Social simulation-based e-commerce dispute resolution method

Abstract

This invention relates to the field of artificial intelligence technology and discloses a method for adjudicating e-commerce disputes based on social simulation. The method includes: acquiring multimodal data of the case to be adjudicated and initializing the environment; employing an active iterative evidence collection mechanism to iteratively perform key evidence location and evidence perception in the evidence, generating a structured evidence fact matrix for both parties in the dispute; introducing a precedent library for adjudication and creating adjudication guidelines; based on the structured evidence fact matrix, executing a three-stage multi-agent adversarial debate including opening statements, cross-rebuttals, and closing arguments, generating a standardized and calibrated debate text sequence; constructing a debate graph, performing logical reasoning through a graph attention network, and outputting a traceable and explainable adjudication result and reasoning. This invention significantly improves processing efficiency and reduces operating costs while ensuring high-precision adjudication, and can efficiently support the automated processing of massive e-commerce disputes, ensuring a high degree of unity between adjudication efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and specifically to a method for adjudicating e-commerce disputes based on social simulation. Background Technology

[0002] The key to achieving intelligent adjudication of e-commerce disputes lies in the fine-grained understanding of multimodal, adversarial, and fragmented evidence, and in completing causal reasoning and fair judgments based on flexible rules such as platform agreements and transaction practices. However, most existing dispute adjudication technologies are concentrated in the legal and judicial field, using professional judges and lawyers as simulation subjects, and making judgments based on structured legal documents and explicit legal provisions. These technologies differ fundamentally from e-commerce disputes in terms of evidence form, reasoning logic, and adjudication basis, and cannot be directly applied. Therefore, this invention proposes an e-commerce dispute adjudication method based on social simulation, aiming to overcome existing technological bottlenecks and achieve fine-grained analysis of multimodal evidence, adversarial logical reasoning, flexible rule alignment, and interpretable fair judgments, providing an efficient, stable, and reliable automated adjudication solution for e-commerce disputes.

[0003] In recent years, multi-agent technology based on large language models has developed rapidly, demonstrating powerful capabilities in complex decision-making tasks through role simulation, adversarial debate, and collaborative reasoning, providing new ideas for intelligent dispute adjudication. For example, AgentCourt constructs a multi-agent court trial simulation framework, simulating roles such as judges and lawyers, and continuously strengthening the agents' legal understanding and reasoning abilities through multiple rounds of adversarial debate, achieving accurate judgments on legal disputes. However, current multi-agent research on dispute adjudication focuses on the legal field and is difficult to directly extend to the e-commerce field. On the one hand, the processing objects and evidence of existing judicial adjudication models are limited to structured, unimodal legal documents, while e-commerce disputes rely on multimodal, fragmented, and highly redundant evidence such as text, images, and videos. The models lack the ability to fine-grainedly analyze unstructured multimodal evidence and locate key clues, and cannot clarify the causal logic between evidence and the focus of the dispute. On the other hand, judicial trials are subject to clear and rigid legal constraints, while e-commerce dispute adjudication is based on scattered and dynamically changing rules, mainly relying on flexible rules such as platform conventions and transaction practices. Existing models are susceptible to cognitive biases and cannot guarantee the fairness, consistency, and interpretability of the adjudication. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for adjudicating e-commerce disputes based on social simulation.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A social simulation-based method for adjudicating e-commerce disputes includes: Acquire multimodal data of the case to be adjudicated and initialize the environment; the multimodal data includes transaction information, the claims of the disputing parties, and the evidence set corresponding to the claims; the evidence includes text, images, and videos; A proactive iterative evidence collection mechanism is adopted, taking the demands as the guide, iteratively executing the location and perception of key evidence in the evidence, and independently generating a structured evidence fact matrix for both parties in the dispute; A library of precedents is introduced as a normative constraint, and an adjudication guide is created. Based on a structured evidence fact matrix, a three-part multi-agent adversarial debate, including opening statements, cross-rebuttals, and closing arguments, is executed, generating a normatively calibrated sequence of debate texts. The debate text sequence and case background information consisting of transaction information and the claims of both parties are mapped into a debate graph that includes temporal sequence, adversarial relationship and global anchor relationship. Logical reasoning is performed through graph attention network to output a verifiable and explainable ruling result and reasoning.

