentioned evaluation standard library, design a complete evaluation inference chain, including program legality evaluation, entity handling appropriateness evaluation, and legal document normativity evaluation; Based on the above-mentioned evaluation inference chain, construct a multi-agent system including a case backtracking agent, a quality scoring agent, and an improvement suggestion agent to realize the intelligent evaluation and analysis of case quality; Based on the above-mentioned multi-agent system, apply the inexact alternating direction method of multipliers algorithm to achieve distributed task allocation for unplanned online tasks, and improve the system's processing ability for sudden or unconventional cases; Based on the above-mentioned distributed task allocation results, combine planning and reinforcement learning techniques to solve the problem of multi-agent case association, and realize the intelligent processing of complex associated cases; Based on the above-mentioned multi-agent case association processing results, apply the forced zero method sparsification graph technology to realize effective learning based on case graphs; Based on the effective learning results of the above-mentioned case graphs, construct a case quality evaluation knowledge graph to realize the intelligence and precision of case quality management, and provide data support for legal decision-making. Optionally, based on legal regulatory documents and trial practices, construct an evaluation standard library covering various types of cases, including: Based on legal regulatory documents and trial practices, extract the regulatory requirements and standards related to case quality evaluation through text extraction and semantic analysis technologies, and form a structured dataset of regulatory requirements; Based on the above-mentioned dataset of regulatory requirements, classify according to case types and evaluation dimensions, construct a hierarchical evaluation standard model, and generate an initial evaluation standard library; Based on the above-mentioned initial evaluation standard library, design a dynamic update mechanism including new regulatory document detection, semantic change recognition, and automatic update of standard items to realize the real-time update ability of the evaluation standard library; Based on the above-mentioned dynamic update mechanism, develop a standard conflict detection algorithm and coordination strategy, and solve the possible conflict problems in the process of standard update through semantic similarity analysis and rule priority determination, so as to realize the real-time update and optimization of the evaluation standard library. Optionally, based on legal regulatory documents and trial practices, extract the regulatory requirements and standards related to case quality evaluation through text extraction and semantic analysis technologies, and form a structured dataset of regulatory requirements, including: Based on legal regulatory documents and trial practices, collect regulatory documents to form an original document set; Based on the above-mentioned original document set, preprocess the documents through natural language processing technologies, including text cleaning, word segmentation, part-of-speech tagging, and syntactic analysis, to generate structured text; Based on the structured text, apply named entity recognition and relationship extraction techniques to identify the evaluation elements and standard items in the text, extract the semantic relationships between them, and construct a preliminary network of specification elements; Based on the network of specification elements, through semantic similarity calculation and clustering analysis, merge and standardize duplicate or similar evaluation elements to form a structured dataset of specification requirements. Optionally, based on the evaluation standard library, design a complete evaluation inference chain, including: Based on the procedural norms in the evaluation standard library and combining with the reasoning ability of the multi-agent system, design a sub-chain for evaluating the legality of procedures including case-filing procedures, service procedures, evidence-filing procedures, and trial procedure links to achieve a comprehensive evaluation of the legality of case procedures; Based on the substantive norms in the evaluation standard library and the evaluation results of the legality of procedures, design a sub-chain for evaluating the appropriateness of substantive handling including fact-finding, acceptance of various types of evidence, application of laws, and reasoning in judgments to achieve a precise evaluation of the appropriateness of substantive handling of cases; Based on the document norms in the evaluation standard library and the evaluation results of the appropriateness of substantive handling, design a sub-chain for evaluating the normativity of legal documents including document format, language expression, logical structure, and content integrity to achieve a detailed evaluation of the normativity of legal documents; Based on the sub-chain for evaluating the legality of procedures, the sub-chain for evaluating the appropriateness of substantive handling, and the sub-chain for evaluating the normativity of legal documents, through multi-task learning and chain reasoning techniques of large language models, integrate each evaluation sub-chain and develop an inference control mechanism including evaluation sequence control, intermediate result feedback, and evaluation depth adjustment to form a complete case quality evaluation inference chain. Optionally, based on the evaluation inference chain, construct a multi-agent system including a case backtracking agent, a quality scoring agent, and an improvement suggestion agent to achieve intelligent evaluation and analysis of case quality, including: Based on the evaluation inference chain, develop a case backtracking agent with the ability to analyze the entire process of the case. Through time-series analysis and causal reasoning, achieve backtracking analysis of the entire process of case handling to identify key nodes and existing problems in case handling; Based on the case backtracking analysis results and the evaluation inference chain, develop a quality scoring agent. Through multi-dimensional evaluation and weight adaptive algorithms, achieve quantitative scoring of case quality and generate evaluation results including overall scores and sub-dimensional scores; Based on the quality scoring results and the evaluation standard library, develop an improvement suggestion agent. Through difference analysis and best practice matching, generate specific and operable improvement suggestions for dimensions with scores lower than the first preset threshold to provide guidance for improving case quality; Based on the case retrospective agent, the quality scoring agent, and the improvement suggestion agent, design an agent collaboration framework that includes task allocation, information sharing, and result fusion. Optimize the collaboration strategy through reinforcement learning to achieve efficient collaborative work among agents, form a complete multi-agent system, and realize the intelligent evaluation and analysis of case quality. Optionally, based on the multi-agent system, apply the inexact alternating direction method of multipliers (ADMM) algorithm to achieve distributed task allocation for unplanned online tasks, improving the system's ability to handle sudden or unconventional cases, including: Based on the multi-agent system, extract features and perform mathematical modeling on unplanned online tasks, and construct a task allocation optimization model that includes task features, resource constraints, and time windows. Based on the task allocation optimization model, design an inexact ADMM algorithm framework. Through steps such as problem decomposition, variable update, and multiplier adjustment, transform the global optimization problem into multiple sub-problems that can be solved in parallel. Based on the inexact ADMM algorithm framework, develop an inexact update strategy that includes adaptive precision control, early stopping strategy, and approximate solution. By reducing the computational precision requirements for each iteration, accelerate the algorithm convergence speed. Based on the inexact update strategy, develop a distributed computing framework that supports multi-node parallel computing. Through task partitioning, node coordination, and result aggregation, achieve efficient distributed processing of unplanned online tasks. Optionally, based on the distributed task allocation results, combine planning and reinforcement learning techniques to solve the problem of multi-agent case association and realize the intelligent processing of complex associated cases, including: Based on the distributed task allocation results, construct a case relevance recognition model through techniques such as text similarity analysis, entity recognition, and relationship extraction to achieve automatic recognition of case groups with internal connections. Based on the case group with internal connections, design a hierarchical task planning framework that includes goal decomposition, constraint recognition, and action sequence generation. Through symbolic planning and heuristic search, generate a preliminary action plan for handling associated cases. Based on the preliminary action plan, develop a reinforcement learning optimization model that includes state representation, reward mechanism, and policy network. Through interaction with the environment and policy iteration, optimize the processing strategy for associated cases and improve the processing effect. Based on the reinforcement learning optimization model and the hierarchical task planning framework, realize a dynamic fusion mechanism of planning and learning. Through a two-way feedback mechanism of planning guiding exploration and learning optimizing planning, improve the adaptability of the strategy while ensuring the rationality of the planning, and realize the intelligent processing of complex associated cases. Optionally, based on the multi-agent case association processing results, apply the forced zero method sparsification graph technology to achieve effective learning based on the case graph and improve the system's knowledge representation and reasoning capabilities, including: Based on the multi-agent case association processing results, construct a case knowledge graph containing case elements, legal concepts, and processing procedures through technologies such as entity extraction, relationship recognition, and knowledge fusion, providing a data basis for graph learning; Based on the case knowledge graph, design a graph neural network model containing graph convolutional layers, attention mechanisms, and message passing mechanisms, and achieve preliminary modeling of the case graph structure through node feature extraction and edge relationship learning; Based on the graph neural network model, implement a forced zero method sparsification mechanism containing L0 regularization, gated activation, and gradient estimation, by explicitly controlling the model complexity and the number of activated nodes; Based on the forced zero method sparsification mechanism, develop an adaptive learning algorithm containing meta-learning, knowledge distillation, and incremental training, by dynamically adjusting the learning strategy and model structure to adapt to the graph characteristics of different types of cases and achieve effective learning based on the case graph. Optionally, based on the effective learning results of the case graph, construct a case quality assessment knowledge graph to achieve the intelligence and precision of case quality management and provide data support for legal decision-making, including: Based on the effective learning results of the case graph, integrate multi-source assessment data containing assessment criteria, case characteristics, assessment results, and improvement suggestions through data cleaning, format conversion, and redundancy elimination technologies to form a unified data set; Based on the unified data set, design a knowledge graph schema containing concept hierarchies, relationship types, and attribute definitions, and construct a conceptual framework for the case quality assessment field through ontology engineering and semantic modeling to guide the construction of the knowledge graph; Based on the knowledge graph schema and the unified data set, convert unstructured and semi-structured data into knowledge triples through technologies such as named entity recognition, relationship extraction, and event detection to construct an initial case quality assessment knowledge graph; Based on the initial case quality assessment knowledge graph, develop a knowledge reasoning engine containing rule reasoning, path reasoning, and statistical reasoning, and design a standardized application interface to support multi-dimensional queries, intelligent analysis, and precise recommendations for case quality management, achieving the intelligence and precision of case quality management and providing data support for legal decision-making. An embodiment of the present application also provides a computer system, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above-mentioned intelligent case quality evaluation method based on a large language model. An embodiment of the present application also provides a computer-readable storage medium, which stores computer instructions for causing a computer to execute the above-mentioned intelligent case quality evaluation method based on a large language model. An embodiment of the present application also provides a computer program product, including computer instructions, which implement the steps of the above-mentioned intelligent case quality evaluation method based on a large language model when executed by a processor. The present application has the following technical effects: By constructing a dynamically updated evaluation standard library, the real-time update and optimization of evaluation standards are realized, improving the system's adaptability to newly emerging legal norms and trial practices; By designing a complete evaluation inference chain, a comprehensive evaluation of case quality is realized, covering multiple dimensions such as procedural legality, entity handling appropriateness, and legal document standardization; By constructing a multi-agent system, the intelligent evaluation and analysis of case quality are realized, improving the accuracy and efficiency of evaluation; By applying the inexact alternating direction method of multipliers algorithm, efficient distributed task allocation for unplanned online tasks is realized, improving the system's processing ability for sudden or unconventional cases; By combining planning and reinforcement learning techniques, the intelligent processing of complex related cases is realized, improving the efficiency and consistency of handling related cases; By applying the forced zero method sparsification graph technique, effective learning based on case graphs is realized, improving the system's knowledge representation and reasoning ability; By constructing a case quality evaluation knowledge graph, the intelligent and precise management of case quality is realized, providing data support for legal decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] To more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. The following drawings only show some aspects of the present application.

[0008] Figure 1 is a schematic flowchart of the intelligent case quality evaluation method based on a large language model provided by an embodiment of the present application; Figure 2 is a schematic flowchart of the construction and dynamic update mechanism of the case quality evaluation standard library provided by an embodiment of the present application; Figure 3 It is a schematic flowchart of the reasoning chain for case quality assessment based on large language models provided by an embodiment of the present application; Figure 4 It is a schematic flowchart of a specialized intelligent agent system provided by an embodiment of the present application; Figure 5 It is a schematic flowchart of a distributed task allocation based on the inexact alternating direction method of multipliers provided by an embodiment of the present application. Detailed implementation manners

[0009] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are only some of the embodiments of the present disclosure, rather than all the embodiments. The components of the embodiments of the present disclosure described and illustrated herein generally can be arranged and designed in a variety of different configurations. Therefore, the detailed description of the embodiments of the present disclosure provided herein is not intended to limit the scope of the claimed present disclosure, but merely represents selected embodiments of the present disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of the present disclosure without creative efforts shall fall within the scope of protection of the present disclosure.

[0010] As Figure 1 shown, an embodiment of the present application provides an intelligent case quality assessment method based on large language models, which is applied to an intelligent case quality assessment system and includes. S1: Based on legal normative documents and trial practices, construct an evaluation standard library covering various types of cases. Step S1 aims to construct an evaluation standard library covering various types of cases and design a dynamic update mechanism. This is the basis of the entire intelligent case quality assessment system. In actual legal work, the accuracy and timeliness of evaluation standards directly affect the reliability of evaluation results.

[0011] First, the system needs to collect and organize legal normative documents. Through text extraction and semantic analysis techniques, extract the normative requirements and standards related to case quality assessment from various legal regulations, legal interpretations, guiding opinions and other normative documents.

[0012] Specifically, the system first preprocesses the collected original documents, including text cleaning, word segmentation, part-of-speech tagging, and syntactic analysis. Taking the regulations on the acceptance of various types of evidence in a certain legal article as an example, the system can parse the article "Relevant departments shall review and judge various types of evidence provided by the parties" into a structured form of subject (relevant departments), obligation behavior (review and judge), and object (various types of evidence provided by the parties) through natural language processing techniques.

[0013] Next, the system applies named entity recognition and relation extraction technologies to identify the evaluation elements and criterion items in the text, and extracts the semantic relations between them. For example, it identifies "admissibility of various bases" as an evaluation element, "legitimacy, relevance, and authenticity of various bases" as specific criterion items, and establishes a semantic relation such as "admissibility - includes - judgment of the legitimacy of various bases", thus constructing a preliminary network of normative elements.

[0014] Then, through semantic similarity calculation and clustering analysis, duplicate or similar evaluation elements are merged and standardized. For example, "examination of various bases" and "judgment of various bases" are highly similar semantically and can be merged into a unified element of "examination and judgment of various bases". Through this process, a structured dataset of normative requirements is formed.

[0015] Based on this dataset, the system classifies according to case types (such as civil and commercial cases, etc.) and evaluation dimensions (such as procedural legitimacy, appropriateness of entity handling, normativity of legal documents, etc.), constructs a hierarchical evaluation criterion model, and generates an initial evaluation criterion library. For example, for a certain type of case, an evaluation criterion system can be constructed that includes multiple dimensions such as case - filing procedure, service procedure, mediation procedure, and trial procedure.

[0016] Specifically, when constructing the hierarchical evaluation criterion model, a top - down multi - level classification structure is adopted, and the evaluation criteria are scientifically stratified according to case types and evaluation dimensions. First, at the top level, it is divided into main categories such as civil cases and commercial cases according to the nature of the case; then at the second level, it is further subdivided according to specific fields. For example, civil cases can be subdivided into family disputes, labor disputes, etc.; commercial cases can be subdivided into contract disputes, intellectual property rights, corporate governance, etc. Under each case type, the system then constructs the third level according to evaluation dimensions, which are uniformly divided into three major dimensions: procedural legitimacy, appropriateness of entity handling, and normativity of legal documents. At the fourth level, the system further refines each evaluation dimension into multiple specific indicators. For example, under procedural legitimacy, there are case - filing procedure, service procedure, evidence - presenting procedure, etc.; under appropriateness of entity handling, there are fact - finding, admissibility of various bases, application of law, etc.; under document normativity, there are format requirements, language expression, logical structure, etc.

[0017] The generation of the initial evaluation standard library adopts a method combining templatization and knowledge filling. First, the system designs a standardized data structure template for each leaf node (specific evaluation indicator), including fields such as indicator name, indicator description, scoring criteria, basis source, scope of application, etc. Then, based on the aforementioned specification requirements dataset, the system uses a semantic matching algorithm to match and fill the extracted specification requirements with the corresponding indicator templates, generating structured evaluation standard entries. For example, for the "service procedure" indicator, the system will fill in its scoring criteria with specific requirements such as "the service method complies with legal requirements, the service period complies with regulations, and the service records are complete". To ensure the quality and applicability of the standard library, the system also introduces an expert review mechanism, inviting legal experts to check and optimize the initial standard library, further improving the scientificity and practicality of the evaluation criteria.

[0018] The initially generated evaluation standard library is stored in a graph database, supporting flexible query and correlation analysis while retaining the hierarchical structure. Through the tree-like index structure, the system can quickly locate the applicable evaluation standard set according to the case type and evaluation requirements, providing an accurate standard basis for subsequent case quality evaluation. This hierarchical design of the evaluation standard model not only makes the organization of the evaluation criteria clearer and more systematic but also provides a structural basis for intelligent screening and priority ranking during the evaluation process.

[0019] To maintain the timeliness of the evaluation criteria, the system designs a dynamic update mechanism. This mechanism includes functions such as new specification file detection, semantic change recognition, and automatic update of standard items. When new regulations or legal interpretations are issued, the system can automatically detect and extract relevant evaluation criteria and update them to the standard library. For example, when a relevant department issues a new legal interpretation, the system can automatically identify the content related to case quality evaluation and integrate it into the existing evaluation standard library.

[0020] In addition, the system also develops a standard conflict detection algorithm and coordination strategy to solve possible conflict problems during the standard update process through semantic similarity analysis and rule priority determination. For example, when a new legal interpretation is inconsistent with the original specification, the system can automatically coordinate the conflict based on factors such as the legal effect level and formulation time to ensure the consistency and reliability of the evaluation standard library.

[0021] Through the above steps, the system has successfully constructed an evaluation standard library covering various types of cases, realized its dynamic update and optimization, and provided a reliable standard basis for subsequent case quality evaluation.

[0022] S2: Based on the evaluation standard library, design a complete evaluation inference chain, including program legality evaluation, entity handling appropriateness evaluation, and legal document normativity evaluation.

[0023] In the embodiments of the present application, based on the procedural norms in the evaluation standard library and combined with the reasoning ability of the multi-agent system, a sub-chain for evaluating the legality of procedures is designed. This sub-chain covers the entire process of case handling.

[0024] Taking the case-filing procedure as an example, the system can check whether the case meets the acceptance conditions, whether the case-filing review is completed within the legal time limit, whether the parties' rights and obligations are informed as required, etc., so as to achieve a comprehensive evaluation of the procedural legality of the case.

[0025] Specifically, the system represents the procedural norms as a series of checkpoints and logical relationships, and through the reasoning ability of the large language model, analyzes the procedural descriptions in the case materials to determine whether they conform to the corresponding norms. For example, for the service procedure, the system can extract information such as the service time, service method, and service object from the case materials, compare them with the legal requirements, and evaluate the legality of the service procedure.

