Method for constructing interpretable evidence chain atlas under multi-agent cooperation framework oriented to judicial reasoning

Abstract

The invention discloses a method for constructing an interpretable evidence chain atlas under a judicial reasoning-oriented multi-agent cooperation framework, and relates to the technical field of evidence chain atlas construction. Comprising the following steps: acquiring evidence materials, and extracting entity information and legal relationship types; performing legal attribute labeling on the entity information and the legal relationship type, performing compliance review after performing preliminary judgment on the validity of the fact, and identifying and labeling unacceptable evidences; mining a key reasoning path by using a graph algorithm; dynamic weighting is carried out on each associated edge, a visual evidence chain network is constructed, internal parameters of the visual evidence chain network are corrected through an expert-agent alignment mechanism, and reasoning logic is optimized. According to the method, automatic extraction, legality examination, logic relation modeling and interpretable display of evidences can be realized, and the structural integrity of an evidence chain and the preciseness of legal logic are ensured, so that the automation level, transparency and reliability of judicial reasoning are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of evidence chain graph construction, and in particular to a method for constructing an interpretable evidence chain graph under a multi-agent collaboration framework for judicial reasoning. BACKGROUND

[0002] In judicial practice, evidence is the core basis for case adjudication, and its authenticity, legality and relevance directly affect the rigor of judicial reasoning and the fairness of the adjudication result. However, the sources of evidence in real cases are complex and diverse, including documentary evidence, witness testimony, electronic data, audio and video materials, as well as new types of evidence such as bank transaction records, transaction records, and surveillance images. These evidences often show characteristics of fragmentation, heterogeneity and decentralization, making it extremely difficult to manually sort and logically integrate. Existing evidence chain construction mainly relies on the experience and manual comparison of judicial personnel, which is inefficient and prone to logical gaps or reasoning biases, especially when dealing with large-scale electronic evidence and complex causal chains, it is difficult to ensure the completeness and interpretability of reasoning.

[0003] With the development of artificial intelligence, knowledge graph and multi-agent technology, some research and application attempts have introduced intelligent analysis into the judicial field. For example, existing methods achieve task decomposition and reasoning chain synthesis through multi-agent dynamic scheduling to improve the scheduling efficiency of computing tasks; there are also multi-agent collaboration-based contract compliance review methods that use rule knowledge graphs and compliance review agents to identify risks and automatically generate reports for contract clauses. This kind of method has certain advantages in task allocation and legal document review, but its application scenarios are still limited to task scheduling or contract text, lacking a systematic solution for judicial evidence, especially in evidence legality review, logical chain modeling, reasoning interpretability and expert knowledge integration.

[0004] Therefore, to provide a method for constructing an interpretable evidence chain graph under a multi-agent collaboration framework for judicial reasoning to solve the difficulties existing in the prior art is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] Therefore, to provide a method for constructing an interpretable evidence chain graph under a multi-agent collaboration framework for judicial reasoning to solve the difficulties existing in the prior art is a problem that needs to be solved by those skilled in the art.

[0006] To achieve the above purpose, the present application adopts the following technical solutions: A method for constructing an interpretable evidence chain graph under a multi-agent collaboration framework for judicial reasoning, comprising the following steps: Obtaining evidence materials, extracting entity information and legal relationship types by using a language model, and generating standardized knowledge triples; Determining a rule base and a case base, marking the legal attributes of entity information and legal relationship types, preliminarily determining the legality of facts, and obtaining recognized evidence; Conducting compliance review on the recognized evidence, identifying and marking inadmissible evidence, and obtaining reviewed evidence; Generating an evidence network, mining key reasoning paths by using a graph algorithm, locating cornerstone evidence and logical weak links; Based on the evidence credibility, correlation strength and legal relevance, dynamically weighting each correlation edge to realize the quantitative evaluation of the strength of the evidence chain; Constructing a visual evidence chain network, correcting internal parameters of the visual evidence chain network through an expert-agent alignment mechanism, and optimizing reasoning logic.

[0007] Optionally, the entity information includes parties, actions, time and location, and the legal relationship type includes causal relationship, spatiotemporal correlation and subordinate relationship.