[0006] In one embodiment, the proactive iterative evidence collection mechanism, guided by the claims, iteratively performs key evidence location and evidence perception within the evidence, independently generating a structured evidence fact matrix for both parties in the dispute, specifically including: The proactive iterative evidence collection mechanism operates independently and in parallel from the perspectives of both parties in the dispute, continuing until sufficient clues are available or the maximum number of iterations is reached. Each iteration includes: The key evidence location step is used to select from the evidence set the evidence most relevant to the current point of contention and the claims of both parties in the dispute. The fine-grained evidence perception step is used to directly analyze the images of selected evidence or to analyze the videos of selected evidence frame by frame to extract key clues and form evidence of argumentation. After the iteration is completed, a structured evidence fact matrix for both parties in the dispute is output.

[0007] In one embodiment, the key evidence location step is used to select from the evidence set the evidence most relevant to the current point of contention and the claims of both parties, specifically including: For the buyer, in the first During the round of iteration, based on the case background information The evidence includes a collection of texts from the buyer's evidence, from which the most relevant piece of evidence is selected to verify the truthfulness of the buyer's statements. ; in, For the index of the selected evidence, For the first Buyer evidence selected in rounds of iterations. Indicates the first Index of round iteration, Indicates as of the date The wheel is a collection of texts consisting of all the evidence presented by the selected buyers. This refers to the text set consisting of all the text in the buyer's evidence. Indicates that during the first The text set of buyer evidence that was not selected during the round of iteration; The prompts indicate how to extract the points of contention based on case background information and evidence text, and how to select evidence. For the seller, in the first During the round of iteration, based on the case background information The evidence includes a collection of texts from the seller's evidence. From this collection, select the piece of evidence that best reflects the point of contention to verify the seller's statement. ; For the first Seller evidence selected in rounds of iteration; Indicates as of the date A round is a collection of texts consisting of all selected seller evidence. A collection of texts representing all the seller's evidence. Indicates that during the first The set of texts of seller evidence that were not selected during the round of iteration.

[0008] In one embodiment, the fine-grained evidence perception step is used to directly analyze the image of the selected evidence or to analyze the video of the selected evidence frame by frame to extract key clues and form evidence of argumentation, specifically including: If the selected evidence is an image, the image is directly analyzed; if the selected evidence is a video, the video is analyzed frame by frame. For the buyer, it is crucial to extract key leads directly related to the dispute. And provide the key clue that can support or weaken the buyer's argument. : ; in, Indicates the first In the round of iterations, key clues from the buyer Buyer's justification The output set is composed of Clue words for intelligent agents to perceive key cues in a fine-grained manner; For the seller, it is crucial to extract key leads directly related to the dispute. And provide the seller's arguments that support or weaken the seller's statement based on this key clue. : ; in, Indicates the first In the round of iterations, key clues from the seller Seller's justification The output set; After the iteration is completed, a structured evidence fact matrix for both parties in the dispute is output, specifically including: By summarizing the evidence perceptions from both sides of the dispute, a structured matrix of evidence facts is generated. ; This represents the evidence perception result obtained by the buyer after completing all iteration rounds. It integrates the key clues and arguments of the buyer from each round of fine-grained evidence perception output from the buyer's perspective; This represents the evidence perception result obtained by the seller after completing all iteration rounds. It integrates the key clues and arguments of the seller from each round of fine-grained evidence perception output from the seller's perspective; This represents the maximum number of iteration rounds.