[0026] Secondly, based on the substantive norms in the evaluation standard library and the evaluation results of the procedural legality, the system designs a sub-chain for evaluating the appropriateness of substantive handling. This sub-chain focuses on the substantive handling process of the case, including fact-finding, adoption of various types of evidence, application of laws, reasoning in the judgment, etc.

[0027] For example, in the link of adopting various types of evidence, the system can evaluate whether the legality, relevance, and probative force of various types of evidence in the case have been fully considered, and whether the contradictions between various types of evidence have been reasonably resolved, etc.

[0028] In implementation, the system analyzes the substantive content in the case materials through the powerful semantic understanding and logical reasoning ability of the large language model to evaluate the appropriateness of its handling. For example, for the link of applying laws, the system can analyze whether the legal basis in a document is accurate and complete, whether there are errors or omissions in the application of laws, so as to evaluate the appropriateness of the application of laws.

[0029] Thirdly, based on the document norms in the evaluation standard library and the evaluation results of the appropriateness of substantive handling, the system designs a sub-chain for evaluating the normativity of legal documents. This sub-chain focuses on the form and content norms of legal documents, including aspects such as document format, language expression, logical structure, and content integrity.

[0030] For example, the system can evaluate whether the format of a document is standardized, whether the language is accurate and concise, whether the argumentation is well-structured, and whether the reasoning is sufficient, etc.

[0031] In implementation, the system conducts multi-dimensional evaluations on legal documents through the text analysis ability of the large language model.

[0032] For example, by analyzing the paragraph structure, keyword distribution, and language style of the document, the logical structure and language expression of the document are evaluated; by comparing the content of the document with the case facts and legal provisions, the integrity and accuracy of the document content are evaluated.

[0033] Finally, the system integrates the above-mentioned evaluation sub-chains and forms a complete case quality evaluation inference chain through the multi-task learning and chain reasoning technologies of the large language model. The system also develops an inference control mechanism that includes evaluation sequence control, intermediate result feedback, and evaluation depth adjustment to ensure the efficiency and accuracy of the evaluation process.

[0034] For example, the system first evaluates the procedural legality of the case. If procedural defects are found, the evaluation depth of the subsequent links is correspondingly reduced; if the evaluation result of procedural legality is good, the system further deeply evaluates the appropriateness of entity handling and the normativity of legal documents. Through this dynamically adjusted evaluation strategy, the system can concentrate evaluation resources on key issues and improve evaluation efficiency.

[0035] Through the above steps, the system has successfully designed a case quality evaluation inference chain based on the large language model, achieving a comprehensive and accurate evaluation of case quality, and providing a methodological basis for subsequent intelligent evaluation and analysis.

[0036] S3: Based on the evaluation inference chain, construct a multi-agent system including a case backtracking agent, a quality scoring agent, and an improvement suggestion agent to realize the intelligent evaluation and analysis of case quality.

[0037] Among them, the multi-agent system consists of a case backtracking agent, a quality scoring agent, and an improvement suggestion agent. Each agent is an independent artificial intelligence model with specific functions and structural designs.

[0038] Taking the case backtracking agent as an example, its internal structure includes the following key components: Data preprocessing module: Responsible for processing the input case materials, including text cleaning, word segmentation, feature extraction, etc.

[0039] Specifically, the NLTK and spaCy libraries in Python are used to perform word segmentation, part-of-speech tagging, and named entity recognition on the text to extract key information.

[0040] Time series analysis module: Use a sequence model based on the Transformer architecture to identify the key events and time points of case processing.

[0041] The sequence model contains 6 layers of Transformer encoders, with a hidden layer dimension of 512, 8 attention heads, and a feed-forward network dimension of 2048. The model is trained using the cross-entropy loss function, the Adam optimizer, with an initial learning rate set to 0.0001, and the learning rate is adjusted using cosine annealing. The batch size is 32, and the number of training epochs is 30. To avoid overfitting, weight decay (weight of 0.01) and dropout (ratio of 0.1) techniques are used.

[0042] Causal inference module: Based on the causal graph model and the rule engine, analyze the causal relationships between events.

[0043] The causal graph model uses a method based on structural equation modeling (SEM), which includes three steps: variable selection, edge direction determination, and weight estimation. Variable selection uses feature engineering based on domain knowledge; edge direction determination uses a scoring-based algorithm (such as the BIC score); weight estimation uses the maximum likelihood estimation method. The rule engine uses the Rete algorithm to achieve efficient pattern matching. The rule base contains approximately 200 domain-specific rules, which are written by experts and updated regularly.

[0044] Among them, the approximately 200 domain-specific rules contained in the rule base mainly cover four categories: case processing procedure rules, causal relationship inference rules, case element association rules, and quality assessment inference rules. Case processing procedure rules are used to describe the standard steps and sequential relationships in the case processing flow, such as "material review must be completed within 7 days after acceptance" and "the verification of the parties' identities must be completed before a certain procedure". Such rules account for approximately 30% of the total and are mainly used to judge whether the case processing process meets the procedural requirements.

[0045] Causal relationship inference rules are used to analyze the causal connections between various links in the case processing process, such as "if the evidence collection in the case investigation stage is insufficient, it will lead to difficulties in fact-finding" and "if the parties' rights and obligations are not fully informed, it may lead to doubts about the fairness of the procedure". Such rules account for approximately 25% of the total and are mainly used to identify the key issues and their roots in case processing.

[0046] Case element association rules describe the logical relationships between different elements in the case, such as "if multiple pieces of evidence point to the same fact and corroborate each other, the reliability of the fact-finding increases" and "if the applicable legal standard has an obvious relevance to the case facts, the accuracy of legal application increases". Such rules account for approximately 20% of the total and are mainly used to evaluate the internal logic and consistency of case processing.

[0047] The quality assessment inference rules are directly aimed at the comprehensive judgment of case quality. For example, "If the score of procedural legality is lower than 70 points, there is a significant risk of procedural defects in the case", "If the logical structure of the judgment document is chaotic and the reasoning is insufficient, the quality rating of the document shall not be higher than level C", etc. Such rules account for about 25% of the total, and are mainly used to derive the comprehensive conclusion of case quality based on the evaluation results of various indicators.

[0048] These rules are represented in formal language, including a precondition and a conclusion, and each rule is given a different confidence weight to reflect the universality and importance of rule application. The rule base is constructed through the expert knowledge engineering method and is continuously optimized and expanded based on the analysis of actual cases. During the system operation, the rule engine realizes efficient pattern matching and rule triggering through the Rete algorithm, supports forward chaining inference and backward chaining inference, and can automatically activate relevant rules according to case characteristics and processing status, realizing the intelligent analysis and quality assessment of the case processing process.

[0049] Result integration module: Adopts a weighted voting mechanism to integrate the results of time series analysis and causal reasoning, and generates a final retrospective analysis report. The weights are dynamically adjusted according to the historical accuracy of each module, and the initial weights are set as time series analysis: causal reasoning = 0.4:0.6.

[0050] The training of the entire multi-agent system adopts a phased strategy: First, each agent is trained independently, and then collaborative optimization is carried out.

[0051] In the collaborative optimization stage, a reinforcement learning method based on the Actor-Critic architecture is used. Both the Actor network and the Critic network adopt a multi-layer perceptron structure. The Actor network has 3 hidden layers (sizes are 256, 128, 64 respectively), uses the ReLU activation function, and the output layer uses Softmax activation; the Critic network has 2 hidden layers (sizes are 256, 128 respectively), uses the ReLU activation.

[0052] The training uses the Proximal Policy Optimization (PPO) algorithm, the discount factor γ = 0.99, the advantage estimation parameter λ = 0.95, the learning rate is 3e-4, and the training batch is 10,000 cases. The reward function design comprehensively considers the evaluation accuracy, efficiency and consistency. Specifically: R = 0.5 × accuracy + 0.3 × efficiency score + 0.2 × consistency score.

[0053] Specifically, in step S3, the system develops a case retrospective agent based on the evaluation inference chain. This agent has the ability to analyze the entire process of the case and realizes the retrospective analysis of the entire process of case processing through time series analysis and causal reasoning.

[0054] In terms of specific implementation, the intelligent agent uses timeline analysis and event extraction techniques to extract the key events and time points of case handling from case materials and construct a chronological graph of case handling. Then, through causal reasoning techniques, it analyzes the causal relationships between events to identify the key nodes and potential problems in case handling. For example, the intelligent agent can find that the deficiencies in various evidence collection links have led to difficulties in subsequent fact-finding, or that insufficient pretrial preparation has led to a reduction in trial efficiency.

[0055] In addition, the system develops a quality scoring intelligent agent based on the case backtracking analysis results and the evaluation reasoning chain. This intelligent agent realizes the quantitative scoring of case quality through multi-dimensional evaluation and weight adaptive algorithms. For example, for a civil case, the intelligent agent can evaluate from three dimensions: procedural legality, appropriateness of entity handling, and standardization of legal documents, and further break them down into multiple sub-dimensions such as case-filing procedure, service procedure, fact-finding, legal application, document format, and argument logic. Weights are assigned to each dimension and sub-dimension to calculate the overall score and sub-dimension scores.

[0056] In implementation, the intelligent agent uses a multi-level evaluation model to organize the evaluation indicators into a hierarchical structure and dynamically adjusts the weights of each indicator through an adaptive weight algorithm.

[0057] For example, according to the case type and complexity, the intelligent agent may increase the weights of the fact-finding and legal application dimensions, or according to the case's dispute focus, increase the weights of relevant evaluation dimensions to ensure that the scoring results objectively reflect the case quality.

[0058] The multi-level evaluation model is a structured evaluation framework that comprehensively and systematically evaluates case quality by hierarchically organizing evaluation indicators. This model is designed in a tree structure and consists of four levels: the overall scoring level, the dimension scoring level, the sub-dimension scoring level, and the specific indicator level. The overall scoring level represents the comprehensive score of case quality and is the final output of the entire evaluation; the dimension scoring level includes three basic dimensions: procedural legality, appropriateness of entity handling, and standardization of legal documents, which respectively evaluate different aspects of case handling; the sub-dimension scoring level further breaks down each basic dimension into several key links, such as case-filing procedure, service procedure, and evidence presentation procedure under procedural legality; the specific indicator level contains the finest-grained evaluation items, such as "completeness of case-filing materials" and "compliance of case-filing time limit" under the case-filing procedure.

[0059] Evaluation indicators are the basic evaluation elements at the bottom layer of the multi-level evaluation model. Each indicator is designed for a specific aspect of case handling and has clear evaluation content, scoring criteria, and weight values. For example, in the sub-dimension of service procedures under the dimension of procedural legality, specific indicators such as "legality of service methods", "timeliness of service time", and "standardization of service records" are included. Each indicator sets a clear scoring range (usually 0-100 points) and detailed scoring rules. For example, the scoring rules for the "legality of service methods" indicator may stipulate that 100 points are obtained for using legal service methods; 80 points are obtained for using alternative service methods with sufficient reasons; 50 points are obtained for using non-standard service methods; and 0 points are obtained for completely failing to serve according to regulations.

[0060] The process of establishing a multi-level evaluation model includes four main steps: indicator system design, weight assignment, scoring rule formulation, and model verification. In the indicator system design stage, the system determines the evaluation elements and hierarchical relationships at each level based on legal norms and expert knowledge, forming a complete indicator tree structure. In the weight assignment stage, the system combines the Analytic Hierarchy Process (AHP) and the Delphi method to determine the initial weight values of the indicators at each level, reflecting the relative importance of different indicators. For example, in the overall scoring, the initial weights of the three dimensions of procedural legality, properness of entity handling, and standardization of legal documents may be set at 0.3, 0.5, and 0.2 respectively, indicating that the properness of entity handling plays a more important role in case quality evaluation.

[0061] The adaptive weight algorithm can dynamically adjust the weights of each indicator according to case characteristics and evaluation scenarios, making the evaluation results more accurate. The algorithm adjusts the weights based on three key factors: case type characteristics, case complexity, and focus of disputes. For different types of cases, the algorithm will strengthen the weights of the corresponding key dimensions. For example, for commercial contract dispute cases, the weight of the sub-dimension of law application may be increased; for a certain type of civil cases, the weight of the sub-dimension of fact finding may be increased. For case complexity, the algorithm calculates a complexity coefficient based on indicators such as the amount of case materials and the number of parties involved, and adjusts the weights accordingly. For the focus of disputes, the algorithm will identify the main disputes in the case and increase the weights of the relevant evaluation dimensions to ensure that the evaluation results can reflect the core quality issues of the case.

[0062] In addition, the algorithm also has learning ability and can continuously optimize the weight adjustment strategy based on historical evaluation data and expert feedback. Through this adaptive mechanism, the multi-level evaluation model can provide more accurate and targeted quality evaluations for different types and characteristics of cases, greatly improving the reliability and practical value of the evaluation results.

[0063] The system develops an intelligent agent for improvement suggestions based on the quality scoring results and the evaluation standard library.

[0064] Through difference analysis and best practice matching, the agent generates specific and actionable improvement suggestions for dimensions with scores lower than the first preset threshold. For example, for cases with scores lower than the first preset threshold in the dimension of legal application, the agent can point out existing legal application errors or omissions and provide correct legal application suggestions; for cases with scores lower than the first preset threshold in the document standardization dimension, the agent can provide specific methods for optimizing the document structure or enriching the reasoning.

[0065] In terms of specific implementation, the agent uses difference analysis technology to compare the differences between the case handling process and best practices, and identify deficiencies. Then, through best practice matching technology, it retrieves cases from the knowledge base that are similar to the current case but have higher handling quality, extracts the successful experiences from them, and generates targeted improvement suggestions. For example, the agent can recommend suitable methods for collecting various types of evidence, ideas for fact-finding, or legal application cases according to the characteristics of the case to help improve the quality of case handling.

[0066] Finally, based on the above three agents, the system designs an agent collaboration framework, optimizes the collaboration strategy through reinforcement learning, and realizes the efficient collaborative work among agents. For example, the analysis results of the case backtracking agent are directly input into the quality scoring agent, and the scoring results of the quality scoring agent are then input into the improvement suggestion agent, forming a closed-loop of information flow; at the same time, the output of the improvement suggestion agent can be fed back to the case backtracking agent to help it improve the analysis method, forming a feedback optimization mechanism.

[0067] In terms of implementation, the system uses a collaboration strategy optimization method based on reinforcement learning to continuously adjust the interaction methods and information sharing strategies among agents to improve the overall collaboration efficiency. For example, the system can learn when in-depth information exchange is needed among agents and when they can work in parallel, so as to improve the evaluation efficiency while ensuring the evaluation quality.

[0068] Through the above steps, the system has successfully constructed a professional multi-agent system, realized the intelligent evaluation and analysis of case quality, and provided strong technical support for improving the efficiency and quality of legal affairs work. S4: Based on the multi-agent system, apply the inexact alternating direction method of multipliers algorithm to achieve distributed task allocation for unplanned online tasks, and improve the system's ability to handle sudden or unconventional cases.

[0069] Step S4 aims to achieve distributed task allocation for unplanned online tasks through the inexact alternating direction method of multipliers algorithm. In actual legal affairs work, there are often situations where sudden or unconventional cases need to be handled urgently, and the system needs to be able to efficiently allocate resources to ensure that these tasks are processed in a timely manner.

[0070] First, the system, based on the multi-agent system, extracts features and conducts mathematical modeling on unscheduled online tasks.

[0071] For example, for a batch of emergency case quality assessment tasks, the system can extract the features of each task, such as case type, complexity, priority, estimated processing time, etc., and build a task allocation optimization model that includes task features, resource constraints, and time windows.

[0072] The system uses multi-source data fusion and feature engineering methods to extract case task features. For the case type feature, the system first automatically identifies the category to which the case belongs through text classification technology. Using a hierarchical classification model based on BERT, it maps the case document content into a predefined case type system, such as major categories like civil and commercial, administrative, etc., and further subdivides into specific subcategories. This classification model adopts a transfer learning strategy. Based on the legal text pre-training model, it is fine-tuned using labeled data, and the classification accuracy reaches over 95%.

[0073] The hierarchical classification model based on BERT is used for case type identification, and its construction and training process are as follows: The model architecture adopts a two-level classification structure. The first level corresponds to the classification of major case categories (such as civil, administrative, enforcement, etc.), and the second level corresponds to the detailed categories (such as contract disputes under civil cases). The basic pre-training model selects Chinese Legal BERT (Legal-BERT-Chinese), which is obtained by performing domain adaptation pre-training on the general Chinese BERT using 5 million legal documents. It contains 12 layers of Transformer encoders, with a hidden layer dimension of 768, 12 attention heads, and a total number of parameters of approximately 110M.

[0074] On this basis, a hierarchical classifier is constructed: The first-level classifier uses the output of the [CLS] token of BERT connected to a fully connected layer (768×5, corresponding to 5 major categories) and uses the softmax activation; the second level contains multiple classifiers, each corresponding to a detailed category under a major category. The structure is the output of the [CLS] token of BERT connected to a fully connected layer (the dimension is determined according to the number of subcategories under each major category. For example, if there are 12 subcategories under civil cases, it is 768×12), and the softmax activation is also used.

[0075] The training process is divided into two stages: In the first stage, the first-level classifier is trained. The parameters of the bottom 9 layers of BERT are frozen, and only the top 3 layers and the classification layer are fine-tuned. The weighted cross-entropy loss function is used (the weights are set inversely proportional to the number of samples in each category to solve the problem of class imbalance).

[0076] In the second stage, each second-level classifier is trained separately. Similarly, the parameters of the bottom of BERT are frozen, and the weighted cross-entropy loss function is used.

[0077] Training hyperparameter range: Learning rate: [1e-5, 2e-5, 3e-5, 5e-5], finally select 2e-5; Batch size: [16, 32, 64], finally select 32; Number of training epochs: [3, 4, 5, 6], finally select 4 epochs; Weight decay: [0.01, 0.05, 0.1], finally select 0.01; Learning rate warm-up ratio: [0.05, 0.1, 0.15], finally select 0.1; Dropout rate: [0.1, 0.2, 0.3], finally select 0.1; The training data contains 50,000 labeled cases, which are divided into training set, validation set and test set according to the ratio of 8:1:1. AdamW optimizer is used for training, and cosine learning rate scheduling strategy is adopted. On RTX 3090 GPU, the training time of the first stage is about 4 hours, and the training time of each classifier in the second stage is about 1 - 2 hours. The overall accuracy of the model on the test set reaches 95.3%, and the F1 values of each major category are all above 92%.