[0008] Optionally, generating the evidence network includes: Determining nodes, edges and directionality, determining the weight of the edge based on evidence credibility, correlation strength and legal relevance, and constructing a weighted directed graph model; Using a symbolic display to highlight different edges and nodes to form traceable data; Using a local subgraph evolution algorithm to incrementally update the graph, and combining a closedness determination to ensure that the evidence chain forms a complete logical closed loop.

[0009] Optionally, the node represents an evidence entity or a fact to be proven, the edge represents the support or corroboration relationship between the evidence, and the directionality reflects the causality or logical dependence.

[0010] Optionally, mining the key reasoning path includes: Using an algorithm to analyze the node centrality, sorting the evidence nodes by indicators, and taking the nodes with high PageRank as cornerstone evidence; Taking the final fact to be proven as a target node, and taking the cornerstone evidence node as a starting point, generating the best reasoning path by algorithm, and marking it as the core logical chain of the case; After determining the core logical chain, analyzing the vulnerability of each correlation edge in the main chain to locate the logical weak link; Through depth-first traversal or reachability matrix, judging the node connectivity and edge reachability of the evidence chain of the logical weak link to ensure logical self-consistency.

[0011] Optionally, the expert-agent alignment mechanism includes: Collecting evidence selection, reasoning path and fact determination annotation of legal experts in typical cases to construct an expert experience dataset; The expert experience dataset is used as a verification set to correct the visual evidence chain network and dynamically adjust the edge weight.

[0012] Compared with the prior art, the application provides a construction method of an interpretable evidence chain graph under a multi-agent collaboration framework for judicial reasoning, which has the following advantages: 1) the application can automatically extract key information from massive fragmented evidence and perform multi-dimensional logical modeling through the cooperation of five types of agents, namely evidence extraction, fact determination, review, connection and scoring, so as to realize the automatic construction of the evidence chain and greatly improve the processing efficiency; 2) the application can not only mark the legal properties of evidence and relationships through the double mechanism of the fact determination agent and the review agent, but also identify illegal evidence collection, invalid evidence and other situations, so as to ensure that the final evidence chain meets the strict requirements of judicial practice; 3) the application uses a weighted directed graph model to intuitively present the evidence chain, and the weights of the nodes and edges are calculated based on the evidence credibility, correlation strength and legal relevance, so that judicial personnel can trace back the reasoning basis and legal provisions through an interactive interface, thereby realizing the transparency and interpretability of the whole reasoning process; 4) the application introduces a local subgraph evolution and closedness determination algorithm to support the dynamic updating of the evidence chain, so that the graph can be adjusted in real time when new evidence is added or modified, ensuring the closedness and integrity of the logical chain and improving the adaptability of the system to complex cases; 5) the application uses an expert-agent alignment mechanism to dynamically adjust the reasoning path by using expert preference data and reinforcement learning, gradually solidifies the judicial logic, realizes the intelligent inheritance of legal knowledge, and effectively avoids the black box of AI reasoning; 6) the application is designed for evidence analysis and reasoning of judicial cases, and can be used as an auxiliary judgment tool, as well as in case analysis, judicial training and other scenarios, and has high application value and promotion prospect. BRIEF DESCRIPTION OF DRAWINGS

[0013] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only constitute the embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor based on the provided drawings.

[0014] Figure 1 A construction method of an interpretable evidence chain graph under a multi-agent collaboration framework for judicial reasoning according to the application is disclosed. DETAILED DESCRIPTION

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Reference Figure 1 As shown, this invention discloses a method for constructing an interpretable evidence chain graph within a multi-agent collaborative framework for judicial reasoning, comprising the following steps: Obtain evidence materials, use pre-trained language models such as BERT to extract entity information and legal relationship types, and generate standardized knowledge triples; Establish a rule base and a case base, label entity information and legal relationship types with legal attributes, such as criminal acts and legal transactions, make a preliminary judgment on the legality of facts, and obtain identified evidence; Conduct compliance reviews on the identified evidence, identify and mark inadmissible evidence, and obtain the reviewed evidence; Generate an evidence network, use graph algorithms to mine key reasoning paths, and locate foundational evidence and logical weaknesses; Each link in the evidence chain is dynamically weighted based on its credibility, relevance, and legal relevance, enabling a quantitative assessment of the strength of the evidence chain. A visual evidence chain network is constructed, and the internal parameters of the visual evidence chain network are corrected through an expert-agent alignment mechanism to optimize the reasoning logic.