[0009] In one embodiment, the introduction of a library of adjudication precedents as normative constraints and the creation of adjudication guidelines specifically include: Trial Precedent Library , For adjudicated cases, To extract structured adjudication rules from adjudicated cases; through a semantic encoder. Treating adjudicated cases The text and the claims of both parties in the dispute are characterized using a retrieval function. Calculate similarity and return A number of similar cases serve as guidelines for adjudication. : .

[0010] In one embodiment, the three-part multi-agent adversarial debate based on a structured evidence fact matrix, comprising an opening statement, cross-refutation, and closing statement, specifically includes: The three-stage multi-agent adversarial debate is executed sequentially by the buyer's lawyer agent and the seller's lawyer agent: During the opening statement phase, each side declares its arguments based on its perceived evidence, generating the content of its opening statement; In the cross-rebuttal phase, the logical flaws and evidentiary validity of the opposing side's opening statement are refuted and examined. During the closing arguments phase, both parties filter out invalid arguments, retain valid evidence that has been verified through cross-examination, and submit their requests for a ruling.

[0011] In one embodiment, during the opening statement phase, the arguments are stated based on the applicant's perception of evidence, specifically including: First Buyer's Lawyer Intelligent Agent With the first seller's lawyer's intelligent agent Based on their respective perceptions of evidence, each speaker takes turns to present their arguments, extracting supporting factual evidence from a structured evidence matrix to generate their opening statement: ; ; This refers to the opening statement of the first buyer's lawyer's AI agent. This refers to the opening statement of the first seller's lawyer's AI agent. This indicates that the debate generation prompts enable the lawyer agent to generate the debate text based on the given prompts. For case background information, This represents the buyer's perception of complete evidence within the structured evidence fact matrix. This represents the seller's perceived completeness of evidence in the structured evidence fact matrix. Adjudication Guidelines; The cross-rebuttal phase involves refuting and examining the logical flaws and evidentiary validity of the opposing side's opening statements, specifically including: ; ; This represents the cross-rebuttal content of the second buyer's lawyer's AI agent. This represents the cross-rebuttal content of the second seller's lawyer's intelligent agent. For the second buyer's lawyer intelligent agent, For the second seller's lawyer intelligent agent; Both parties filtered out invalid arguments, retained valid evidence verified through cross-examination, and submitted a request for a ruling, specifically including: ; ; This refers to the closing statement of the third-party buyer's legal counsel. This refers to the closing statement of the third-party seller's lawyer agent. This refers to the third-party buyer's lawyer's intelligent agent. This represents the third-party seller's lawyer's intelligent agent; Will Combined into a debate text sequence .

[0012] In one embodiment, mapping the debate text sequence and case background information consisting of transaction information and the claims of both parties to the dispute into a debate graph that includes temporal sequence, adversarial relationships, and global anchor relationships specifically includes: Mapping debate text sequences with case background information into debate graphs Node set It contains 7 nodes in total. to These are debate nodes, corresponding sequentially to the debate outputs of the two sides in a multi-agent adversarial debate. ; This is a case background information node, used to aggregate transaction information and the demands of both parties in the dispute; edge set It is divided into three categories: intra-role temporal edges, which are used to represent the evolutionary relationship of the arguments of lawyer agents with the same position; cross-role adversarial edges, which are used to explicitly model the rebuttal relationship between the two parties in a dispute; and fully connected anchor edges, which are bidirectionally connected to the case background information nodes and all debate nodes.

[0013] In one embodiment, the step of performing logical reasoning through a graph attention network to output a verifiable and explainable ruling and reasoning specifically includes: In the debate graph, the debate outputs of each round of the dispute between the two sides in the multi-agent adversarial debate are mapped as debate nodes; the case background information is mapped as case background information nodes, and the debate nodes and case background information nodes are collectively referred to as nodes. Semantic features of each node in the debate graph are extracted and latent vectors representing node identity and stage attributes are fused to obtain node features. Using the case background information node as the global logical anchor point, message passing and feature aggregation are carried out between nodes through a multi-layer graph attention network. During the propagation process, the attention coefficient is used to measure the importance of each node to the final decision. Based on the final attention coefficient of each debate node to the case background information node in the graph attention network, high-influence debate nodes are selected to generate traceable decision reasons. The projected case background information node features are used as queries. Interactive attention calculation is performed on the features of each debate node, and the final ruling representation is obtained based on the interactive attention fusion result. The ruling result is then output based on the final ruling representation.