[0078] The extraction of case complexity features is based on the comprehensive calculation of multi-dimensional indicators, including: document volume indicators (total number of pages and words of case materials), entity quantity indicators (number of parties involved in the case), fact complexity indicators (number of disputed points, number of evidence materials), legal application indicators (involving legal requirements, legal application difficulty coefficient). The system uses a weighted summation model to calculate the complexity score, and the weights of each indicator are determined through regression analysis. The complexity is finally standardized to a 1 - 10 scale for subsequent processing.

[0079] The priority feature is evaluated based on the urgency and importance of the case. The system automatically extracts time limit elements (such as legal handling period, applicant required period) from the case information to calculate the urgency, and combines factors such as the scope of case influence, involved amount, and social attention to evaluate the importance. The system uses a decision tree model to map these factors into high, medium, and low three-level priority levels and assigns numerical weights (high: 5, medium: 3, low: 1).

[0080] The estimated processing time is calculated based on historical data and machine learning models. The system constructs a random forest regression model including features such as case type, complexity, and processor experience. By analyzing the actual processing time of historical similar cases, it predicts the processing time of new cases. The model adopts ten-fold cross-validation, and the average prediction error is controlled within 15%, providing a reliable time estimation basis for task allocation.

[0081] The random forest regression model is used for estimating the case processing time, and its construction and training process are as follows: The model input features include 25 dimensions, which are divided into four categories: Basic case features (7 dimensions): case type (one-hot encoding), involved amount (logarithmic transformation), number of parties, whether it is foreign-related, etc.; Complexity features (8 dimensions): number of evidences, number of legal provisions cited, number of dispute points, professional and technical difficulty score, etc.; Historical statistical features (6 dimensions): average processing time of similar cases, historical efficiency of processing personnel, volume of concurrent cases, etc.; Procedure features (4 dimensions): whether identification is required, whether a hearing is required, expected number of times of a certain procedure, etc.; The model is constructed using the scikit-learn library, and the core parameter setting ranges are: Number of decision trees (n_estimators): [100, 200, 300, 500, 1000], and finally 500 is selected; Maximum tree depth (max_depth): [10, 15, 20, 25, 30, None], and finally 20 is selected; Minimum number of samples required for splitting (min_samples_split): [2, 5, 10, 15], and finally 5 is selected; Minimum number of samples in leaf nodes (min_samples_leaf): [1, 2, 4, 8], and finally 2 is selected; Feature sampling ratio (max_features): ['auto','sqrt', 'log2', 0.7, 0.8], and finally'sqrt' is selected; Sample sampling ratio (bootstrap): [True, False], and finally True is selected; Using sample weights (sample weights are set based on case timeliness): [True, False], and finally True is selected The training data contains 100,000 historical case records. After data cleaning and outlier handling (removing outliers of processing time using the IQR method), it is split into a training set (80%) and a test set (20%) in chronological order. Parameter optimization uses grid search combined with 5-fold cross-validation, and the evaluation metrics are root mean square error (RMSE) and mean absolute percentage error (MAPE).

[0082] The MAPE of the final model on the test set is 13.7%, and the RMSE is 2.6 days. Feature importance analysis shows that case type, number of evidences, and number of disputed points are the three most significant factors affecting the processing time. The model training time is about 15 minutes (on a 16-core CPU), and the time to predict a single case is less than 10 milliseconds, meeting the requirements of real-time applications.

[0083] The task assignment optimization model is constructed based on a mixed-integer linear programming framework. The model defines the decision variable x ij to represent whether task i is assigned to agent j, and the objective function is designed for multi-objective optimization: minimizing the weighted sum of the total completion time and the agent load imbalance. The mathematical expression is: minimize α∑ j max(∑ i x ij ×t i )+β∑ j |∑ i x ij ×t i - ∑ i t i / m| where t i represents the estimated processing time of task i, m is the number of agents, and α and β are weight coefficients. The constraint conditions include: each task must be and can only be assigned to one agent (∑ j x ij = 1); agent capacity limit (∑ i x ij ×t i ≤C j , C j is the upper limit of the processing capacity of agent j); priority constraint (high-priority tasks must be assigned first); time window constraint (tasks must be completed within the specified time window). By integrating task characteristics, resource constraints, and time requirements, this model forms a structured optimization problem, providing a clear mathematical framework for subsequent algorithm solving.

[0084] Specifically, the system models the task assignment problem as an optimization problem: minimize f(x)+g(z) subject to Ax + Bz =c where f(x) represents the efficiency objective of task assignment, g(z) represents the resource constraint objective, and Ax + Bz = c represents the constraint condition that tasks must be fully assigned. For example, f(x) can represent the objective of minimizing the task completion time, g(z) can represent the objective of balanced resource utilization, and the constraint conditions ensure that all tasks are assigned and do not exceed the resource capabilities.

[0085] Secondly, based on the task assignment optimization model, the system designs an inexact Alternating Direction Method of Multipliers (ADMM) algorithm framework. Through steps of problem decomposition, variable update, and multiplier adjustment, this algorithm transforms the global optimization problem into multiple sub-problems that can be solved in parallel. For example, in the case where 10 agents need to process 30 evaluation tasks, the system can decompose the global task assignment problem into 10 local problems for each agent, and each agent calculates in parallel the subset of tasks it should receive.

[0086] The core of the algorithm framework is to iteratively solve the problem: Update x: x(k + 1) ≈ argmin(f(x) + ρ / 2||Ax + Bzk - c + uk||²) Update z: z(k + 1) ≈ argmin(g(z) + ρ / 2||Ax(k + 1) + Bz - c + uk||²) Update the multiplier u: u(k + 1) = uk + (Ax(k + 1) + Bz(k + 1) - c) Specifically, the system designs the inexact Alternating Direction Method of Multipliers (ADMM) algorithm framework based on the principles of problem decomposition and distributed optimization. First, the system reformulates the global task assignment optimization problem into a form with a separable structure: minimize f(x) + g(z) subject to Ax + Bz = c where x represents the local variables distributed among each agent, such as the task selection scheme for each agent; z represents the globally shared variable, such as the final task assignment result; f(x) represents the objective function related to the local variables, such as the processing efficiency of the agent; g(z) represents the objective function related to the global variable, such as the load balancing degree; the constraint condition Ax + Bz = c ensures the consistency between the local variables and the global variables.

[0087] The ADMM algorithm framework designed by the system contains three core steps, forming an iterative solution process. The first step is local variable update, where each agent solves its own sub-problem in parallel: x k+1 = argmin x { f(x) + (ρ / 2)||Ax + Bz k - c + u k || 2} where ρ is the penalty parameter, u is the Lagrange multiplier, and k represents the number of iterations. This step enables each agent to optimize its task selection scheme based on the current global assignment scheme and the multiplier.

[0088] The second step is the global variable update, which is executed by the central coordinator: z k+1 = argmin_z { g(z) + (ρ / 2)||Ax k+1 + Bz - c + u k || 2} This step comprehensively considers the latest solutions of all agents, optimizes the global task allocation scheme, and ensures the achievement of the overall system goals (such as load balancing).

[0089] The third step is the multiplier update, which is also executed by the central coordinator: u k+1 = u k + (Ax k+1 + Bz k+1 - c) This step adjusts the Lagrange multiplier, punishes the behavior that violates the constraints, and guides the algorithm to converge to a feasible solution.

[0090] To improve the algorithm efficiency, the system designs an inexact update strategy, allowing approximate solutions to be used in the subproblem solving process. The specific implementations include: (1) the dynamic precision control strategy, using a lower precision threshold ε0 (such as 0.1) at the initial stage of the algorithm, and gradually increasing the precision as the iteration progresses, following the formula εk = max{ε0γ k , ε min}, where γ is the decay factor (such as 0.9), and ε min is the minimum precision requirement (such as 0.001); (2) the early stopping strategy, terminating early when the improvement amplitude of consecutive iterations in the subproblem solving process is less than the threshold δ (such as 0.01); (3) the warm start technique, using the result of the previous iteration as the initial solution of the current iteration to accelerate convergence.

[0091] The algorithm framework also includes an adaptive parameter adjustment mechanism to dynamically adjust the penalty parameter ρ to accelerate convergence. When the ratio of the primal residual to the dual residual exceeds the preset threshold μ (such as 10), increase ρ (ρ = τρ, τ>1); when the ratio of the dual residual to the primal residual exceeds μ, decrease ρ (ρ = ρ / τ). In addition, the system designs a distributed communication protocol based on exponential backoff. When node communication fails, retry at exponentially increasing time intervals, with a maximum retry count of 8 times, to ensure the robustness of the algorithm in a distributed environment.

[0092] Thirdly, based on the ADMM algorithm framework, the system develops an inexact update strategy, which speeds up the algorithm convergence rate by reducing the computational accuracy requirements for each iteration. In the traditional ADMM algorithm, the sub-problems need to be solved precisely in each iteration, resulting in a large computational overhead. The inexact update strategy allows the use of approximate solutions in each iteration as long as certain accuracy requirements are met, greatly improving the algorithm efficiency.

[0093] In terms of specific implementation, the system uses adaptive precision control, early stopping strategy, and approximate solution methods. For example, in the initial stage of the algorithm, solutions with lower precision can be used, and as the iteration approaches convergence, the precision requirements can be gradually increased; or an iterative method with early termination can be used to solve the sub-problems, and the iteration stops as long as the improvement of the solution is less than a certain threshold. These strategies greatly reduce the computational complexity and improve the system's response ability to real-time tasks.

[0094] Finally, based on the inexact update strategy, the system develops a distributed computing framework that supports multi-node parallel computing. This framework realizes the efficient distributed processing of unplanned online tasks through task partitioning, node coordination, and result aggregation. For example, the system can process sub-problems in parallel on multiple computing nodes, synchronize the calculation results of each node through a central coordinator, and adjust the global allocation scheme.

[0095] In terms of implementation, the system adopts a master-slave architecture. The central node is responsible for task decomposition and result aggregation, and the slave nodes are responsible for solving the sub-problems. The nodes communicate with each other through a message passing mechanism, sharing necessary variables and results. Through this distributed architecture, the system can make full use of computing resources, improving the processing throughput and response speed.

[0096] Through the above steps, the system has successfully realized the distributed task allocation based on the inexact ADMM algorithm, improving the system's processing ability for sudden or unconventional cases, and ensuring the efficient allocation and processing of tasks under limited resources. S5: Based on the distributed task allocation results, combined with planning and reinforcement learning techniques, solve the problem of multi-agent case association, and realize the intelligent processing of complex associated cases.

[0097] Step S5 aims to combine planning and reinforcement learning techniques to solve the problem of multi-agent case association and realize the intelligent processing of complex associated cases. In legal practice, there are often situations where there are associated relationships between multiple cases, and collaborative processing of these associated cases is crucial for improving legal efficiency and consistency.

[0098] First, based on the distributed task allocation results, the system constructs a case relevance recognition model through technologies such as text similarity analysis, entity recognition, and relationship extraction. This model can automatically identify case groups with internal connections, providing a basis for subsequent collaborative processing. For example, the system can identify the following association patterns: entity overlap association (such as multiple cases of the same case-related person), factual association (such as multiple cases caused by a single traffic accident), and legal relationship association (such as multiple cases caused by a company dispute).

[0099] In terms of specific implementation, the system uses a text similarity algorithm to calculate the similarity between cases, extracts key entities (such as parties, locations, events) in cases through named entity recognition technology, and analyzes the relationships between entities through relationship extraction technology to discover potential case associations. For example, the system can find that the same case-related person "Zhang" appears in multiple cases, or that multiple cases are all related to the same traffic accident, and thus determine that there is an association between the cases.

[0100] The case relevance recognition model is constructed using multi-modal feature fusion and deep learning methods. This model consists of four key components: a text similarity analysis module, an entity recognition and association module, a relationship extraction module, and a multi-feature fusion classification module, forming an end-to-end case relevance recognition process.

[0101] The text similarity analysis module first preprocesses the content of case documents, including operations such as word segmentation, stop word removal, and synonym replacement, and then uses the Sentence-BERT model to extract text semantic features and generate a text vector representation with a dimension of 768. For any two cases, calculate the cosine similarity and Euclidean distance of their text vectors, and use the threshold method (similarity greater than 0.7 or distance less than 0.3) to preliminarily screen potential associated case pairs. To improve efficiency, the system uses the Locality-Sensitive Hashing (LSH) technique to construct an index, reducing the similarity calculation complexity from O(n²) to O(n log n), supporting fast retrieval of large-scale case libraries.

[0102] The entity recognition and association module is responsible for extracting key entities from case texts and establishing associations between entities. This module uses a BiLSTM-CRF (Bidirectional Long Short-Term Memory Network - Conditional Random Field) sequence labeling model to identify entity types such as person names, organizations, locations, times, legal provisions, etc. in cases. The model training adopts a transfer learning strategy, adapting to the domain using 10,000 labeled cases based on a general Chinese NER model, and the F1 value of entity recognition reaches 92%. In the entity association stage, a method combining rules and learning is used: for entities with the same name, the system determines whether they are the same entity through attribute comparison (such as ID card numbers, registered addresses, etc.); for entities with different names, the system identifies potential associations through knowledge graph queries and entity linking techniques. The system calculates the entity overlap rate as an important feature for case association.

[0103] The BiLSTM-CRF model is used for entity recognition in case texts, and its construction and training process are as follows: The model architecture consists of four layers: Word embedding layer: Concatenate pre-trained legal domain word vectors (300 dimensions) with character vectors (200 dimensions), and at the same time fuse part-of-speech features (20-dimensional one-hot encoding); Feature extraction layer: A bidirectional LSTM network to extract context-related feature representations; Nonlinear mapping layer: A time-distributed fully connected layer to map LSTM features to the label space; Label decoding layer: Conditional Random Field (CRF), considering the dependencies between labels to decode the optimal label sequence; The target entity types include 12 categories: person names, organization names, locations, times, amounts, legal provisions, case causes, legal documents, judicial organs, case numbers, evidence types, and litigation requests.

[0104] Training hyperparameter range: LSTM hidden layer dimension: [128, 256, 384, 512], finally select 256 (512 dimensions after bidirectional stacking); Number of LSTM layers: [1, 2, 3], finally select 2; Dropout rate: [0.3, 0.4, 0.5, 0.6], finally select 0.5; Learning rate: [0.0005, 0.001, 0.002, 0.005], finally select 0.001; Batch size: [16, 32, 64, 128], finally select 64; L2 regularization strength: [1e-5, 1e-4, 1e-3], finally select 1e-4; Gradient clipping thresholds: [1.0, 3.0, 5.0, 10.0], and finally 5.0 was selected; The training data contains 10,000 manually annotated legal documents, approximately 5 million words, and the BIO annotation scheme is used for annotation. The data is divided into a training set, a validation set, and a test set in a ratio of 8:1:1. Adam optimizer is used for training, and an early stopping strategy is adopted (stopping if the F1 value on the validation set does not improve for 5 consecutive rounds). The model is trained on a single RTX 2080Ti GPU for about 8 hours, with a total of 30 rounds of iteration.

[0105] The overall F1 value of the model on the test set reaches 92.3%, among which the F1 value of person name entities is 94.7%, the F1 value of legal clause entities is 95.2%, the F1 value of time entities is 96.8%, the F1 value of case number entities is 98.1%, and the F1 value of organization names is 89.3%. In terms of inference speed, it takes about 200 milliseconds on average to process 1000-word text on the CPU.

[0106] The relationship extraction module analyzes the semantic relationships between entities in the case and identifies key relationship types that may indicate case associations. This module adopts a BERT-based relationship classification model to identify relationship types such as "different parties in the same event", "multiple cases of the same party", "a case and its derivative cases", etc. The model training uses the remote supervision method to automatically construct the training set and fine-tune it in combination with high-quality manually annotated data, and the relationship classification accuracy reaches 85%. To handle the long-distance entity relationships in long texts, the system introduces dependency syntactic analysis and the shortest dependency path features to enhance the model's ability to capture relationships in complex contexts.

[0107] The multi-feature fusion classification module integrates the output features of the previous three modules to construct a final judgment model for case relevance. This module adopts the Gradient Boosting Decision Tree (GBDT) algorithm, and the input features include: text similarity scores (cosine similarity, Euclidean distance, Jaccard coefficient), entity overlap features (number of common entities, matching degree of important entities), relationship features (key relationship types and their confidence levels), time features (closeness of case time), case attribute features (similarity of case types, relevance of involved amounts), etc. The model output is the type of association (such as entity association, factual association, legal relationship association) between case pairs and their confidence levels.

[0108] The Gradient Boosting Decision Tree (GBDT) model is used for case relevance judgment, and its construction and training process are as follows: The model input features contain 35 dimensions, which are divided into five categories: Text similarity features (7 dimensions): TF-IDF cosine similarity, bag-of-words Jaccard similarity, BM25 similarity, BERT semantic similarity, etc.; Entity overlap features (10 dimensions): the number and proportion of co-owned personal entity names, the number and proportion of co-owned organizational entity names, the number and proportion of co-owned location entity names, the temporal proximity of co-owned time entities, the number and proportion of co-owned legal provisions, etc.; Relationship features (6 dimensions): the strength of the association between case parties, the similarity of legal relationships, the degree of association of case facts, etc.; Temporal features (4 dimensions): the time difference between case occurrences, the time difference between case filings, the time difference between case closings, the degree of temporal overlap, etc.; Case attribute features (8 dimensions): the similarity of case types, the relevance of involved amounts, the identity of handling agencies, the identity of judges, etc.; The model is implemented using the XGBoost library, and the core parameter settings are as follows: Number of weak learners (n_estimators): [100, 200, 300, 500, 1000], and 500 is finally selected; Learning rate (learning_rate): [0.01, 0.05, 0.1, 0.2], and 0.05 is finally selected; Maximum tree depth (max_depth): [3, 4, 5, 6, 7, 8], and 6 is finally selected; Subsample ratio (subsample): [0.7, 0.8, 0.9, 1.0], and 0.8 is finally selected; Column sampling ratio (colsample_bytree): [0.7, 0.8, 0.9, 1.0], and 0.8 is finally selected; Minimum sum of child node weights (min_child_weight): [1, 3, 5, 7], and 3 is finally selected; L1 regularization parameter (alpha): [0, 0.001, 0.01, 0.1, 1], and 0.01 is finally selected; L2 regularization parameter (lambda): [0.1, 1, 10, 100], and 1 is finally selected; Early stopping rounds (early_stopping_rounds): [10, 20, 30, 50], and 30 is finally selected; The training data contains 50,000 pairs of case pairs, each pair is labeled as one of the four categories: "unrelated", "weakly related", "moderately related", or "strongly related", as well as the specific type of association (entity association, fact association, legal relationship association). The data is divided into a training set, a validation set, and a test set in a 7:1:2 ratio. The problem of class imbalance is solved by increasing the weights of minority class samples, and the weights are inversely proportional to the class frequencies.