[0017] Furthermore, the entity information includes: parties involved, actions, time and place; the types of legal relationships include: causal relationship, spatiotemporal relationship, and subordinate relationship. Compliance review includes examining the legality and eligibility of evidence to identify inadmissible evidence such as illegally obtained surveillance footage.

[0018] Furthermore, graph algorithms include PageRank, shortest path, and centrality analysis.

[0019] Furthermore, the evidence-generating network includes: Determine nodes, edges, and directionality; determine edge weights based on evidence credibility, correlation strength, and legal relevance; and construct a weighted directed graph model. Different edges and nodes are highlighted using symbolic methods to create traceable data; The local subgraph evolution algorithm is used to incrementally update the graph, and the closure determination is combined to ensure that the evidence chain forms a complete logical closed loop.

[0020] Specifically, the system uses symbolic methods such as color, node size, and line thickness to highlight core evidence, key nodes, and weak links. Judicial personnel can trace the source evidence, applicable legal provisions, and system logic interpretation of each edge and node through the interactive interface, thereby achieving transparency and traceability of the reasoning process.

[0021] Furthermore, incremental updates include: Using newly introduced or changed evidence as seed nodes, the algorithm expands outward from the seed nodes according to a preset hop count threshold k, extracting a local subgraph containing related evidence entities and their supporting or corroborating relationships. This local subgraph is then used as the area to be updated. The hop count threshold k can be dynamically adjusted based on the type of evidence, the complexity of the case, and the level of reasoning to avoid unnecessary recalculation of the global graph. The following evolutionary operations are performed sequentially in the local subgraph: 1) Exclusion of legality: For evidence nodes or relational edges that are confirmed to be illegally obtained, lack eligibility, or have questionable sources, mark them as inadmissible evidence and reset the weight of the associated edge corresponding to the node to zero; 2) Weight recalculation: Based on the latest evidence credibility, correlation strength and legal relevance indicators, the weights of each related edge in the local subgraph are recalculated to obtain the updated weighted directed subgraph; 3) Structural reinforcement: When the fact node to be proved lacks sufficient supporting paths or presents logical breaks in the local subgraph, potential supporting evidence is automatically retrieved from the rule base and case base to form candidate reinforcement items, and the reinforcement suggestions are submitted to the expert-agent alignment module for confirmation. 4) Graph structure simplification: For redundant nodes and weakly related edges with weights below the set threshold or that do not affect path connectivity, weight reduction or visualization weakening are performed to improve the clarity of the evidence chain graph structure and the salience of dependent paths. The evolved local subgraphs are subjected to closure tests, which include three dimensions: structural integrity, logical consistency, and reasoning coherence. 1) Structural integrity is used to determine whether the facts to be proved are supported by a sufficient number of valid reasoning paths; 2) Logical consistency is used to identify whether there are contradictory pieces of evidence or opposing chains of reasoning in a local subgraph; 3) Reasoning coherence is used to confirm whether an uninterrupted causal or logical dependency link is formed between each evidence node.

[0022] If the closure determination score reaches the preset threshold, the update result of the local subgraph is written back to the global evidence chain graph; otherwise, a reinforcement suggestion task is generated and updated again after manual or expert verification.

[0023] Furthermore, nodes represent evidence entities or facts to be proven, edges represent supporting or corroborating relationships between pieces of evidence, and directionality reflects causal or logical dependence.