[0014] In one embodiment, the extraction of semantic features from each node in the debate graph and the fusion of latent vectors representing node identity and stage attributes to obtain node features specifically include: Through semantic encoder Extracting semantic features of nodes: , ; in, Indicates the first The debate text embedding of each debate node, , Text embedding indicating case background information; Background information for the case; For the first The debate output corresponding to each debate node; Through role perception, perform semantic alignment of roles: ; in, The latent vector representing the identity and stage attributes of each node. For a complete character representation vector, ; Learnable role-aware linear transformation weight matrix; Node features are obtained through feature concatenation and fusion. : ; The process uses case background information nodes as global logical anchors, and employs a multi-layer graph attention network for message passing and feature aggregation among nodes. During propagation, attention coefficients are used to measure the importance of each node to the final ruling. Based on the final attention coefficients of each debate node to the case background information nodes in the graph attention network, high-influence debate nodes are selected, and traceable ruling reasons are generated. Specifically, this includes: A multi-layer graph attention network is used to complete message passing and feature aggregation between nodes: ; in, For the i-th node in the... Features of the layer For activation function, Let be the set of neighboring nodes of the i-th node. For the first The learnable linear transformation matrix of the layer; This is the attention coefficient used to measure the logical importance of the j-th neighboring node to the current i-th node; extract the... The final attention coefficient between the case background information nodes and the debate nodes after the layered graph attention network is used to filter high-impact debate nodes and generate the ruling reasoning. : ; These are prompt words used to generate the reasons for the ruling. Select an operator for sorting the attention coefficient set from largest to smallest and return the node numbers corresponding to the top n attention coefficients; Indicates the process After the layered graph attention network operation, the case background information nodes With the Debate Nodes The final attention coefficient between them; This indicates the number of high-impact debate nodes selected. The process involves using the projected case background information node features as queries, performing interactive attention calculations on the features of each debate node, obtaining the final ruling representation based on the interactive attention fusion result, and outputting the ruling result based on the final ruling representation. Specifically, this includes: The feature matrices of the 7 nodes output by the multi-layer graph attention network are denoted as the logical representation. ,Will Projected to a unified dimension: , ; in, This represents the global feature matrix after projection of all nodes. This represents the global projection weight matrix, which consists of learnable parameters. This represents the feature vector after individual projection of the case background information node. This represents the case node-specific projection weight matrix, which consists of independent learnable parameters. The features of the case background information nodes after being updated by a multi-layer graph attention network; then... For the query vector, with the global feature matrix The projected feature submatrices corresponding to the 6 debate nodes obtained from the extraction are used as keys and values ​​to perform interactive attention calculation: ; Indicates from the global feature matrix The projected feature submatrices corresponding to the 6 debate nodes are extracted from the data. The query vector representing interactive attention. Keys representing interactive attention The value representing the interaction attention; ; This represents the output features after interactive attention fusion. Indicates the feature projection dimension. This is a matrix transpose operation. Represents the Softmax function; After that Pooling and focus interaction features are concatenated to obtain the final decision representation: ; Represents global aggregated features. This represents the graph pooling function. The final decision obtained after feature cascading is represented. This indicates a feature concatenation operation; The final decision result is output by the classifier. : ; It is a multilayer perceptron.