[0109] Parameter optimization uses the Bayesian optimization method, and the evaluation metric is the weighted F1 score (considering the importance of each category). The weighted F1 score of the final model on the test set reaches 89.7%. Among them, the accuracy of strongly associated categories is 94.3% and the recall rate is 92.1%; the accuracy of moderately associated categories is 88.2% and the recall rate is 86.5%; the accuracy of weakly associated categories is 82.7% and the recall rate is 80.4%; the accuracy of non-associated categories is 95.8% and the recall rate is 97.2%.

[0110] Feature importance analysis shows that BERT semantic similarity, the proportion of co-owned named entities, the strength of the case subject association, and the proximity of the case occurrence time are the four most important features for judging case relevance. The model training time is about 30 minutes (on a 16-core CPU), and the time taken to predict the relevance of a single pair of cases is about 20 milliseconds, supporting batch processing and real-time applications.

[0111] To enhance the interpretability of the model, the system designs an associated evidence extraction mechanism that can automatically identify the key text fragments, entities, and relationships leading to the association judgment and generate explanatory notes. For example, for two cases determined to be "factually associated", the system can extract and identify the paragraphs describing the events involved in common, highlighting the key time and location information to help users understand the basis for the association judgment. The model adopts an online learning mechanism that can be continuously optimized based on user feedback, and the accuracy of the association judgment has been improved from the initial 83% to over 90%, effectively supporting the intelligent identification and management of complex associated cases.

[0112] Secondly, based on the identified case groups with internal connections, the system designs a hierarchical task planning framework. This framework generates a preliminary action plan for associated case processing through goal decomposition, constraint identification, and action sequence generation. For example, for a group of interrelated financial cases, the system can design a hierarchical processing plan: the first layer determines the case processing order and dependencies, the second layer plans the specific processing steps for each case, and the third layer arranges resource allocation and time scheduling.

[0113] In implementation, the system uses symbolic planning and heuristic search methods. Symbolic planning represents case processing as a series of states and actions, and constructs a state transition model for case processing by defining preconditions and effects. Heuristic search uses domain knowledge to design heuristic functions to guide the search process and find high-quality preliminary action plans. For example, the system may plan to process key core cases first and then dependent cases to ensure the efficiency and consistency of the overall processing.

[0114] The design of the heuristic function is based on the professional knowledge and experience rules in the case handling field and is a key component guiding the search algorithm to explore the solution space. The heuristic function h(n) developed by the system evaluates the estimated cost from the current state node n to the target state and adopts the form of multi-feature weighted combination: h(n) = w1f1(n) + w2f2(n) +... + w k f k (n), where f1 to f k are feature functions, and w1 to w k are the corresponding weights. The core feature functions include: the processing step distance function (estimating the cost of completing the remaining necessary steps), the resource matching degree function (evaluating the matching degree of the current resource allocation and the case requirements), the quality risk function (predicting the quality problems that may be caused by the current path), and the processing efficiency function (estimating the time efficiency of handling the case according to the current path).

[0115] The processing step distance function is based on the critical path analysis of case handling and calculates the minimum distance in processing steps from the current state to the target state. For example, for a case in the "evidence investigation" stage, the system identifies the necessary steps such as "fact determination", "legal application", and "judgment document production" that still need to be completed to reach the closed case state through a case type-specific processing flow template, and estimates the standard processing cost of each step based on historical data and accumulates to obtain the total distance estimate. The resource matching degree function evaluates the support degree of the current resource configuration for case handling, considering the matching situation between the expertise field and experience level of the handling personnel and the case type and complexity. The higher the matching degree, the lower the estimated cost.

[0116] The quality risk function predicts the risk of possible quality problems, such as the risk of procedural defects, unclear fact determination, and incorrect legal application, through case characteristics and the current processing state. This function is implemented based on a random forest model trained with historical data of case quality assessment. By inputting case characteristics and current state characteristics, it outputs a quality risk score. The processing efficiency function is evaluated based on the time efficiency of the current processing path, considering case complexity, resource utilization, and the possibility of parallel processing steps, and estimates the time cost required to reach the target state.

[0117] The weight setting of the heuristic function adopts an adaptive mechanism, which dynamically adjusts the weights of each feature function according to the case type and processing stage. For example, for an urgent case approaching the expiration date, the system will increase the weight of the processing efficiency function; for a case with great social impact, the weight of the quality risk function will be increased. The weight adjustment follows the Bayesian optimization method and continuously optimizes the weight configuration based on the processing effects of historical cases. The design of the heuristic function ensures the admissible (not overestimating the actual cost) and consistent (satisfying the triangle inequality) characteristics, guaranteeing the optimality and efficiency of the search algorithm.

[0118] The heuristic function guides the search process implemented by the A* algorithm, which combines heuristic evaluation and actual known costs. It selects the nodes to expand according to the principle of f(n) = g(n) + h(n), where g(n) is the known cost from the initial state to the current node n, and h(n) is the estimated cost from n to the target state. To handle large-scale search spaces, the system adopts the iterative deepening A* variant, which finds a sub-optimal solution within a limited time by gradually increasing the search depth limit and supports interrupting at any time to return the current best solution, meeting the real-time response requirements. During the search process, the system uses pruning techniques to reduce the search space, including pruning invalid paths based on domain knowledge, pruning infeasible paths based on resource constraints, and pruning sub-optimal paths based on upper bound estimation.

[0119] The generation of a high-quality preliminary action plan adopts a three-stage strategy: skeleton planning, detail enrichment, and quality optimization. In the skeleton planning stage, based on the standard processing templates and key decision points of the case type, the system uses heuristic search to determine the main processing paths and key nodes, forming the plan skeleton. In the detail enrichment stage, the system assigns specific resources, time, and execution methods to each processing step in the skeleton, filling in the complete action details. In the quality optimization stage, the system applies quality inspection rules to evaluate the quality risk points of the preliminary plan and makes targeted adjustments, such as adding review steps for key links, optimizing resource allocation, and adjusting the processing order.

[0120] To handle the complexity of related cases, the system introduces a hierarchical planning mechanism. First, it plans the overall processing strategy for the group of related cases, determines the priority order and dependency relationships between cases, then plans the specific processing paths for each case separately, and finally conducts cross-case coordination and optimization to ensure the consistency and efficiency of handling related cases. Through this heuristic search method combined with domain knowledge, the system can generate high-quality case processing action plans within a reasonable time, effectively supporting the intelligent handling of complex cases.

[0121] Third, based on the preliminary action plan, the system develops a reinforcement learning optimization model. This model optimizes the processing strategy of related cases through state representation, reward mechanism, and policy network.

[0122] Specifically, the system takes the case processing state as the environment, the actions of the agent include case analysis, evaluation, and recommendation generation, and the reward signal comes from the processing efficiency and consistency metrics. Through interacting with the environment and policy iteration, the system continuously adjusts and optimizes the processing strategy to improve the processing effect.

[0123] In implementation, the system uses deep reinforcement learning methods, such as deep Q-learning or policy gradient methods. The system designs appropriate state representations to capture the key features of case handling; designs a multi-objective reward function to balance processing efficiency, quality, and consistency; and uses a deep neural network as the policy network to learn the mapping from states to optimal actions. By continuously trying different strategies and learning from experience, the system can discover the optimal processing solutions.

[0124] Finally, the system implements a dynamic fusion mechanism for planning and learning. This mechanism improves the adaptability of the strategy while ensuring the rationality of the plan through two-way feedback of using planning to guide exploration and learning to optimize planning.

[0125] For example, the initial plan provides the overall framework and constraints for handling cases. Reinforcement learning explores and optimizes specific strategies within this framework, and at the same time, the learning results can in turn adjust and improve the plan.

[0126] In specific implementation, the system adopts a hybrid architecture, which organically combines symbolic planning and reinforcement learning. The planning module provides high-level structured knowledge to guide the exploration direction of reinforcement learning; the reinforcement learning module discovers better execution strategies through actual interaction experience and feeds back the learning results to the planning module to optimize subsequent plans.

[0127] For example, the system may discover through reinforcement learning that when handling related financial cases, it is more efficient to first handle the core case with the largest amount and then handle its derivative cases. This discovery can be integrated into the planning knowledge to guide the handling of future similar cases.

[0128] Through the above steps, the system has successfully achieved the intelligent handling of complex related cases, improving the efficiency and consistency of handling related cases, and providing strong support for legal work.

[0129] S6: Based on the multi-agent case association processing results, apply the forced zero method sparse graph technology to achieve effective learning based on the case graph. Step S6 aims to apply the forced zero method sparse graph technology to achieve effective learning based on the case graph and improve the knowledge representation and reasoning ability of the system. Case data usually presents complex graph structure characteristics, including a large number of entities and relationships. Effective learning of this graph structure is crucial for case quality assessment.

[0130] First, the system constructs a case knowledge graph based on the multi-agent case association processing results.

[0131] This graph contains information such as case elements, legal concepts, and processing procedures, providing a data basis for graph learning. For example, for a certain type of intellectual property case, the graph may contain the following nodes: case associated with Company A, case associated with Company B, Patent X, Witness C, Expert Report D, etc.

[0132] In terms of specific implementation, the system identifies key entities from case materials through entity extraction technology, discovers the associations between entities through relationship recognition technology, and integrates information from different sources into a unified knowledge graph through knowledge fusion technology. For example, the system can extract entities such as parties, various bases, and legal standards from certain documents, identify the citation relationships, support relationships, etc. between them, and construct a knowledge graph representing the case structure.

[0133] Secondly, based on the case knowledge graph, the system designs a graph neural network model.

[0134] This graph neural network model includes graph convolutional layers, attention mechanisms, and message passing mechanisms, and realizes the preliminary modeling of the case graph structure through node feature extraction and edge relationship learning. The typical architecture of a graph neural network includes: an input layer (initial node features), graph convolutional layers (aggregating adjacent node information), attention mechanisms (focusing on important nodes and edges), and an output layer (node representation or graph representation).

[0135] In implementation, the system adopts advanced graph neural network architectures such as Graph Convolutional Network (GCN) or Graph Attention Network (GAT).

[0136] For example, through GCN, the system can aggregate the neighbor information of nodes and learn the representation of nodes; through GAT, the system can focus on important nodes and edges and improve the learning efficiency. These models can capture the complex patterns and structural features in the case graph and provide effective representations for subsequent tasks.

[0137] The graph attention network is used for the structure learning of the case knowledge graph, and its construction and training process are as follows: The model architecture adopts a multi-layer graph attention network structure: Input layer: The initial node feature vector (128 dimensions) is composed of the concatenation of text semantic features (64 dimensions) and structural features (64 dimensions); The first layer of graph attention layer: 8 attention heads, each head outputs 32-dimensional features, and after merging, it is 256 dimensions; Batch normalization layer: Standardize the feature distribution to stabilize the training process; The second layer of graph attention layer: 8 attention heads, each head outputs 32-dimensional features, and after merging, it is 256 dimensions; Global pooling layer: Use graph pooling with attention mechanisms to compress the graph-level features into a 384-dimensional vector; Output layer: Designed according to downstream tasks, such as using a Softmax output layer for case classification tasks; Key parameters for implementing the attention mechanism: Attention vector dimension: 64; LeakyReLU negative slope: 0.2; Attention dropout rate: 0.6; Feature dropout rate: 0.5; L0 sparsity regularization parameter: Temperature parameter (initial value): 2.0; Lower temperature limit: 0.5; Temperature annealing rate: 0.99; Gating variable initialization method: Uniform distribution U(-0.1, 0.1); L0 regularization coefficient (λ): 0.001; Training hyperparameter range: Learning rate: [0.0001, 0.0005, 0.001, 0.005], finally selected 0.001; Learning rate scheduling: ReduceLROnPlateau, factor is 0.5, patience value is 10; Batch size: [16, 32, 64], finally selected 32 (according to GPU memory limit); Number of training epochs: 200, using early stopping strategy, patience value is 30; Gradient clipping threshold: 1.0; Weight decay: [0.0001, 0.0005, 0.001], finally selected 0.0005; The training data contains 10,000 case subgraphs. Each subgraph contains an average of 30 nodes and 80 edges. Node types include cases, people, organizations, legal standards, facts, documents, etc. Downstream tasks include case classification, association prediction, and quality assessment, etc. The data is divided into training set, validation set, and test set according to the ratio of 7:1:2.

[0138] The training process implements a dynamic temperature adjustment mechanism. As the training progresses, the temperature parameter gradually decreases, making the distribution of gating variables gradually shift from continuous to discrete, promoting the formation of a sparse structure. At the same time, the model adopts a weight initialization technique, assigning larger values to the initial gating variables of edges with high importance to accelerate convergence.

[0139] The model is trained on the GPU for about 4 hours. The case classification accuracy on the test set reaches 94.2%, the F1 value of association prediction reaches 91.3%, and the mean squared error of quality assessment is 0.043. The sparsification effect is significant. Finally, about 25% of the original edges are retained. The number of model parameters is reduced from the original 3.5M to 1.2M, and the inference speed is increased by about 3 times. The average time for processing medium-scale case graphs (50 nodes) on the CPU is reduced from 120 milliseconds to 40 milliseconds.

[0140] Importantly, the sparsified model retains the key case knowledge structure. Through visual analysis, it is found that the retained edges are mainly concentrated in the connections between core legal relationships, key factual bases, and case entities, greatly improving the interpretability and practical value of the model.

[0141] Third, the system implements a forced zero method sparsification mechanism based on the graph neural network model. This mechanism includes L0 regularization, gated activation, and gradient estimation. By explicitly controlling the model complexity and the number of activated nodes, it reduces the redundancy of the graphical representation and improves the learning efficiency. Different from traditional L1 regularization, the forced zero method sparsification directly sets unimportant connections to exactly zero through L0 regularization, thus significantly reducing the model complexity.

[0142] In terms of specific implementation, the system adds a binary gated variable to each edge to determine whether to retain the edge; realizes the differentiable optimization of L0 regularization through hard gating approximation techniques (such as Concrete Distribution); automatically learns the optimal sparse structure during the training process, and in typical cases, most edges can be pruned while maintaining performance.

[0143] For example, when the system analyzes a complex commercial litigation case, the original case graph may contain more than 300 entities and more than 2,000 relationships. After applying the forced zero method sparsification, only about 400 key relationships may be retained, but these relationships accurately capture the core legal relationships of the case.

[0144] Finally, the system develops an adaptive learning algorithm based on the forced zero method sparsification mechanism. This algorithm includes meta-learning, knowledge distillation, and incremental training. By dynamically adjusting the learning strategy and model structure, it adapts to the graphical characteristics of different types of cases. For example, when dealing with a new type of environmental pollution liability case, the system can quickly adapt to the graph structure characteristics of the new case type through meta-learning, extract key knowledge from the complex model using knowledge distillation, and continuously optimize the model through incremental training.

[0145] In implementation, the system uses the meta-learning method to learn the common characteristics of different types of cases, accelerating the adaptation process for new case types; uses knowledge distillation technology to transfer the knowledge of complex models to simpler and more efficient models; adopts an incremental training strategy to efficiently update the model when new case data arrives, avoiding the overhead of retraining. These technologies together improve the learning efficiency and adaptability of the system.

[0146] Through the above steps, the system has successfully achieved effective learning based on case graphs, improving the computational efficiency and model interpretability, and providing strong technical support for case quality assessment.

[0147] S7: Based on the effective learning results of the case graph, construct a case quality assessment knowledge graph to realize the intelligence and precision of case quality management and provide data support for legal decision-making.

[0148] As a powerful knowledge representation and reasoning tool, the knowledge graph can support complex query, analysis, and recommendation functions and provide data support for legal decision-making.

[0149] In the embodiment of the present application, the system integrates multi-source evaluation data based on the effective learning results of the case graph. This data includes information such as evaluation criteria, case characteristics, evaluation results, and improvement suggestions. Through technologies such as data cleaning, format conversion, and redundancy elimination, a unified data set is formed.

[0150] For example, the system can integrate the results from different evaluation agents, including case retrospective analysis, quality scoring, improvement suggestions, etc., convert them into a unified data format, eliminate redundancy and inconsistencies, and form a high-quality data foundation.

[0151] Specifically, the system uses data integration technology to process heterogeneous data sources, uses data cleaning technology to identify and correct errors, uses data conversion technology to unify data formats, and uses redundancy elimination technology to improve data efficiency. For example, for multiple evaluation results of the same case, the system can extract common points and differences, merge them into a more comprehensive evaluation record, and improve data quality and usage efficiency.

[0152] Secondly, based on the unified data set, the system designs a knowledge graph schema. This schema includes concept hierarchies, relationship types, and attribute definitions. Through ontology engineering and semantic modeling, a conceptual framework for the case quality assessment field is constructed to guide the construction of the knowledge graph. For example, the system can define concepts such as cases, evaluation dimensions, evaluation indicators, evaluation results, etc., as well as the hierarchical and associative relationships between them to form a structured knowledge system.

[0153] In implementation, the system uses ontology engineering methods to define domain concepts and relationships and uses semantic modeling techniques to construct semantic connections between concepts. For example, the system can define sub-concepts such as "civil and commercial cases" under the "case" concept, define the relationship between the "evaluation" concept and the "case" concept, define the attributes and value ranges of "evaluation results", and construct a complete domain knowledge framework.

[0154] Thirdly, based on the knowledge graph schema and the unified data set, the system realizes knowledge extraction and graph construction. This process uses technologies such as named entity recognition, relationship extraction, and event detection to convert unstructured and semi-structured data into knowledge triples and construct an initial case quality assessment knowledge graph.