[0024] Furthermore, identifying key reasoning paths includes: The algorithm is used to perform node centrality analysis and the evidence nodes are ranked by indicators. Nodes with high in-degree are usually the core facts that multiple pieces of evidence point to, while nodes with high out-degree correspond to key evidence that can support multiple inferences. Nodes with high PageRank have global influence in the overall reasoning chain and are regarded as cornerstone evidence. Taking the final fact to be proven as the target node and the cornerstone evidence node as the starting point, the Dijkstra algorithm is used to search for the shortest logical path from the evidence node to the fact to be proven. At the same time, the edge weight is used as a "persuasive reward" to search for the path with the largest total weight, and the best reasoning path supporting the judgment is obtained. The best reasoning path is marked as the core logical chain of the case. After identifying the core logical chain, vulnerability analysis is performed on each related edge in the main chain to pinpoint weak points in the logic. Using edge weight as a quantitative indicator of the strength of evidence, the following tests are performed on the main chain: 1) Identify the associated edge with the lowest weight and determine it as the most questionable evidence support point in the reasoning chain; 2) If a fact to be proved is supported by only a single low-weight edge, then mark the fact node as a logically weak node; 3) If a fact node supported by a low-weight edge also has a high out-degree, it indicates that it plays a key transit role in the inference chain, and it is marked as a structural risk node. By using depth-first traversal or reachability matrices, the connectivity of nodes and the reachability of edges in the evidence chain of logically weak links are judged to ensure logical consistency.

[0025] Specifically, the connectivity of nodes and the reachability of edges in the evidence chain are assessed to identify the following structural features: 1) Factual points to be proven that are not supported by any evidence: If a node representing a fact to be proven has no incoming edges, it indicates that the fact lacks direct evidence and is marked as an insufficiently supported node. 2) Cases where the reasoning path involves jumps: If there is no reachable path between a certain evidence node and the fact to be proved, it is considered that the reasoning chain is broken, and the intermediate reasoning links need to be supplemented by human intervention or additional evidence. 3) There exists a loop that leads to mutually contradictory conclusions: If a cyclic subgraph with opposite conclusions is detected in the graph, the nodes involved will be marked as regions of reasoning conflict.

[0026] Furthermore, the edges connecting any two nodes in the evidence chain graph are dynamically weighted. These weights consist of three parts: evidence credibility C, association strength R, and legal relevance L. The comprehensive weight calculation formula is as follows: W=αC+βR+γL Among them, α, β, and γ are weighting coefficients that can be adaptively adjusted according to the type of case and the standard of proof.

[0027] Specifically, the credibility of evidence is primarily assessed based on its source, the compliance of its acquisition process, and the integrity of its medium. Evidence with official sources, obtained through legal procedures, and possessing metadata verification chains or digital watermarks receives higher credibility scores. Conversely, evidence with unclear sources, missing links in its chain, flawed acquisition procedures, or deemed by judicial practice to be "potentially subject to external interference" automatically receives a lower credibility score, and the system marks the corresponding edge or node in the graph as "questionable." Credibility scoring is determined by the reviewing agent in conjunction with the RAG rule base to ensure the output has legal basis and auditability. The strength of association is determined by the consistency among evidence in terms of subject, temporal and spatial sequence, and content description. If multiple pieces of evidence point to the same subject's behavior, the strength of the association increases with improved consistency. If there are no contradictions in the information about time or location among the evidence, the algorithm treats them as links with a continuous chain of reasoning, assigning them a high strength of association. When evidence from different sources corroborates each other in terms of content expression, the system further increases the strength of that association. Conversely, if there are contradictions, content conflicts, or logical jumps among the evidence, the strength of the association will automatically decrease. The strength of association is also modulated by the centrality algorithm. When a piece of evidence is pointed to by multiple high-weight nodes, its strength of association will be further enhanced, thus helping to identify the supporting chains of core and weak evidence. Legal relevance measures the degree to which evidence aligns with the facts to be proven in a case. Evidence that directly proves key facts in the legal requirements, such as "action taken," "result occurring," or "causation established," carries a higher weight in legal relevance. Evidence that merely serves as supplementary narration, background information, or explanation of events carries a relatively lower weight. In scenarios involving circumstantial evidence, legal relevance is dynamically adjusted based on the interpretability of the logical deduction chain. If the system identifies multiple levels of inference and a lack of supporting intermediate evidence, it reduces the legal relevance of that link to the evidence, alerting judicial personnel to the potential risk of a broken chain of proof.