[0015] Compared with the prior art, the beneficial technical effects of the present invention are: 1. Filling the technological gap in intelligent adjudication of multimodal e-commerce disputes: This invention addresses the characteristics of multimodal, adversarial, and flexible rule-based disputes in real-world e-commerce scenarios. For the first time, it constructs a complete integrated dispute adjudication framework based on proactive evidence perception, multi-agent adversarial debate, and interpretable adjudication using graph attention networks. This solves the problems of long adjudication cycles and high costs associated with traditional manual small court mechanisms. Experimental results show that on real-world multimodal e-commerce dispute datasets, the accuracy of this invention significantly outperforms existing mainstream large language models, multimodal large models, and court simulation methods. Compared to the best-performing large language models and multimodal large models, performance is improved by 4.71% and 4.61%, respectively; compared to existing court simulation frameworks, performance is improved by 5.61%.

[0016] 2. Achieved deep perception of multimodal evidence and precise causal logic reasoning: This invention innovatively proposes an active iterative evidence collection mechanism based on location and perception. It upgrades the existing model's passive, one-time, and full-volume evidence input coarse-grained processing method to an iterative key clue location and fine-grained analysis guided by the focus of the dispute. This enables the precise extraction of decisive visual and textual clues such as product defects, abnormal status, timestamps, and logistics information from massive amounts of redundant information, significantly alleviating the pressure on the context window of large multimodal models and greatly improving the efficiency of evidence utilization. Simultaneously, through a three-stage adversarial debate consisting of an opening statement, cross-refutation, and concluding remarks, fragmented, adversarial, and unstructured evidence is transformed into a logically rigorous and progressively developing chain of reasoning. This precisely clarifies the causal relationship between the evidence and the focus of the dispute, ensuring that the ruling no longer relies on simple pattern matching but is based on complete, traceable, and verifiable logical deduction, significantly improving the professionalism, rigor, and persuasiveness of the ruling.

[0017] 3. Ensuring the fairness, consistency, and stability of adjudication under dynamic and flexible rules: E-commerce disputes lack rigid, legally binding regulations and rely heavily on platform practices, common sense in transactions, and group consensus. Traditional models are susceptible to pre-training biases, positional drift, and arbitrary logic, leading to chaotic adjudication standards and uncontrollable results. This invention follows a precedent mechanism by introducing a precedent library, extracting accepted adjudication norms formed from historical cases into explicit constraints. This provides agents with a unified, objective, and interpretable adjudication benchmark, correcting cognitive biases and positional deviations at the source. Simultaneously, by combining multi-agent adversarial debate and graph attention network structured reasoning, the adjudication process fully incorporates diverse perspectives, avoiding the one-sided judgments of a single agent. This ensures that in complex, flexible, and dynamically changing e-commerce transaction scenarios, the adjudication standards remain consistent, the position is neutral, and the results are stable, significantly improving the fairness and credibility of the adjudication.

[0018] In summary, this invention achieves comprehensive breakthroughs in multimodal evidence understanding, causal reasoning, and fair adjudication. While ensuring high-precision adjudication, it significantly improves processing efficiency and reduces operating costs, and can efficiently support the automated processing of massive e-commerce disputes, ensuring a high degree of unity between adjudication efficiency and accuracy. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the method of the present invention.

[0020] Figure 2 This is a schematic diagram of the framework structure proposed by the method of this invention. Detailed Implementation

[0021] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, a social simulation-based e-commerce dispute adjudication method of the present invention includes the following steps: S1. Acquire multimodal data of the case to be adjudicated and initialize the environment; the multimodal data includes transaction information, the claims of the disputing parties, and the evidence set corresponding to the claims; the evidence includes text, images, and videos; S2 adopts an active iterative evidence collection mechanism, taking the demands as the guide, iteratively executing the location and perception of key evidence in the evidence, and independently generating a structured evidence fact matrix for both parties in the dispute. S3 introduces a library of precedents as normative constraints and creates adjudication guidelines. Based on a structured evidence fact matrix, it executes a three-part multi-agent adversarial debate, including opening statements, cross-rebuttals, and closing arguments, generating a normatively calibrated sequence of debate texts. S4 maps the debate text sequence and case background information consisting of transaction information and the claims of both parties to the dispute into a debate graph containing temporal, adversarial and global anchor relationships. Logical reasoning is performed through graph attention network to output a verifiable and explainable ruling result and reasoning.