[0155] For example, the system can extract information such as case ID, evaluation dimension, score, and problem description from the evaluation report, and transform it into knowledge triples such as "Case A - in Dimension B - with a score of C", "Case A - has a problem - Problem D".

[0156] In terms of specific implementation, the system uses natural language processing technology to process text data, uses information extraction technology to identify entities and relationships, and uses knowledge representation technology to organize information into knowledge triples.

[0157] For example, the system can use named entity recognition to identify entities such as case numbers, names of case handlers, and evaluation indicators in the text, and use relation extraction to identify relationships such as "evaluation" and "has a problem", and construct a triple network representing case quality assessment knowledge.

[0158] Finally, based on the initial knowledge graph, the system develops knowledge reasoning and application interfaces. This part includes a knowledge reasoning engine for rule reasoning, path reasoning, and statistical reasoning, as well as standardized application interfaces, supporting multi-dimensional queries, intelligent analysis, and precise recommendations for case quality management. For example, the system can infer the root cause of case quality problems based on the knowledge graph, analyze common problems of different types of cases, recommend targeted improvement plans, and provide data support for legal decision-making.

[0159] In terms of implementation, the system uses rule reasoning to process explicit logical relationships, uses path reasoning to discover implicit association relationships, and uses statistical reasoning to analyze data patterns and trends. For example, through rule reasoning, the system can infer "high risk of procedural violation" based on "low score for procedural legality"; through path reasoning, the system can discover the association between "cases handled by Case Handler A" and "high evaluation scores"; through statistical reasoning, the system can analyze the evaluation distribution of different types of cases and discover potential problems and optimization directions.

[0160] At the same time, the system develops standardized application interfaces to support various query and analysis functions. For example, the interface can support querying evaluation results according to conditions such as case type, evaluation dimension, time period, etc.; support analyzing the time trend and type distribution of case quality; support recommending improvement plans for specific problems, etc. These functions provide powerful decision-making support tools for legal affairs managers, realizing the intelligence and precision of case quality management.

[0161] Through the above steps, the system has successfully constructed a case quality assessment knowledge graph, realized the intelligence and precision of case quality management, provided data support for legal decision-making, and promoted the digital transformation and quality improvement of legal affairs work.

[0162] Among them, combining Figure 1 and Figure 2 , step S1 specifically includes: S1.1: Based on legal regulatory documents and trial practices, extract the regulatory requirements and standards related to case quality assessment through text extraction and semantic analysis techniques, and form a structured dataset of regulatory requirements.

[0163] In actual work, the accuracy and timeliness of the evaluation criteria directly affect the reliability of the evaluation results. The system first collects regulatory documents such as relevant regulations, interpretive documents, and guiding opinions to form an original document set. These documents contain various standards and requirements for case handling and are the basic data sources for constructing the evaluation criteria library.

[0164] For example, extract various specific indicators for case evaluation from relevant case quality inspection measures, and extract the regulatory requirements for case handling from guiding cases issued by authoritative departments.

[0165] In a specific embodiment, S1.1 may include: S1.1.1: Based on legal regulatory documents and trial practices, collect various regulatory documents such as legal regulations, legal interpretations, and guiding opinions to form an original document set.

[0166] S1.1.2: Based on the original document set, preprocess the documents through natural language processing techniques, including text cleaning, word segmentation, part-of-speech tagging, and syntactic analysis, to generate structured text. S1.1.3: Based on the structured text, apply named entity recognition and relationship extraction techniques to identify the evaluation elements and standard items in the text, and extract the semantic relationships between them to construct a preliminary network of regulatory elements. S1.1.4: Based on the network of regulatory elements, through semantic similarity calculation and clustering analysis, merge and standardize the repeated or similar evaluation elements to form a structured dataset of regulatory requirements. S1.2: Based on the dataset of regulatory requirements, classify them according to case types and evaluation dimensions, construct a hierarchical evaluation criteria model, and generate an initial evaluation criteria library. Classify and organize the extracted evaluation criteria according to different case types (such as people's livelihood cases, commercial cases, etc.) and evaluation dimensions (such as procedural normativity, appropriateness of entity handling, document normativity, etc.) to construct a clearly hierarchical evaluation criteria system. For example, for commercial dispute cases, an evaluation criteria system can be constructed that includes multiple dimensions such as acceptance procedures, service procedures, mediation procedures, trial procedures, and certain documents. Each dimension contains multiple specific indicators to form a complete criteria system.

[0167] S1.3: Based on the initial evaluation criteria library, design a dynamic update mechanism that includes detection of new regulatory documents, identification of semantic changes, and automatic update of standard items to achieve the real-time update ability of the evaluation criteria library. Step S1.3 is to design a dynamic update mechanism based on the initial evaluation standard library, which includes the detection of new specification documents, the identification of semantic changes, and the automatic update of standard items, so as to realize the real-time update ability of the evaluation standard library.

[0168] To maintain the timeliness of the evaluation criteria, the system needs to be able to automatically identify and process newly released regulatory documents. When new regulations or interpretive documents are released, the system can automatically detect and extract relevant evaluation criteria and update them to the standard library. For example, when an authoritative department releases a new interpretive document, the system can automatically identify the content related to case quality evaluation therein and integrate it into the existing evaluation standard library to ensure the timeliness and authority of the standard library.

[0169] S1.4: Based on the dynamic update mechanism, develop a standard conflict detection algorithm and coordination strategy. Through semantic similarity analysis and rule priority determination, solve the possible conflict problems in the standard update process, and realize the real-time update and optimization of the evaluation standard library.

[0170] During the standard update process, there may be inconsistencies or conflicts between the old and new specifications, and these conflicts need to be automatically detected and coordinated through algorithms. The system identifies potential conflicts through semantic similarity analysis and automatically coordinates conflicts through rule priority determination (such as based on factors such as the level of effectiveness and the formulation time), ensuring the consistency and reliability of the evaluation standard library. For example, when a new interpretive document is inconsistent with the original specification, the system can automatically decide which version of the standard to adopt based on factors such as the document's effectiveness level and release time, ensuring the authority and consistency of the standard library.

[0171] The standard conflict detection algorithm is an automated mechanism for identifying potential contradictions or inconsistencies in the evaluation standard library. This algorithm first represents the evaluation criteria as a triple structure <subject, attribute, value> using a formal method, such as <service procedure, time limit requirement, within 3 days>. The algorithm realizes conflict detection through three core steps: semantic equivalence analysis, specification requirement comparison, and logical conflict derivation. Semantic equivalence analysis uses a word vector model and a thesaurus to identify standard items with different expressions but similar semantics. For example, "service time requirement" and "service deadline regulation" are identified as semantically equivalent. In the specification requirement comparison stage, it compares whether the values of semantically equivalent standard items are consistent. If it is found that one standard stipulates that "the service deadline is within 3 days", while another standard stipulates that "the service time does not exceed 5 days", it is marked as a potential conflict. Logical conflict derivation uses automated reasoning techniques to identify indirect conflicts based on predefined conflict patterns. For example, if standard A requires that "evidence exchange must be completed before a certain procedure", standard B requires that "evidence exchange must be carried out after the expiration of the time limit for presenting evidence", and standard C requires that "the deadline for presenting evidence shall not be later than the day before a certain procedure", then through reasoning, it can be found that there is a logical conflict among these three standards.

[0172] The standard conflict detection algorithm is implemented using a graph structure representation and path analysis method. The system represents the evaluation criteria as a knowledge graph, with nodes being the subjects and values, and edges being the attribute relationships. Conflict detection is transformed into a search problem for specific patterns in the graph, and subgraph matching is performed through six designed conflict pattern templates (direct contradiction, numerical inconsistency, inclusion conflict, mutually exclusive conditions, temporal conflict, derivation conflict) to identify potential conflicts. The time complexity of the algorithm is O(n²), where n is the number of standard items, and it can be reduced to O(n log n) through index optimization, supporting real-time conflict detection for large-scale standard libraries.

[0173] The coordination strategy is a systematic approach to resolving identified standard conflicts and consists of four key components: conflict classification, rule priority determination, expert consultation mechanism, and conflict resolution record. Conflict classification divides the detected conflicts into hard conflicts (completely mutually exclusive regulations) and soft conflicts (partially overlapping or ambiguous regulations) according to their nature. Rule priority determination is based on a five-level determination criterion: legal effect level (higher-level laws prevail over lower-level laws), time of enactment (new rules prevail over old rules), professional pertinence (specific regulations prevail over general regulations), source authority (regulations issued by authoritative institutions take precedence), and scope of application (regulations with a clear scope of application take precedence). The system assigns a priority score to each standard and determines the priority order between conflicting standards through weighted calculation.

[0174] For hard conflicts, the coordination strategy adopts the automatic replacement method, retaining the high-priority standard and discarding the low-priority standard. For example, when a new regulation issued by a certain department in 2023 conflicts with an old regulation in 2018, the system automatically adopts the new regulation and marks the old regulation as discarded. For soft conflicts, the coordination strategy adopts the merging and reconciliation method, attempting to retain the reasonable parts of both and eliminate the contradictions. For example, when one standard stipulates that "the quality assessment ratio of major cases is not less than 15%", and another standard stipulates that "the overall quality assessment ratio is controlled between 10% and 20%", the system will automatically merge them into "the quality assessment ratio of major cases is 15% - 20%, and that of other cases is 10% - 15%".

[0175] When automatic coordination cannot resolve complex conflicts, the system activates the expert consultation mechanism, pushing the conflict details, conflict analysis, and possible solutions to experts in the designated field for manual judgment. Expert decisions are recorded by the system and form precedents for conflict resolution, which are used to guide the handling of future similar conflicts. The system maintains a conflict resolution knowledge base, recording all conflict cases and their solutions, supporting case-based reasoning, and realizing the continuous optimization and learning of the coordination strategy. Through this multi-level coordination strategy, the system can effectively handle various conflicts during the standard update process, ensuring the consistency and reliability of the evaluation standard library.

[0176] Combined with Figure 1 and Figure 3, Step S2 specifically includes: S2.1: Based on the procedural norms in the evaluation standard library and combined with the reasoning ability of the multi-agent system, design a sub-chain for evaluating the legality of procedures, including case-filing procedures, service procedures, evidence presentation procedures, and trial procedure links, to achieve a comprehensive evaluation of the legality of case procedures. This sub-chain covers the entire process of case handling, including key links such as acceptance procedures, service procedures, evidence presentation procedures, and trial procedures. Taking the acceptance procedure as an example, the system can check whether the case meets the acceptance conditions, whether the review is completed within the specified time limit, and whether the rights and obligations of the parties are informed as required, so as to achieve a comprehensive evaluation of the standardization of case procedures. The system represents procedural norms as a series of checkpoints and logical relationships, and through the reasoning ability of the multi-agent system, analyzes the procedural descriptions in the case materials to determine whether they conform to the corresponding norms.

[0177] S2.2: Based on the substantive norms in the evaluation standard library and the evaluation results of the legality of procedures, design a sub-chain for evaluating the appropriateness of substantive handling, including fact-finding, adoption of various types of evidence, application of laws, and reasoning in judgments, to achieve a precise evaluation of the appropriateness of substantive handling of cases. In S2.2, based on the substantive norms in the evaluation standard library and the evaluation results of procedural standardization, design a sub-chain for evaluating the appropriateness of substantive handling. This sub-chain focuses on the substantive handling process of cases, including fact-finding, adoption of various types of evidence, application of regulations, and reasoning in judgments. For example, in the link of adopting various types of evidence, the system can evaluate whether the compliance, relevance, and probative force of various types of evidence in the case have been fully considered, and whether the contradictions between various types of evidence have been reasonably resolved.

[0178] The system analyzes the substantive content in the case materials through its powerful semantic understanding and logical reasoning abilities to evaluate the appropriateness of its handling. For example, in the link of applying regulations, the system can analyze whether the evidence in certain documents is accurate and complete, whether there are application errors or omissions, so as to evaluate the appropriateness of applying regulations.

[0179] S2.3: Based on the document norms in the evaluation standard library and the evaluation results of the appropriateness of substantive handling, design a sub-chain for evaluating the standardization of legal documents, including document format, language expression, logical structure, and content integrity, to achieve a detailed evaluation of the standardization of legal documents. This sub-chain focuses on the form and content norms of documents, including document format, language expression, logical structure, content integrity, etc. For example, the system can evaluate whether the format of certain documents is standardized, whether the language is accurate and concise, whether the argumentation is well-structured, and whether the reasoning is sufficient.

[0180] The system conducts multi-dimensional evaluations on documents through text analysis capabilities. For example, by analyzing the paragraph structure, keyword distribution, and language style of the documents, it evaluates the logical structure and language expression of the documents; by comparing the content of the documents with the case facts and relevant regulations, it evaluates the integrity and accuracy of the document content.

[0181] S2.4: Based on the sub-chain for evaluating procedural legality, the sub-chain for evaluating the appropriateness of entity handling, and the sub-chain for evaluating the normativity of legal documents, through the multi-task learning and chain reasoning techniques of the large language model, integrate each evaluation sub-chain, and develop an inference control mechanism that includes evaluation sequence control, intermediate result feedback, and evaluation depth adjustment to form a complete case quality evaluation inference chain.

[0182] By integrating the sub-chain for evaluating procedural normativity, the sub-chain for evaluating the appropriateness of entity handling, and the sub-chain for evaluating document normativity, and developing an inference control mechanism such as evaluation sequence control, intermediate result feedback, and evaluation depth adjustment, a comprehensive and orderly evaluation of case quality is achieved. For example, the system first evaluates the procedural normativity of the case. If procedural defects are found, the evaluation depth of subsequent links is correspondingly reduced; if the evaluation result of procedural normativity is good, it further deeply evaluates the appropriateness of entity handling and document normativity. Through this dynamically adjusted evaluation strategy, the system can concentrate evaluation resources on key issues, improve evaluation efficiency, and at the same time ensure the comprehensiveness and accuracy of the evaluation.

[0183] Through the above steps, the system has successfully constructed an evaluation standard library covering various types of cases and its dynamic update mechanism, and designed a complete case quality evaluation inference chain, providing a solid foundation for subsequent intelligent evaluation and analysis. These technological innovations effectively solve problems such as untimely standard updates and incomplete evaluations in traditional evaluation methods, and greatly improve the accuracy and efficiency of case quality evaluation.

[0184] Combined with Figure 1 and Figure 4 , step S3 specifically includes: S3.1: Based on the evaluation inference chain, develop a case backtracking intelligent agent with the ability to analyze the entire process of the case. Through time series analysis and causal reasoning, achieve backtracking analysis of the entire process of case handling, and identify key nodes and existing problems in case handling.

[0185] Step S3.1 is to develop a case backtracking intelligent agent with the ability to analyze the entire process of the case based on the evaluation inference chain. This intelligent agent realizes backtracking analysis of the entire process of case handling through time series analysis and causal reasoning, and identifies key nodes and existing problems in case handling.

[0186] For example, in the case of a complex commercial case, the intelligent agent can sort out the time sequence of the whole process from case acceptance, preliminary preparation, investigation and discussion to conclusion formation, identify key nodes, such as the exchange of various evidences, the determination of major dispute focuses, the testimonies of key witnesses, etc., and analyze the processing quality of these nodes. Specifically, the intelligent agent uses timeline analysis and event extraction techniques to extract key events and time points of case processing from case materials and construct a time sequence map of case processing. Then, through causal reasoning techniques, it analyzes the causal relationships between events, identifies key nodes and potential problems in case processing. For example, the intelligent agent can find that the deficiencies in the collection link of various evidences lead to difficulties in subsequent fact-finding, or insufficient preliminary preparation leads to a decrease in processing efficiency.

[0187] S3.2: Based on the case backtracking analysis results and the evaluation reasoning chain, develop a quality scoring intelligent agent, and through multi-dimensional evaluation and weight adaptive algorithm, realize the quantitative scoring of case quality and generate an evaluation result including the overall score and sub-dimensional scores. Step S3.2 is to develop a quality scoring intelligent agent based on the case backtracking analysis results and the evaluation reasoning chain. This intelligent agent realizes the quantitative scoring of case quality through multi-dimensional evaluation and weight adaptive algorithm.

[0188] For example, in the case of a people's livelihood case, the intelligent agent can evaluate from three dimensions: procedural standardization, entity handling appropriateness, and document standardization, and further break them down into multiple sub-dimensions such as acceptance procedure, service procedure, fact-finding, regulation application, document format, and argument logic. Weights are assigned to each dimension and sub-dimension, and the overall score and sub-dimensional scores are calculated. In implementation, the intelligent agent uses a multi-level evaluation model to organize evaluation indicators into a hierarchical structure and dynamically adjusts the weights of each indicator through an adaptive weight algorithm.

[0189] For example, according to the case type and complexity, the intelligent agent may increase the weights of the fact-finding and regulation application dimensions, or according to the case dispute focus, increase the weights of relevant evaluation dimensions to ensure that the scoring results objectively reflect the case quality. This flexible weight adjustment mechanism enables the evaluation system to adapt to the characteristics of different types of cases and provide more accurate quality evaluation.

[0190] S3.3: Based on the quality scoring results and the evaluation standard library, develop an improvement suggestion intelligent agent, and through difference analysis and best practice matching, generate specific and actionable improvement suggestions for the dimensions with scores lower than the first preset threshold, providing guidance for improving case quality. This intelligent agent generates specific and actionable improvement suggestions for the dimensions with scores lower than the first preset threshold (such as 70 points) through difference analysis and best practice matching.

[0191] For example, for cases where the specified applicable dimension score is lower than the first preset threshold, the intelligent agent can point out existing application errors or omissions and provide correct application suggestions; for cases where the document standardization score is lower than the first preset threshold, the intelligent agent can provide specific methods for optimizing the document structure or enriching the reasoning.

[0192] In terms of specific implementation, the intelligent agent uses difference analysis technology to compare the differences between the case processing process and best practices, and identify deficiencies. Then, through best practice matching technology, it retrieves cases from the knowledge base that are similar to the current case but have higher processing quality, extracts the successful experiences from them, and generates targeted improvement suggestions. For example, the intelligent agent can recommend suitable methods for collecting various types of evidence, ideas for fact-finding, or applicable cases according to the characteristics of the case to help improve the quality of case processing. These specific and feasible suggestions not only contribute to improving the handling of the current case but also provide opportunities for the handlers to learn and improve.