[0028] Furthermore, the expert-agent alignment mechanism includes: Collect evidence selection, reasoning paths, and fact-finding annotations from legal experts in typical cases to construct an expert experience dataset; The expert experience dataset is used as a validation set to calibrate the visualized evidence chain network, and the weight parameters are dynamically adjusted and the edge weights are dynamically adjusted based on the Direct Preference Optimization (DPO) method.

[0029] Specifically, when the generated reasoning path deviates from expert experience, the system automatically corrects entity weights or relationship types and optimizes reasoning logic in subsequent cases. Through continuous alignment iterations, the intelligent agent gradually solidifies the expert's legal logic, achieves consistency between reasoning results and judicial practice, effectively solves the "black box" problem of AI reasoning, and improves the interpretability and reliability of the system.

[0030] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing an interpretable chain of evidence graph within a multi-agent collaborative framework for judicial reasoning, characterized in that, Includes the following steps: Obtain evidence materials, extract entity information and legal relationship types using language models, and generate standardized knowledge triples; Establish a rule base and a case base, label entity information and legal relationship types with legal attributes, make a preliminary judgment on the legality of facts, and obtain identified evidence; Conduct compliance reviews on the identified evidence, identify and mark inadmissible evidence, and obtain the reviewed evidence; Generate an evidence network, use graph algorithms to mine key reasoning paths, and locate foundational evidence and logical weaknesses; Each link in the evidence chain is dynamically weighted based on its credibility, relevance, and legal relevance, enabling a quantitative assessment of the strength of the evidence chain. A visual evidence chain network is constructed, and the internal parameters of the visual evidence chain network are corrected through an expert-agent alignment mechanism to optimize the reasoning logic.

2. The method for constructing an interpretable evidence chain graph within a multi-agent collaborative framework for judicial reasoning, as described in claim 1, is characterized in that... Entity information includes: parties involved, actions, time and place; legal relationship types include: causal relationship, spatiotemporal relationship and subordinate relationship.

3. The method for constructing an interpretable evidence chain graph within a multi-agent collaborative framework for judicial reasoning, as described in claim 1, is characterized in that... Evidence generation networks include: Determine nodes, edges, and directionality; determine edge weights based on evidence credibility, correlation strength, and legal relevance; and construct a weighted directed graph model. Different edges and nodes are highlighted using symbolic methods to create traceable data; The local subgraph evolution algorithm is used to incrementally update the graph, and the closure determination is combined to ensure that the evidence chain forms a complete logical closed loop.

4. The method for constructing an interpretable evidence chain graph within a multi-agent collaborative framework for judicial reasoning, as described in claim 3, is characterized in that... Nodes represent evidence entities or facts to be proven, edges represent supporting or corroborating relationships between pieces of evidence, and directionality reflects causal or logical dependence.

5. The method for constructing an interpretable chain of evidence graph within a multi-agent collaborative framework for judicial reasoning, as described in claim 1, is characterized in that... Key reasoning paths include: The algorithm is used to perform node centrality analysis, and the evidence nodes are ranked by indicators. The nodes with high PageRank are used as foundational evidence. The final facts to be proven are taken as the target nodes, and the cornerstone evidence nodes are taken as the starting points. The best reasoning path is generated by the algorithm and marked as the core logical chain of the case. After identifying the core logic chain, vulnerability analysis is performed on each related edge in the main chain to pinpoint the weak points in the logic; By using depth-first traversal or reachability matrices, the connectivity of nodes and the reachability of edges in the evidence chain of logically weak links are judged to ensure logical consistency.

6. The method for constructing an interpretable evidence chain graph within a multi-agent collaborative framework for judicial reasoning, as described in claim 1, is characterized in that... Expert-agent alignment mechanisms include: Collect evidence selection, reasoning paths, and fact-finding annotations from legal experts in typical cases to construct an expert experience dataset; The expert experience dataset is used as a validation set to correct the visualized evidence chain network and dynamically adjust the edge weights.

Images