[0023] This invention proposes a framework integrating multi-agent adversarial debate and graph attention network reasoning to achieve efficient and fair intelligent adjudication of e-commerce disputes. In the debate module, the invention first proposes an active iterative evidence-gathering mechanism based on localization and perception, performing fine-grained analysis and key clue extraction of multimodal evidence to accurately establish causal relationships between evidence and disputes, significantly improving the accuracy and interpretability of reasoning. Subsequently, the invention constructs a three-stage adversarial debate process, breaking down the adjudication process into opening statements, cross-refutations, and closing statements. Through multiple rounds of game-playing between the lawyer agents of both sides, the invention deeply explores the points of contention in the evidence. Simultaneously, at the normative constraint level, the invention introduces a precedent library, providing platform consensus-level rule guidance to the agents by retrieving similar historical cases, effectively correcting model cognitive biases and ensuring consistency and fairness in adjudication. In the adjudication module, the invention constructs a structured debate graph from the debate content and case background information, utilizes graph attention networks to model temporal dependencies and adversarial relationships, and completes logical fusion through a focus-driven attention mechanism, ultimately outputting a traceable and interpretable adjudication result and reasoning.

[0024] The present invention will be described in detail below in several parts.

[0025] 1. Task request and environment initialization.

[0026] During the initialization phase, the system first acquires full multimodal data of the dispute case to be adjudicated, including transaction information submitted by both parties, their claims, and corresponding text, image, and video evidence. Simultaneously, it initializes the lawyer's intelligent agent configuration and decision-making environment, specifically including: loading the multimodal large model and graph attention network parameters, constructing and loading the adjudication precedent library index, and completing evidence preprocessing to provide a standardized input environment for subsequent debate and adjudication. The intelligent agent is driven by the multimodal large model.

[0027] Specifically, regarding dispute cases It contains case background information. (Including transaction information and the demands of both parties), and also includes Buyer's Evidence ,as well as Seller's Evidence For buyer evidence, each buyer's evidence , containing text Image collection and video collection Similarly, each seller's evidence , containing text Image collection and video collection ; .

[0028] 2. Proactively iterate evidence perception.

[0029] The core of adjudication is to accurately locate key clues from redundant, heterogeneous, and unstructured multimodal evidence and establish causal relationships between the evidence and the focus of the dispute. Existing methods generally adopt a coarse-grained processing mode of one-time full input, which not only occupies a large number of context windows but also fails to analyze key details in images and videos with fine granularity, making it difficult to achieve reliable causal inferences. To address this, this invention proposes an active iterative evidence collection mechanism based on localization and perception, breaking down e-commerce dispute evidence processing into two progressive stages: key evidence localization and fine-grained evidence perception, achieving a cognitive deepening from full coarse encoding to key evidence focusing. Specifically, the system employs the following mechanisms: 1) Key Evidence Location: Focusing on the points of contention and the claims of both parties, the system proactively selects and locates the most representative multimodal evidence blocks from the evidence set, discarding redundant and irrelevant information to avoid wasting resources from full input; 2) Fine-Grained Evidence Perception: The selected image and video evidence undergoes meticulous analysis, region by region and frame by frame, extracting key clues such as visual information, state characteristics, timestamps, and logistics information directly related to the dispute. This forms arguments that can support or weaken claims, and the system iterates independently from the perspectives of both parties until sufficient clues are available or the maximum number of iterations is reached. Finally, a structured evidence fact matrix is ​​output for both parties. Through this mechanism, the system transforms massive amounts of redundant multimodal raw evidence into high signal-to-noise ratio structured facts that can be directly used in debate, effectively alleviating the context window pressure of large multimodal models and improving the reliability of causal reasoning.