[0193] S3.4: Based on the case backtracking intelligent agent, the quality scoring intelligent agent, and the improvement suggestion intelligent agent, design an intelligent agent collaboration framework that includes task assignment, information sharing, and result integration. Optimize the collaboration strategy through reinforcement learning to achieve efficient collaborative work among the intelligent agents, form a complete multi-intelligent agent system, and realize intelligent evaluation and analysis of case quality.

[0194] Based on the above three intelligent agents, design an intelligent agent collaboration framework, optimize the collaboration strategy through reinforcement learning, and achieve efficient collaborative work among the intelligent agents.

[0195] For example, the analysis results of the case backtracking intelligent agent are directly input into the quality scoring intelligent agent, and the scoring results of the quality scoring intelligent agent are then input into the improvement suggestion intelligent agent, forming a closed loop of information flow; at the same time, the output of the improvement suggestion intelligent agent can be fed back to the case backtracking intelligent agent to help it improve the analysis method, forming a feedback optimization mechanism.

[0196] In implementation, the system uses a collaboration strategy optimization method based on reinforcement learning. By continuously adjusting the interaction methods and information sharing strategies among the intelligent agents, it improves the overall collaboration efficiency. For example, the system can learn when in-depth information exchange is required among the intelligent agents and when they can work in parallel, thus improving the evaluation efficiency while ensuring the evaluation quality. This intelligent collaboration framework enables the entire evaluation system to operate as an organic whole, giving full play to the expertise of each intelligent agent and maximizing the overall effectiveness.

[0197] Combined with Figure 1 and Figure 5 , step S4 specifically includes: S4.1: Based on the multi-agent system, extract features and perform mathematical modeling on the unplanned online tasks, and construct a task allocation optimization model that includes task features, resource constraints, and time windows. Step S4.1 is to extract features and perform mathematical modeling on the unplanned online tasks based on the multi-agent system, and construct a task allocation optimization model that includes task features, resource constraints, and time windows.

[0198] In actual work, there are often situations where sudden or unconventional cases need to be handled urgently, and the system needs to be able to allocate resources efficiently to ensure that these tasks are processed in a timely manner.

[0199] For example, for a batch of sudden case quality assessment tasks, the system can extract the features of each task, such as case type, complexity, priority, estimated processing time, etc., and construct a mathematical model for task allocation.

[0200] Specifically, the system models the task allocation problem as an optimization problem: minimize f(x) + g(z), where f(x) represents the efficiency objective of task allocation (such as the total completion time), g(z) represents the resource constraint objective (such as workload balance), and at the same time satisfy the constraint conditions that tasks must be fully allocated and do not exceed the resource capacity. This mathematical modeling provides a theoretical basis for subsequent algorithm design, ensuring the scientificity and rationality of task allocation.

[0201] S4.2: Based on the task allocation optimization model, design an inexact alternating direction method of multipliers (ADMM) algorithm framework, and transform the global optimization problem into multiple sub-problems that can be solved in parallel through steps of problem decomposition, variable update, and multiplier adjustment.

[0202] Step S4.2 is to design an inexact alternating direction method of multipliers (ADMM) algorithm framework based on the task allocation optimization model.

[0203] This algorithm transforms the global optimization problem into multiple sub-problems that can be solved in parallel through steps of problem decomposition, variable update, and multiplier adjustment.

[0204] For example, for the case where 10 agents need to handle 30 assessment tasks, the system can decompose the global task allocation problem into local problems for 10 agents, and each agent calculates the subset of tasks it should receive in parallel.

[0205] The core of the algorithm framework is to iteratively solve the problem: first update the task allocation variable x, then update the resource usage variable z, and finally update the Lagrange multiplier u. Through this iterative method, the algorithm can efficiently solve the task allocation problem in a distributed environment and meet the requirements of large-scale task allocation. This method of decomposition and solution greatly reduces the computational complexity, enabling the system to handle a large number of sudden tasks.

[0206] S4.3: Based on the inexact alternating direction method of multipliers algorithm framework, develop an inexact update strategy that includes adaptive precision control, early stopping strategy, and approximate solution. By reducing the computational precision requirements for each iteration, the algorithm convergence speed is accelerated. Step S4.3 is to develop an inexact update strategy based on the inexact alternating direction method of multipliers algorithm framework. This strategy accelerates the algorithm convergence speed by reducing the computational precision requirements for each iteration.

[0207] In traditional algorithms, the subproblems need to be solved exactly in each iteration, resulting in a large computational overhead. The inexact update strategy allows the use of approximate solutions in each iteration as long as certain precision requirements are met, greatly improving the algorithm efficiency.

[0208] Specifically, the system uses adaptive precision control, early stopping strategy, and approximate solution methods. For example, at the initial stage of the algorithm, solutions with lower precision can be used, and as the iteration approaches convergence, the precision requirements are gradually increased; or an iterative method with early termination is used to solve the subproblems, and the iteration is stopped as long as the improvement of the solution is less than a certain threshold. These strategies greatly reduce the computational complexity and improve the system's response ability to real-time tasks, especially suitable for handling sudden tasks that require quick responses.

[0209] S4.4: Based on the inexact update strategy, develop a distributed computing framework that supports multi-node parallel computing. Through task partitioning, node coordination, and result aggregation, efficient distributed processing of unscheduled online tasks is achieved.

[0210] Step S4.4 is to develop a distributed computing framework that supports multi-node parallel computing based on the inexact update strategy. This framework achieves efficient distributed processing of unscheduled online tasks through task partitioning, node coordination, and result aggregation.

[0211] For example, the system can process subproblems in parallel on multiple computing nodes, synchronize the calculation results of each node through a central coordinator, and adjust the global allocation scheme. In implementation, the system adopts a master-slave architecture. The central node is responsible for task decomposition and result aggregation, and the slave nodes are responsible for solving subproblems. Communication between nodes is carried out through a message passing mechanism to share necessary variables and results. Through this distributed architecture, the system can make full use of computing resources, improving the processing throughput and response speed.

[0212] For example, when 10 high-priority cases enter the system simultaneously, the distributed framework can assign them to 3 agents with different specializations within a few seconds. Each agent receives a suitable subset of tasks according to its own specialization and current load, while ensuring the optimal overall evaluation quality and the shortest completion time. Compared with traditional methods, this allocation can significantly improve the processing efficiency and ensure that important tasks are processed in a timely manner.

[0213] Exemplarily, taking a practical application as an example, assume that the system suddenly receives a batch of 10 new cases that need to be urgently processed on Monday morning. These cases come from different business entities and require a preliminary assessment to be completed within 48 hours.

[0214] The system first extracts features from these cases, including case type (such as commercial disputes, intellectual property related, etc.), case complexity (calculating a score based on the amount of documents, the number of disputed points, etc.), priority (based on time urgency and importance), etc. For example, case 1 may be an intellectual property related case with a complexity score of 8.5 (out of 10), high priority, and an estimated processing time of 6 hours; case 2 may be a commercial contract dispute with a complexity score of 6.3, medium priority, and an estimated processing time of 4 hours.

[0215] At the same time, the system also needs to consider the status of available resources. For example, there are 5 agents available currently, and each agent has different expertise areas and remaining processing capabilities. The system constructs these information into a mathematical optimization model: the objective function is to minimize the weighted sum of the total completion time and the load imbalance degree, and the constraint conditions include that all cases must be assigned, the load of each agent does not exceed the capacity limit, and high-priority cases must be processed first, etc. This mathematical modeling method formalizes the complex task allocation problem into a solvable optimization problem, laying a foundation for subsequent algorithm applications.

[0216] Taking the scenario of allocating the above 10 cases to 5 agents as an example, the traditional method needs to consider all possible allocation combinations, and the computational complexity is extremely high. While using the ADMM algorithm, the system decomposes the problem into: each agent independently decides which cases it should process (local sub-problem), and then through a central coordination mechanism, ensures that the global constraints are satisfied (such as non-repeated case allocation). Specifically, the algorithm framework contains three main steps: first, update the agent variable x (indicating which cases each agent should process), then update the resource allocation variable z (indicating the final allocation plan of the cases), and finally update the Lagrange multiplier u (adjusting the consistency between the local solution and the global solution). In mathematical form, at the k-th iteration:

[0217]

[0218]

[0219] where fi(xi) represents the local objective function of agent i (such as minimizing the processing time), g(z) represents the global objective function (such as load balancing), A and B are constraint matrices, c is a constraint vector, and ρ is a penalty parameter. Through this decomposition, the system can solve the sub-problems of each agent in parallel on different computing nodes, greatly improving the computational efficiency.

[0220] In practical applications, accurately solving each sub-problem usually incurs a large computational cost and is often unnecessary, especially in the initial stage of algorithm iteration. The inexact update strategy allows the use of approximate solutions under certain accuracy conditions, thus accelerating the entire algorithm process. Taking the above case allocation problem as an example, assume that agent 1 needs to solve its local sub-problem to decide which cases to handle. Traditional methods may require accurate solutions through methods such as quadratic programming, which takes a long time to calculate. By adopting the inexact update strategy, the agent can use the gradient descent method for a finite number of iterations (e.g., 10 steps) to obtain an approximate solution to the local sub-problem. Specifically, the system designs three inexact update techniques: Adaptive precision control: Use a lower precision (e.g., relative error of 0.1) in the initial stage of the algorithm, and gradually increase the precision requirement as the iteration approaches convergence (eventually reaching 0.001). For example, for the case allocation problem, in the initial few rounds of iteration, it may only be necessary to roughly determine which agents are suitable for handling which types of cases, and then fine-tune the specific allocation plan later.

[0221] Early stopping strategy: During the solution process of the sub-problem, terminate the iteration in advance when the improvement amplitude of the solution is less than a preset threshold (e.g., 0.01). For example, if after 3 consecutive iterations of an agent, the change in its case allocation plan does not exceed 5%, it can be considered that the optimal solution has been approached, and the solution of this sub-problem can be terminated in advance.

[0222] Approximate solution method: Use heuristic algorithms or simplified models with lower computational complexity to solve the sub-problem. For example, for the case selection sub-problem of the agent, a greedy algorithm can be used to make a preliminary allocation based on the case-agent matching degree, rather than solving the complete combinatorial optimization problem.

[0223] Through these inexact update strategies, in actual tests, although the number of algorithm iterations may increase slightly (e.g., from 50 to 60 times), the computational time for each iteration is significantly reduced (e.g., from 2 seconds to 0.2 seconds), and the overall computational time can be reduced by more than 80%, greatly improving the system's response speed to sudden tasks.

[0224] In an actual system, when faced with a large number of sudden cases (such as 100 related cases entering the system simultaneously due to a major commercial event), single-machine computing may not be able to meet the real-time processing requirements. The distributed computing framework allows the system to parallelly process the task allocation problem on multiple servers. The specific architecture adopts a master-slave design: the central node (master node) is responsible for problem decomposition, task allocation, and result aggregation; multiple computing nodes (slave nodes) are responsible for parallelly solving sub-problems. Taking a specific case as an example, when the system needs to handle the allocation problem of the above 100 cases, the master node first decomposes the problem into local sub-problems of 20 agents; then allocates these sub-problems to 5 computing nodes, and each node processes the sub-problems of 4 agents; each computing node parallelly solves and returns the results; the master node aggregates these results, updates the global variables z and multiplier u, and then enters the next iteration.

[0225] To ensure the efficiency and reliability of distributed computing, the system has implemented the following key mechanisms: Dynamic load balancing: Dynamically adjust the task allocation according to the performance and current load of the computing nodes. For example, if it is found that a certain computing node has a slow processing speed, its task volume will be reduced in the next iteration, or its tasks will be reallocated to other nodes.

[0226] Fault tolerance mechanism: When a certain computing node fails, the system can automatically reallocate its tasks to other nodes to ensure that the algorithm continues. For example, if node 3 suddenly disconnects during the processing, the 4-agent sub-problems it is responsible for will be automatically transferred to other available nodes for processing.

[0227] Asynchronous communication: Adopt an asynchronous communication mechanism, and different nodes can process sub-problems at different speeds without strict synchronization in each iteration. This further improves the flexibility and efficiency of the system. For example, after a computing node completes solving a sub-problem, it can immediately return the result to the master node without waiting for other nodes.

[0228] Sparse communication: Optimize the communication content between nodes, only transmit the necessary variable updates, and reduce the communication overhead. For example, if the allocation scheme of a certain agent has not changed in this iteration, there is no need to transmit its complete scheme to the master node, and only the "no change" flag needs to be sent.

[0229] Through this distributed computing framework, the system can complete the optimization of large-scale task allocation within seconds to minutes (depending on the problem scale and complexity), meeting the real-time processing requirements. For example, for the above allocation problem of 100 cases, the traditional centralized algorithm may take 30 minutes to complete, while using this distributed framework, with a configuration of 5 computing nodes, the optimization process can be completed within 2 minutes, significantly improving the system's processing ability for sudden or unconventional cases.

[0230] In practical applications, this distributed task allocation method based on inexact ADMM has been successfully applied to multiple scenarios.

[0231] For example, in the handling of a large number of related cases triggered by a certain business activity, the system completed the intelligent allocation of 78 cases within 5 minutes, optimized the matching according to case types, complexity, and agent expertise, and the processing efficiency was increased by about 60% compared with traditional manual allocation. At the same time, it ensured the load balance of each agent and avoided resource bottlenecks. Another important advantage of this method is its strong scalability. When the number of cases or agents increases, only computing nodes need to be added to maintain high computing performance, which provides strong support for the system to handle unplanned tasks of various scales.

[0232] Through the above steps, the system successfully constructed a professional multi-agent system, realized the intelligent evaluation and analysis of case quality, and developed a distributed task allocation mechanism based on the inexact alternating direction method of multipliers, improving the system's ability to handle sudden or unconventional cases. These innovative technologies effectively solved the deficiencies of traditional methods in dealing with complex and changeable cases and sudden tasks, providing strong technical support for improving work efficiency and quality.

[0233] Step S5 specifically includes: S5.1: Based on the distributed task allocation results, through technologies such as text similarity analysis, entity recognition, and relationship extraction, construct a case relevance recognition model to realize the automatic recognition of case groups with internal connections. Step S5.1 is to construct a case relevance recognition model based on the distributed task allocation results through technologies such as text similarity analysis, entity recognition, and relationship extraction.

[0234] In actual work, there are often cases where there are associated relationships between multiple cases. Collaboratively handling these associated cases is crucial for improving work efficiency and consistency. This model can automatically identify case groups with internal connections, providing a basis for subsequent collaborative processing.

[0235] For example, the system can identify the following association patterns: entity coincidence association (such as multiple cases of the same party), factual association (such as multiple cases triggered by one event), and rule relationship association (such as multiple cases triggered by a business dispute). Specifically, the system uses a text similarity algorithm to calculate the similarity between cases, extracts key entities (such as parties, locations, events) in cases through named entity recognition technology, and analyzes the relationships between entities through relationship extraction technology to discover potential case associations. This automatic association recognition greatly reduces the workload of manual screening and improves the efficiency and accuracy of discovering associated cases.

[0236] S5.2: Based on the case groups with internal connections, design a hierarchical task planning framework that includes goal decomposition, constraint identification, and action sequence generation. Through symbolic planning and heuristic search, generate a preliminary action plan for handling related cases. Step S5.2 is to design a hierarchical task planning framework based on the identified case groups with internal connections. This framework generates a preliminary action plan for handling related cases through goal decomposition, constraint identification, and action sequence generation.

[0237] For example, for a group of interrelated commercial dispute cases, the system can design a hierarchical processing plan: the first layer determines the case processing order and dependencies, the second layer plans the specific processing steps for each case, and the third layer arranges resource allocation and time scheduling.

[0238] In implementation, the system uses symbolic planning and heuristic search methods. Symbolic planning represents case handling as a series of states and actions, and constructs a state transition model for case handling by defining preconditions and effects. Heuristic search uses domain knowledge to design a heuristic function to guide the search process and find a high-quality preliminary action plan. For example, the system may plan to handle core cases first and then dependent cases to ensure the efficiency and consistency of overall handling. This hierarchical planning method makes the complex problem of handling related cases structured and manageable, greatly improving the handling efficiency.

[0239] The state transition model of case handling is a computational model that formally represents the case handling process and is constructed using a state-action-transition framework. This model is represented using a five-tuple <S, A, T, I, G>, where S is the state space, A is the action set, T is the transition function, I is the initial state set, and G is the goal state set. The state space S is designed as a multi-dimensional feature vector, including key dimensions such as the case handling stage, completed processing steps, resource allocation status, and document generation status. For example, the state of a commercial case can be represented as [stage = fact-finding, completed steps = {evidence collection, evidence exchange}, resource allocation = {Judge A, Clerk B}, document status = {generated: notice of case-filing, service receipt}]. The design of the state space fully considers the domain characteristics of case handling to ensure the completeness and compactness of state representation.

[0240] The action set A includes all the operations that can be executed during the case handling process, which are functionally divided into procedural actions (such as case acceptance, scheduling a certain procedure, adjourning a hearing), substantive actions (such as evidence investigation, fact determination), and management actions (such as resource allocation, priority adjustment, related case handling). Each action is represented in a structured form, including an action name, a parameter list, preconditions, and effects. For example, the "scheduling a certain procedure" action can be represented as: Name = Scheduling a certain procedure, Parameters = {Date, Location, Participants}, Preconditions = {Case acceptance completed = Yes, Service completed = Yes, Limitation period for evidence submission expired = Yes}, Effects = {Stage = Trial preparation, Date of a certain procedure = Set value, Trial plan = Generated}.

[0241] The transition function T defines the evolution rules between states, adopting a mapping relationship of T: S×A→S. Given the current state s and the executed action a, the transition function checks whether the preconditions of a are satisfied in s. If satisfied, it applies the effects of a to generate a new state s'. The implementation of the transition function uses a rule-based inference engine, supporting operations such as condition judgment, value assignment, and list update. To handle the uncertainty in case handling, the model supports probabilistic transitions, using T: S×A×S→[0,1] to represent the probability of transitioning from state s by executing action a to state s', which is applicable to situations where the processing result is uncertain, such as the success probability of a mediation attempt.

[0242] The initial state set I defines the starting state of case handling, usually corresponding to the case acceptance stage. The system automatically generates an appropriate initial state representation based on the case type and characteristics. The target state set G defines the final states of case handling, usually including various possible ways of closing a case, such as closing a case by judgment, closing a case by mediation, withdrawing a lawsuit to close a case, etc. The definition of the target state considers both procedural completion indicators and quality compliance indicators.

[0243] The construction of the state transition model adopts a three-stage method: pattern extraction, rule formalization, and model verification. In the pattern extraction stage, the system analyzes a large number of historical case handling records and uses process mining techniques to identify common handling paths and decision points. In the rule formalization stage, combined with legal expert knowledge, the extracted handling patterns are transformed into structured states, actions, and transition rules. In the model verification stage, historical case data is used to verify the completeness and correctness of the model to ensure that the model can accurately represent the handling processes of various types of cases.

[0244] To enhance the practicality of the model, the system has implemented a state visualization tool that transforms the abstract state representation into an intuitive flowchart to help users understand the current state and possible paths of case handling. In addition, the model supports dynamic updates and can automatically adjust the state definitions and transition rules according to changes in regulations and the evolution of handling practices to ensure that the model always reflects the latest case handling norms and best practices.

[0245] S5.3: Based on the preliminary action plan, develop a reinforcement learning optimization model that includes state representation, reward mechanism, and policy network. Through interaction with the environment and policy iteration, optimize the handling strategies of related cases to improve the handling effect. Step S5.3 is to develop a reinforcement learning optimization model based on the preliminary action plan. This model optimizes the handling strategies of related cases through state representation, reward mechanism, and policy network.

[0246] Specifically, the system takes the case handling status as the environment. The actions of the agent include case analysis, evaluation, and recommendation generation. The reward signal comes from the handling efficiency and consistency metrics. Through interaction with the environment and policy iteration, the system continuously adjusts and optimizes the handling strategies to improve the handling effect.

[0247] In implementation, the system uses deep reinforcement learning methods, such as deep Q-learning or policy gradient methods. The system designs a suitable state representation to capture the key features of case handling; designs a multi-objective reward function to balance handling efficiency, quality, and consistency; and uses a deep neural network as the policy network to learn the mapping from states to optimal actions. For example, the system may learn that in related cases, handling various cases with the most sufficient evidence first can provide a reference standard for other cases; or in event-related cases, liability determination cases should be given priority. This data-driven optimization method can learn from actual handling experience, continuously improve handling strategies, and adapt to complex and changing case situations.

[0248] The following details the construction and training of this model: Model architecture design: The reinforcement learning optimization model adopts a deep reinforcement learning architecture based on Advantage Actor-Critic (A2C), which includes two core components: a policy network (Actor) and a value network (Critic).

[0249] Structure of the policy network (Actor): Input layer: Receives a state vector with a dimension of 128; First fully connected layer: 128 → 256, using ReLU activation; Second fully connected layer: 256 → 128, using ReLU activation; Third fully connected layer: 128 → 64, using ReLU activation; Output layer: 64 → action space dimension, using Softmax activation to output the probability distribution of each action; Structure of the value network (Critic): Input layer: Receives a state vector with a dimension of 128; The first fully connected layer: 128 → 256, using ReLU activation; The second fully connected layer: 256 → 128, using ReLU activation; Output layer: 128 → 1, without an activation function, directly outputting the state value estimate; State representation design: The state vector (128 - dimensional) contains the following parts of information: Case features (64 - dimensional): including case type (one - hot encoded, 8 - dimensional), complexity score (scalar), urgency level (scalar), processing stage (one - hot encoded, 6 - dimensional), associated strength matrix (compressed representation, 48 - dimensional); Resource status (16 - dimensional): including the current load of each processor (vector), expertise matching degree (vector); Time status (16 - dimensional): including the processed time of each case, remaining time window, global time progress; Historical actions (32 - dimensional): including the encoded representation of the recent action history and its results; Action space design: The action space defines the operations that the agent can take, including: Case priority assignment: Assigning a processing priority (5 priority levels) to each case; Resource allocation: Assigning processors to specific cases; Processing strategy selection: Selecting a processing strategy for the case (such as "in - depth investigation", "quick processing", etc.); Associated case coordination: Adjusting the information sharing and processing dependency relationships between associated cases; Reward mechanism design: The reward function comprehensively considers multiple factors, The formula is: R = w1*R efficiency + w2*R quality + w3*R consistency - w4*R penalty Where: R efficiency : Efficiency reward, based on the ratio of the completion time to the benchmark time, the formula is: max(0, 1 - actual time / benchmark time) * 10 R quality : Quality reward, based on the case processing quality score, ranging from 0 - 10; R consistency : Consistency reward, based on the consistency measure of the processing results of associated cases, ranging from 0 - 5; R penalty : Violation penalty, for behaviors that violate processing rules or constraints, such as over - allocation of resources, incorrect priorities, etc.; Weight setting: w1 = 0.4, w2 = 0.3, w3 = 0.3, w4 = 1.0, obtained through hyperparameter search and optimization; Implementation of the training algorithm: Use the Proximal Policy Optimization (PPO) algorithm for training, which is more stable than the standard A2C. Key parameter settings: Discount factor (γ): 0.99, controlling the importance of future rewards; GAE parameter (λ): 0.95, used for advantage estimation; Coefficient of the value function: 0.5, controlling the proportion of the value loss in the total loss; Entropy coefficient: 0.01, encouraging exploration; PPO clipping parameter (ε): 0.2, restricting the magnitude of policy updates; Learning rate: 3e-4, using the Adam optimizer; Maximum gradient norm: 0.5, preventing gradient explosion; Design of the training process: The training is carried out in the following steps: Initialization: Use the pre-trained action plan as the initial policy to accelerate the training convergence; Experience collection: Collect the experience of 2048 environmental steps in each round; Batch update: Divide the collected experience into 64 small batches and perform 10 rounds of parameter updates; Policy evaluation: After every 10 rounds of training, evaluate the current policy in the validation environment; Early stopping mechanism: If there is no improvement in 20 consecutive rounds of evaluation, stop the training; Model saving: Save the model parameters with the best performance; Experience replay and sample reuse: Maintain an experience replay buffer with a size of 10,000; Adopt the Prioritized Experience Replay mechanism and allocate the sampling probability according to the TD error; The importance sampling weight parameter β increases linearly from 0.4 to 1.0 to compensate for the sampling bias; Design of the exploration strategy: Use the ε-greedy strategy in the first 1000 rounds, and ε decays linearly from 0.5 to 0.1; Use entropy-based exploration afterwards, and adaptively control the exploration degree by adjusting the entropy coefficient; Increase the exploration probability (curiosity-driven exploration) for complex or rare case types; Implementation of the environment simulator: To implement the reinforcement learning training, a case processing environment simulator is developed: Case Generation: Generate simulated cases based on historical data distribution, including attributes such as type and complexity; Association Relationship Generation: Generate an association relationship network among cases according to the actual association pattern; Processing Time Model: Simulate the time consumption under different types of cases and different processing strategies; Processing Quality Model: Simulate the impact of different processing methods on the final quality; Consistency Evaluation: Calculate the consistency index of the processing results of associated cases; Model Evaluation and Tuning: Performance Metrics: Average completion time, average quality score, consistency score, resource utilization rate; Offline Evaluation: Conduct retrospective testing on historical case data; Comparison Benchmark: Compare with rule-based methods and supervised learning methods; Sensitivity Analysis: Analyze the sensitivity of the model to changes in each hyperparameter; Abnormal Situation Testing: Test the performance of the model under extreme situations (such as resource shortage and sudden high-priority cases); Implementation Technology Stack: Framework: Implement neural networks using PyTorch, and RLlib provides support for reinforcement learning algorithms; Environment Simulation: Use a custom Python environment, following the interface specifications of OpenAI Gym; Distributed Training: Use the Ray framework to achieve parallel data collection and gradient update; Training Hardware: 8 NVIDIA V100 GPUs, and the training time is about 48 hours; Deployment and Online Learning: Model deployment adopts a microservices architecture and provides decision-making services through RESTful APIs; Implement a lightweight online learning mechanism to continuously adjust strategies according to actual feedback; Set a trust threshold. When the model confidence is lower than the threshold, transfer it to manual intervention; Maintain a case library, record typical decision-making scenarios and optimal decisions for continuous improvement.

[0250] S5.4: Based on the reinforcement learning optimization model and the hierarchical task planning framework, implement a dynamic fusion mechanism of planning and learning, through a two-way feedback mechanism of using planning to guide exploration and learning to optimize planning, improve the adaptability of strategies while ensuring the rationality of planning, and achieve intelligent processing of complex associated cases.

[0251] Step S5.4 is to implement a dynamic integration mechanism for planning and learning. Through the two-way feedback of using planning to guide exploration and learning to optimize planning, while ensuring the rationality of planning, it improves the adaptability of the strategy. For example, the initial plan provides the overall framework and constraints for handling cases. Reinforcement learning explores and optimizes specific strategies within this framework, and at the same time, the learning results can in turn adjust and improve the plan.

[0252] In terms of specific implementation, the system adopts a hybrid architecture, which organically combines symbolic planning and reinforcement learning. The planning module provides high-level structured knowledge to guide the exploration direction of reinforcement learning; the reinforcement learning module discovers better execution strategies through actual interaction experience and feeds back the learning results to the planning module to optimize subsequent plans.

[0253] For example, the system may discover through reinforcement learning that when handling related commercial dispute cases, it is more efficient to first handle the core case with the largest amount and then handle its derivative cases. This discovery can be integrated into the planning knowledge to guide the handling of future similar cases. This dynamic integration mechanism of planning and learning combines the advantages of rule-based knowledge and data-driven learning, ensuring both the standardization of handling and the adaptability of the strategy, and is particularly suitable for handling complex related cases.

[0254] Step S6 specifically includes: S6.1: Based on the multi-agent case association processing results, through technologies such as entity extraction, relationship recognition, and knowledge fusion, construct a case knowledge graph containing case elements, legal concepts, and processing procedures, providing a data basis for graph learning. Step S6.1 is to construct a case knowledge graph based on the multi-agent case association processing results. This graph contains information such as case elements, rule concepts, and processing procedures, providing a data basis for graph learning.

[0255] Case data usually exhibits complex graph structure characteristics, containing a large number of entities and relationships. Effectively learning this graph structure is crucial for case quality assessment.

[0256] In terms of specific implementation, the system identifies key entities from case materials through entity extraction technology, discovers the associations between entities through relationship recognition technology, and integrates information from different sources into a unified knowledge graph through knowledge fusion technology. For example, the system can extract entities such as parties, various bases, and applicable regulations from case documents, identify citation relationships, support relationships, etc. between them, and construct a knowledge graph representing the case structure. This structured knowledge representation lays the foundation for subsequent graph learning.

[0257] S6.2: Based on the case knowledge graph, design a graph neural network model that includes a graph convolutional layer, an attention mechanism, and a message passing mechanism. Through node feature extraction and edge relationship learning, achieve a preliminary modeling of the case graph structure; Step S6.2 is to design a graph neural network model based on the case knowledge graph. This model includes a graph convolutional layer, an attention mechanism, and a message passing mechanism. Through node feature extraction and edge relationship learning, achieve a preliminary modeling of the case graph structure.

[0258] The typical architecture of a graph neural network includes: an input layer (initial node features), a graph convolutional layer (aggregating adjacent node information), an attention mechanism (focusing on important nodes and edges), and an output layer (node representation or graph representation).

[0259] In implementation, the system adopts advanced graph neural network architectures such as Graph Convolutional Network (GCN) or Graph Attention Network (GAT). For example, through GCN, the system can aggregate the neighbor information of nodes and learn the representation of nodes; through GAT, the system can focus on important nodes and edges to improve the learning efficiency. These models can capture the complex patterns and structural features in the case graph and provide effective representations for subsequent tasks. Compared with traditional vector representations, graph neural networks can better capture the relationships and structural information between entities and are particularly suitable for processing data types with rich relationships such as case data.

[0260] The graph neural network model is used to learn the structural features of the case knowledge graph. The detailed construction and training process of this model are as follows: Model architecture: Adopt a hierarchical architecture based on Graph Attention Network (GAT), including the following layers: Input layer: Receive the initial feature vector of nodes, with a dimension of 128; The first graph attention layer: 8 attention heads, each head outputs 32-dimensional features, and after merging, it is 256-dimensional; Batch normalization layer: Standardize the feature distribution to stabilize the training process; The second graph attention layer: 8 attention heads, each head outputs 32-dimensional features, and after merging, it is 256-dimensional; Global pooling layer: Use graph pooling with the attention mechanism to compress the graph-level features into a 384-dimensional vector; Fully connected layer: Two layers, 384 -> 192 and 192 -> 64 respectively, using ReLU activation; Output layer: Designed according to specific tasks. For example, in the case classification task, use Softmax to output class probabilities Feature design: Node features include two parts: Text semantic features: Extracted using a pre-trained BERT model, with a dimension of 768, and then reduced to 64 dimensions through linear mapping; Structural features: Include node type (one-hot encoding, dimension depending on the number of node types), node degree, centrality measure, etc., which are merged to 64 dimensions; The final node features are obtained by concatenating the two parts of features to get a 128-dimensional vector; Attention mechanism: Adopt a multi-head self-attention mechanism, and the calculation formula is: e ij = LeakyReLU(aT [Wh i || Wh j ) α ij = softmax j (e ij ) h' i = σ(∑ j α ij W h j ) Among them, W is the weight matrix (128×32), a is the attention vector (64×1), and the negative slope of LeakyReLU is 0.2.

[0261] Forced zero method sparsification mechanism: Achieved by applying L0 regularization to each attention head, and the specific method is as follows: Introduce a gating variable z ij for each edge, parameterized by the Hard Concrete distribution to make it differentiable during training; The parameter θ of the gating variable ij is learned by the model and initialized to 0; During forward propagation, the edge weight is calculated by w ij × z ij , where z ij is 0 or 1; Add an L0 regularization term to the loss function: λ∑ ij P(z ij = 1), where λ is the regularization strength, set to 0.001; Training configuration: Optimizer: Adam, learning rate is 0.001, β1 = 0.9, β2 = 0.999; Learning rate scheduling: Use ReduceLROnPlateau, factor is 0.5, patience value is 10; Batch size: Adjusted according to GPU memory, typical value is 16 - 32; Number of training epochs: 200 epochs, use early stopping strategy, patience value is 30; Loss function: The main task loss (selected according to the specific task, such as cross-entropy for classification tasks) plus the L0 regularization term; Gradient clipping: Set the maximum gradient norm to 1.0 to prevent gradient explosion; Adaptive learning strategy: Meta-learning: Adopt the Model-Agnostic Meta-Learning (MAML) algorithm, with the inner-loop learning rate of 0.01, the outer-loop learning rate of 0.001, and the meta-batch size of 5; Knowledge distillation: The teacher model is the complete GAT, the student model is the sparse GAT, the temperature parameter is 2.0, and the distillation weight is 0.5; Incremental training: Use the Elastic Weight Consolidation (EWC) method, set the importance coefficient to 1000, and retain the original knowledge when fine-tuning on new data; Evaluation metrics: Accuracy (for classification tasks) or mean squared error (for regression tasks); Model sparsity (percentage of retained edges); Inference time (milliseconds per sample); Explanatory score (rated by experts); Implementation technology: Implemented using the PyTorch and PyTorch Geometric libraries, training is carried out on an NVIDIA Tesla V100 GPU, and batch inference can also be efficiently executed on the CPU.

[0262] S6.3: Based on the graph neural network model, implement a forced-zero method sparsification mechanism that includes L0 regularization, gated activation, and gradient estimation, by explicitly controlling the model complexity and the number of activated nodes. Step S6.3 is to implement a forced-zero method sparsification mechanism based on the graph neural network model.

[0263] This mechanism includes L0 regularization, gated activation, and gradient estimation. By explicitly controlling the model complexity and the number of activated nodes, it reduces the redundancy of the graph representation and improves the learning efficiency. Different from traditional L1 regularization, the forced-zero method sparsification directly forces unimportant connections to be exactly zero through L0 regularization, thus significantly reducing the model complexity.

[0264] Specifically, the system adds a binary gating variable for each edge to determine whether to retain the edge; realizes the differentiable optimization of L0 regularization through hard gating approximation techniques (such as Concrete Distribution); automatically learns the optimal sparse structure during training, and typically most edges can be pruned while maintaining performance.

[0265] For example, when the system analyzes a complex commercial dispute case, the original case graph may contain more than 300 entities and more than 2,000 relationships. After applying the forced zero method for sparsification, only about 400 key relationships may be retained, but these relationships accurately capture the core rule relationships of the case. This sparsification process not only improves the computational efficiency but also enhances the interpretability of the model, enabling the system to focus on the key elements and relationships of the case.

[0266] S6.4: Develop an adaptive learning algorithm based on the forced zero method sparsification mechanism, which includes meta-learning, knowledge distillation, and incremental training. By dynamically adjusting the learning strategy and model structure, it adapts to the graphical characteristics of different types of cases and achieves effective learning based on the case graph.

[0267] Step S6.4 is to develop an adaptive learning algorithm based on the forced zero method sparsification mechanism. This algorithm includes meta-learning, knowledge distillation, and incremental training, and adapts to the graphical characteristics of different types of cases by dynamically adjusting the learning strategy and model structure.

[0268] For example, when dealing with new types of environmental protection-related cases, the system can quickly adapt to the graph structure characteristics of the new case type through meta-learning, extract key knowledge from complex models using knowledge distillation, and continuously optimize the model through incremental training.

[0269] In implementation, the system uses the meta-learning method to learn the common characteristics of different types of cases, accelerating the adaptation process for new case types; uses knowledge distillation technology to transfer the knowledge of complex models to simpler and more efficient models; and adopts an incremental training strategy to efficiently update the model when new case data arrives, avoiding the overhead of retraining. These technologies together improve the learning efficiency and adaptability of the system, and are particularly suitable for dealing with diverse case types and changing regulatory requirements. Through adaptive learning, the system can continuously accumulate and utilize processing experience, improving the generalization ability and performance of the model.

[0270] Through the above steps, the system has successfully achieved the intelligent processing of complex associated cases and effective learning based on the case graph, improving the efficiency and consistency of handling associated cases and enhancing the knowledge representation and reasoning ability of the system.