[0030] The proactive iterative evidence collection mechanism mimics the review logic of human reviewers, transforming the understanding of massive amounts of multimodal evidence at once into an iterative key clue localization process guided by the dispute's claims. This effectively alleviates the model's context window load and improves the accuracy of key clue localization. To ensure the fairness of the judgment and eliminate potential prior interference from mixed-source evidence, the evidence perception process operates independently and in parallel from the perspectives of both parties in the dispute. Taking the buyer's t-th round of evidence collection as an example, the specific execution steps are as follows: Key evidence location. First, the agent proactively locates the evidence most likely containing key clues within the evidence set, based on the textual clues in the dispute. This is done in conjunction with case background information. The text set consisting of the text in the buyer's evidence. Actively select one piece of evidence from the evidence set that best reflects the core of the conflict in order to verify the truthfulness of the buyer's statement: ; in, For the selected evidence index, Indicates as of the date A wheel is a collection of texts consisting of all the selected evidence. For the index of the selected evidence, For the first Buyer evidence selected in rounds of iterations, superscript This indicates that the evidence belongs to the buyer, and the bid is placed below. This represents the index of the t-th iteration.

[0031] Fine-grained evidence perception: the agent's perception of selected evidence For fine-grained perception, the agent focuses on visual focal points such as product defects and logistics documents for image evidence; for video evidence, it uses frame extraction for perception, focusing on capturing key frames and matching timestamp semantics. This process aims to extract key clues directly related to the dispute. And provide the evidence to support or weaken the buyer's statement. : ; in, The concept of historical perception is introduced to ensure the coherence and logical progression of clue extraction. This process is driven by the sufficiency of evidence and undergoes multiple iterations. It continues until the identified clues are sufficient to support the final judgment logic, or until the preset maximum number of iterations is reached. At that point, the perception process ceases. The seller's perspective on evidence perception proceeds symmetrically.

[0032] The basis for the argument is that the intelligent agent performs fine-grained analysis of selected images and videos, region by region and frame by frame, extracts key clues, and then combines these key clues with the disputed claims to arrive at a structured logical conclusion. The basis for the argument clarifies the supporting or weakening effect of the extracted key clues on the party's claims, completing the transformation from objective materials to subjective logical argumentation.

[0033] Ultimately, this module summarizes the perceptions from both sides' perspectives and generates a structured matrix of evidence facts. It summarizes the key clues and arguments from both sides' perspectives, condensing the raw and redundant multimodal data into argumentative elements with a high signal-to-noise ratio, providing a solid foundation of clues for the subsequent adversarial debate stage.

[0034] 3. Multi-agent adversarial debate and ruling precedent constraints.

[0035] To deeply explore the points of contention, correct cognitive biases of the agents, and further enhance the logic and credibility of the rulings, this invention constructs a three-stage structured adversarial debate process: opening statement, cross-refutation, and closing statement. It also introduces a precedent library to provide explicit and unified normative constraints for the agents, ensuring that the overall debate process closely resembles the evidence presentation, rebuttal, and trial logic of a real court, guaranteeing that the final decision aligns with platform consensus and social fairness. Based on structured evidence, this invention establishes a professional adversarial debate framework for both sides' lawyer agents: 1) Opening Statement: Both agents clarify their core claims, factual events, and key evidence based on their own evidence, presenting their positions and evidence comprehensively; 2) Cross-Refutation: Both sides cross-examine and refute the other's statements, the validity of their evidence, and the logical rationality, actively pointing out contradictions, loopholes, and unreasonable aspects; 3) Closing Statement. In this process, the present invention strictly adheres to the "following precedent" principle in case law. Through semantic encoding and similarity retrieval, it accurately matches historical cases highly similar to the current case from a pre-set database of precedents. Structured adjudication guidelines conforming to group consensus are injected as external constraints throughout the debate process, guiding the agent to conduct reasoning and rebuttal within a compliant, fair, and consistent framework. This effectively avoids imbalances and inconsistent standards in adjudication caused by inherent model biases, positional biases, or cognitive deviations. Through the dual support of structured debate and precedent norms, the present invention significantly enhances the rigor and fairness of adjudication, ultimately forming a six-round debate text sequence that is chronologically clear, well-argued, logically rigorous, and calibrated according to norms. This provides high-quality, reliable, and interpretable reasoning support for subsequent fair adjudication based on graph attention networks.