[0271] These innovative technologies effectively solve the difficulties of traditional methods in dealing with complex associated cases and learning the case graph structure, providing strong technical support for improving the quality and efficiency of case handling. In particular, the combination of the method of combining planning and reinforcement learning and the forced zero method sparsification graph technology represents the forefront research direction in the field of case handling and analysis, with important theoretical significance and practical value.

[0272] Step S7 specifically includes: S7.1: Based on the effective learning results of the case graph, through data cleaning, format conversion, and redundancy elimination techniques, integrate multi-source evaluation data including evaluation criteria, case characteristics, evaluation results, and improvement suggestions to form a unified dataset. Step S7.1 is to integrate multi-source evaluation data based on the effective learning results of the case graph.

[0273] This data contains information such as evaluation criteria, case characteristics, evaluation results, and improvement suggestions, and through techniques such as data cleaning, format conversion, and redundancy elimination, a unified dataset is formed.

[0274] For example, the system can integrate the results from different evaluation agents, including case retrospective analysis, quality scoring, improvement suggestions, etc., convert them into a unified data format, eliminate redundancy and inconsistencies, and form a high-quality data foundation. In actual implementation, the system first standardizes various data sources, including unifying data formats, normalizing field names, and checking the consistency of value ranges. For example, the date formats from different sources may have differences such as "2023 / 03 / 25", "2023-03-25", or "20230325", and the system will unify them into the ISO standard format.

[0275] Secondly, the system conducts quality checks and cleaning on the data to identify and process missing values, outliers, and duplicate records. For example, for outliers in the scoring data, the system may adopt an outlier detection method based on the interquartile range or a rationality check based on domain rules.

[0276] Finally, the system performs data integration and redundancy elimination, merges data from different sources that describe the same entity or event, and constructs a complete and non-redundant unified dataset to provide a high-quality data foundation for subsequent knowledge graph construction.

[0277] S7.2: Based on the unified dataset, design a knowledge graph schema including concept hierarchy, relationship types, and attribute definitions, and through ontology engineering and semantic modeling, construct a conceptual framework for the case quality evaluation domain to guide the construction of the knowledge graph.

[0278] Step S7.2 is to design a knowledge graph schema based on the unified dataset.

[0279] This schema includes concept hierarchy, relationship types, and attribute definitions, and through ontology engineering and semantic modeling, constructs a conceptual framework for the case quality evaluation domain to guide the construction of the knowledge graph.

[0280] In schema design, the system first defines core concept classes, such as "case", "evaluation", "evaluation dimension", "improvement suggestion", etc. Each concept class has a clearly defined set of attributes. For example, the "case" concept may include attributes such as "number", "type", "processing date", etc., and the "evaluation" concept may include attributes such as "evaluation time", "evaluator", "total score", etc. Secondly, the system defines the types of relationships between concepts, such as "case - belongs to - case type", "evaluation - evaluates - case", "evaluation - contains - evaluation dimension", "evaluation dimension - generates - improvement suggestion", etc.

[0281] Each type of relationship has a clear definition, directionality, and constraints. For example, the "evaluation - evaluates - case" relationship means that the evaluation object must be a case, and a case can have multiple evaluations, while one evaluation can only evaluate one case.

[0282] Finally, the system establishes a hierarchical system of concepts to form a concept classification tree. For example, "evaluation dimension" can be subdivided into sub - concepts such as "procedural standardization", "appropriateness of entity handling", "document standardization", etc., and these sub - concepts can be further subdivided. Through this hierarchical design, the system constructs a structured concept framework, providing a theoretical basis for subsequent knowledge representation and reasoning.

[0283] S7.3: Based on the knowledge graph schema and the unified data set, through techniques such as named entity recognition, relation extraction, and event detection, convert unstructured and semi - structured data into knowledge triples to construct an initial case quality assessment knowledge graph. Step S7.3 is to realize knowledge extraction and graph construction based on the knowledge graph schema and the unified data set. In this process, through techniques such as named entity recognition, relation extraction, and event detection, convert unstructured and semi - structured data into knowledge triples to construct an initial case quality assessment knowledge graph.

[0284] In the knowledge extraction process, the system first uses named entity recognition technology to identify key entities from text data. For example, identify case numbers, evaluation dimension names, score values, etc. from evaluation reports.

[0285] The system adopts a sequence annotation model based on BiLSTM - CRF (Bidirectional Long Short - Term Memory Network - Conditional Random Field), combined with domain - specific dictionaries and rules, to achieve high - precision entity recognition. Secondly, the system uses relation extraction technology to identify the semantic relationships between entities.

[0286] For example, identify the relationship "Case A scores 85 points in the procedural standardization dimension". The system adopts a method based on dependency syntactic analysis and graph pattern matching, combined with distant supervision learning technology, to automatically extract structured relationships from the text.

[0287] Finally, the system organizes the extracted entities and relationships into knowledge triples according to the predefined knowledge graph schema, such as (Case A, Scoring_Procedural_Normativity, 85 points), (Case A, Existing_Problems, Irregular_Service_Procedure), etc., and imports these triples into the graph database to construct a knowledge graph. In this way, the system can transform a large amount of unstructured and semi-structured data into a structured knowledge representation, laying a foundation for subsequent knowledge reasoning and applications.

[0288] S7.4: Based on the initial case quality assessment knowledge graph, develop a knowledge reasoning engine that includes rule reasoning, path reasoning, and statistical reasoning, and design a standardized application interface to support multi-dimensional queries, intelligent analysis, and precise recommendations for case quality management, realizing the intelligence and precision of case quality management and providing data support for legal decision-making.

[0289] Step S7.4 is to develop knowledge reasoning and application interfaces based on the initial knowledge graph. This part includes a knowledge reasoning engine for rule reasoning, path reasoning, and statistical reasoning, as well as a standardized application interface to support multi-dimensional queries, intelligent analysis, and precise recommendations for case quality management.

[0290] In terms of knowledge reasoning, the system first implements a rule-based reasoning engine to support deductive reasoning based on ontology axioms and rule bases. For example, if the rule "If the procedural normativity score of a case is lower than 70 points, then there is a risk of procedural problems in the case" is defined, the system can automatically infer which cases have a risk of procedural problems.

[0291] Secondly, the system implements path-based reasoning to discover implicit relationships by analyzing path patterns in the knowledge graph. For example, the system can discover the association between Handler A and the handling of high-quality cases by analyzing the path patterns of "Handler A - Handles - Case Set B" and "Case Set B - Scoring_Higher_Than - 90 points".

[0292] Finally, the system implements statistical-based reasoning to discover data patterns and trends through aggregation analysis. For example, by analyzing the scoring distribution of different types of cases, it can be found which types of cases have more quality problems.

[0293] In terms of the application interface, the system develops a standardized RESTful API to support flexible queries according to conditions such as case type, time period, and scoring range; supports time trend and distribution analysis of case quality; supports functions such as recommending improvement suggestions based on similar cases. These interfaces provide rich data services for upper-layer applications, realizing the intelligence and precision of case quality management.

[0294] In summary, the intelligent case quality assessment method based on the large language model provided by this application constructs a dynamically updated assessment standard library, designs a complete assessment inference chain, constructs a professional multi-agent system, applies the inexact alternating direction multiplier method to achieve distributed task allocation, combines planning and reinforcement learning to solve the case association problem, applies the forced zero method to sparse the graph technology to achieve case graph learning, constructs a case quality assessment knowledge graph, realizes the intelligence and precision of case quality management, effectively solves the problems of untimely update of assessment standards, incomplete assessment logic chain, and low system intelligence level in the prior art, improves the accuracy, comprehensiveness and efficiency of case quality assessment, and provides strong technical support for legal affairs work.

[0295] The embodiment of this application also provides an intelligent case quality assessment system based on the large language model. The computer system includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the above-mentioned intelligent case quality assessment method based on the large language model. The embodiment of this application also provides a computer-readable storage medium that stores computer instructions for causing a computer to execute the above-mentioned intelligent case quality assessment method based on the large language model. The embodiment of this application also provides a computer program product, including computer instructions that implement the steps of the above-mentioned intelligent case quality assessment method based on the large language model when executed by a processor. The above are only the preferred embodiments of this application, and do not limit the patent scope of this application. Any equivalent structural transformation made by using the content of the specification and drawings of this application under the inventive concept of this application, or direct / indirect application in other related technical fields, shall be equally included in the patent protection scope of this application.

Claims

1. An intelligent case quality assessment method based on large language models, characterized in that, Including: Based on legal regulatory documents and trial practices, construct an evaluation standard library covering various types of cases; Based on the evaluation standard library, design a complete evaluation reasoning chain, including evaluation of procedural legality, evaluation of propriety of entity handling, and evaluation of legal document standardization; Based on the evaluation reasoning chain, construct a multi-agent system including case backtracking agent, quality scoring agent, and improvement suggestion agent to realize intelligent evaluation and analysis of case quality; Based on the multi-agent system, apply the inexact alternating direction method of multipliers algorithm to achieve distributed task allocation for unplanned online tasks, and improve the system's handling ability for sudden or unconventional cases; Based on the distributed task allocation results, combine planning and reinforcement learning techniques to solve the problem of multi-agent case association, and realize intelligent processing of complex associated cases; Based on the multi-agent case association processing results, apply the forced zero method sparsification graph technology to achieve effective learning based on case graphs; Based on the effective learning results of the case graph, construct a case quality evaluation knowledge graph to realize the intelligence and precision of case quality management and provide data support for legal decision-making.

2. The method according to claim 1, wherein The construction of an evaluation standard library covering various types of cases based on legal regulatory documents and trial practices includes: Based on legal regulatory documents and trial practices, extract the regulatory requirements and standards related to case quality evaluation through text extraction and semantic analysis technologies, and form a structured dataset of regulatory requirements; Based on the dataset of regulatory requirements, classify according to case types and evaluation dimensions, construct a hierarchical evaluation standard model, and generate an initial evaluation standard library; Based on the initial evaluation standard library, design a dynamic update mechanism including detection of new regulatory documents, identification of semantic changes, and automatic update of standard items to realize the real-time update ability of the evaluation standard library; Based on the dynamic update mechanism, develop a standard conflict detection algorithm and coordination strategy, and solve the possible conflict problems in the standard update process through semantic similarity analysis and rule priority determination to realize the real-time update and optimization of the evaluation standard library.

3. The method according to claim 2, wherein The extraction of regulatory requirements and standards related to case quality evaluation through text extraction and semantic analysis technologies based on legal regulatory documents and trial practices includes: Based on legal regulatory documents and trial practices, collect regulatory documents to form an original document set; Based on the original document set, preprocess the documents through natural language processing technologies, including text cleaning, word segmentation, part-of-speech tagging, and syntactic analysis, to generate structured text; Based on the structured text, apply named entity recognition and relation extraction technologies to identify the evaluation elements and standard items in the text, and extract the semantic relationships between them to construct a preliminary network of regulatory elements; Based on the network of regulatory elements, merge and standardize duplicate or similar evaluation elements through semantic similarity calculation and clustering analysis to form a structured dataset of regulatory requirements.

4. The method according to claim 1, characterized in that The design of a complete evaluation reasoning chain based on the evaluation standard library includes: Based on the procedural norms in the evaluation standard library and combined with the reasoning ability of the multi-agent system, design a sub-chain for evaluating the procedural legality of a case, which includes the case-filing procedure, service procedure, evidence presentation procedure, and trial procedure links, to achieve a comprehensive evaluation of the procedural legality of the case; Based on the substantive norms in the evaluation standard library and the evaluation results of the procedural legality, design a sub-chain for evaluating the appropriateness of the substantive handling of a case, which includes fact-finding, the acceptance of various types of evidence, legal application, and judgment reasoning links, to achieve an accurate evaluation of the appropriateness of the substantive handling of the case; Based on the document norms in the evaluation standard library and the evaluation results of the appropriateness of the substantive handling, design a sub-chain for evaluating the standardization of legal documents in terms of document format, language expression, logical structure, and content integrity, to achieve a detailed evaluation of the standardization of legal documents; Based on the sub-chain for evaluating the procedural legality, the sub-chain for evaluating the appropriateness of the substantive handling, and the sub-chain for evaluating the standardization of legal documents, through the multi-task learning and chain reasoning technologies of the large language model, integrate each evaluation sub-chain, and develop an inference control mechanism that includes evaluation sequence control, intermediate result feedback, and evaluation depth adjustment, to form a complete inference chain for evaluating the case quality.

5. The method according to claim 1, characterized in that, Based on the evaluation inference chain, construct a multi-agent system that includes a case backtracking agent, a quality scoring agent, and an improvement suggestion agent, to achieve the intelligent evaluation and analysis of the case quality, including: Based on the evaluation inference chain, develop a case backtracking agent with the ability to analyze the entire process of a case. Through time series analysis and causal reasoning, achieve the backtracking analysis of the entire process of case handling, and identify the key nodes and existing problems in case handling; Based on the case backtracking analysis results and the evaluation inference chain, develop a quality scoring agent. Through multi-dimensional evaluation and weight adaptive algorithms, achieve the quantitative scoring of the case quality, and generate an evaluation result that includes the overall score and sub-dimensional scores; Based on the quality scoring results and the evaluation standard library, develop an improvement suggestion agent. Through difference analysis and best practice matching, generate specific and operable improvement suggestions for the dimensions with scores lower than the first preset threshold, to provide guidance for improving the case quality; Based on the case backtracking agent, the quality scoring agent, and the improvement suggestion agent, design an intelligent agent collaboration framework that includes task allocation, information sharing, and result fusion. Optimize the collaboration strategy through reinforcement learning to achieve the efficient collaborative work among agents, form a complete multi-agent system, and achieve the intelligent evaluation and analysis of the case quality.

6. The method according to claim 1, characterized in that, Based on the multi-agent system, apply the inexact alternating direction method of multipliers algorithm to achieve distributed task allocation for unplanned online tasks, and improve the system's ability to handle sudden or unconventional cases, including: Based on the multi-agent system, extract features and perform mathematical modeling on unplanned online tasks, and construct a task allocation optimization model that includes task features, resource constraints, and time windows; Based on the task allocation optimization model, design an inexact alternating direction method of multipliers algorithm framework. Through steps of problem decomposition, variable update, and multiplier adjustment, transform the global optimization problem into multiple sub-problems that can be solved in parallel; Based on the inexact alternating direction method of multipliers algorithm framework, develop an inexact update strategy that includes adaptive precision control, early stopping strategy, and approximate solution. By reducing the computational precision requirements for each iteration, the algorithm convergence speed is accelerated. Based on the inexact update strategy, develop a distributed computing framework that supports multi-node parallel computing. Through task partitioning, node coordination, and result aggregation, efficient distributed processing of unplanned online tasks is achieved.

7. The method according to claim 1, wherein Based on the distributed task allocation results, combine planning and reinforcement learning techniques to solve the problem of multi-agent case association and achieve intelligent processing of complex associated cases, including: Based on the distributed task allocation results, construct a case relevance recognition model through techniques such as text similarity analysis, entity recognition, and relationship extraction to achieve automatic recognition of case groups with internal connections. Based on the case group with internal connections, design a hierarchical task planning framework that includes goal decomposition, constraint recognition, and action sequence generation. Through symbolic planning and heuristic search, generate a preliminary action plan for associated case processing. Based on the preliminary action plan, develop a reinforcement learning optimization model that includes state representation, reward mechanism, and policy network. Through interaction with the environment and policy iteration, optimize the processing strategy of associated cases and improve the processing effect. Based on the reinforcement learning optimization model and the hierarchical task planning framework, implement a dynamic fusion mechanism of planning and learning. Through a two-way feedback mechanism of planning guiding exploration and learning optimizing planning, improve the adaptability of the strategy while ensuring the rationality of the planning, and achieve intelligent processing of complex associated cases.

8. The method according to claim 1, wherein Based on the multi-agent case association processing results, apply the forced zero method sparsification graph technology to achieve effective learning based on the case graph and enhance the knowledge representation and reasoning ability of the system, including: Based on the multi-agent case association processing results, construct a case knowledge graph that includes case elements, legal concepts, and processing processes through techniques such as entity extraction, relationship recognition, and knowledge fusion, providing a data basis for graph learning. Based on the case knowledge graph, design a graph neural network model that includes graph convolutional layers, attention mechanisms, and message passing mechanisms. Through node feature extraction and edge relationship learning, achieve preliminary modeling of the case graph structure. Based on the graph neural network model, implement a forced zero method sparsification mechanism that includes L0 regularization, gated activation, and gradient estimation, by explicitly controlling the model complexity and the number of activated nodes. Based on the forced zero method sparsification mechanism, develop an adaptive learning algorithm that includes meta-learning, knowledge distillation, and incremental training. By dynamically adjusting the learning strategy and model structure, adapt to the graph characteristics of different types of cases and achieve effective learning based on the case graph.

9. The method according to claim 1, wherein Based on the effective learning results of the case graph, construct a case quality assessment knowledge graph to achieve the intelligence and precision of case quality management and provide data support for legal decision-making, including: Based on the effective learning results of the case graphics, through data cleaning, format conversion, and redundancy elimination techniques, multi-source evaluation data including evaluation criteria, case features, evaluation results, and improvement suggestions are integrated to form a unified dataset; Based on the unified dataset, a knowledge graph schema including concept hierarchy, relationship type, and attribute definition is designed. Through ontology engineering and semantic modeling, a conceptual framework in the field of case quality evaluation is constructed to guide the construction of the knowledge graph; Based on the knowledge graph schema and the unified dataset, through techniques such as named entity recognition, relationship extraction, and event detection, unstructured and semi-structured data are transformed into knowledge triples to construct an initial case quality evaluation knowledge graph; Based on the initial case quality evaluation knowledge graph, a knowledge reasoning engine including rule reasoning, path reasoning, and statistical reasoning is developed, and a standardized application interface is designed to support multi-dimensional queries, intelligent analysis, and precise recommendations for case quality management, realizing the intelligence and precision of case quality management and providing data support for legal decision-making.

10. An intelligent case quality evaluation system based on a large language model, characterized in that, Including: A processor and a memory, the memory stores computer-executable instructions, and the processor executes the computer-executable instructions to implement the method according to any one of claims 1 to 9.

Images