[0036] Specifically, in the multi-agent adversarial debate and precedent constraint stage, in order to correct the cognitive biases of agents, standardize the debate logic, and ensure the consistency of the basis for adjudication, this invention introduces a precedent library to provide normative guidance at the level of social consensus for the debate process, and constructs a three-stage structured adversarial debate mechanism, enabling agents to complete evidence examination and logical game within a compliant, fair, and unified framework.

[0037] To eliminate subjective biases and logical deviations introduced by pre-training corpora, this invention constructs a precedent database based on the principle of "following precedent." ,in It is a historical precedent (i.e., a real case). This is a structured set of adjudication rules extracted from case law. It addresses cases pending adjudication. The system uses a semantic encoder The case text and the claims of both parties are represented by features, and then retrieved using a retrieval function (cosine similarity). Calculate similarity and return the adjudication guidelines for the top-K similar cases. : .

[0038] Injecting this ruling guideline into the decision-making process of lawyers' intelligent agents can force the debate positions to align with the platform rules, effectively correct cognitive biases, and improve the fairness and consistency of the rulings.

[0039] Adversarial debate is the core game-theoretic reasoning element of this invention. This involves obtaining a structured evidence fact matrix. Subsequently, this invention constructs a set of buyer's lawyer intelligent agents. With seller's lawyer intelligent agent set The adversarial system is based on evidence and guided by precedent. To ensure consistency, the debate follows a three-part format: opening statement, cross-referencing, and closing statement.

[0040] Define the debate text sequence as The overall debate process is as follows: Opening Statement Phase: Used to construct facts and establish a position. (Buyer's Lawyer Intelligent Agent) With the seller's lawyer AI agent Based on their respective perceptions of evidence, they spoke in turn, transforming visual cues and chains of facts into normative arguments: ; .

[0041] The cross-refutation phase: used for cross-examination and logical deconstruction, this is the stage with the highest adversarial intensity and the greatest information gain. Both agents refute each other's opening statements, evidence, and logical flaws, explicitly revealing the contradictions and conflicts. ; .

[0042] Concluding remarks phase: Used to achieve logical convergence and decision request. Based on the dynamics of the entire debate, both agents eliminate invalid arguments, retain valid facts that have been examined, and finalize their claims. ; .

[0043] Through a three-stage structured adversarial debate and the constraints of precedent, the system ultimately generates a six-round debate text sequence that is chronologically complete, logically rigorous, and has standardized positions, providing a high-quality and highly reliable semantic foundation for subsequent interpretable adjudication based on graph attention networks.

[0044] 4. Interpretable decision output based on graph attention network.

[0045] To model the temporal dependencies, adversarial relationships, and logical connections between debate stages and achieve a fair, traceable, and explainable final ruling, this invention designs a ruling module based on graph attention networks and interactive attention. The specific process is as follows: 1) Debate graph construction: Mapping the content of the six rounds of debate and case background information into a debate graph containing 7 functional nodes, setting temporal dependency edges, cross-role adversarial edges, and global anchor edges to explicitly model the debate logical structure; 2) Role-aware feature fusion: fusing the semantic features of the debate with identity features and stage features through a role-aware encoder to enhance feature distinguishability; 3) Graph network logical reasoning. 4) Focus-driven interactive attention fusion: Using the core claims of the case as the query vector, the interactive attention mechanism focuses on key debate content and fuses global pooling features and focus features; 5) Adjudication classification and reason generation: The fused features are input into the classifier to output the probability of the buyer and seller winning, determine the final winning party, select high-influence debate nodes according to attention weights, and fuse the corresponding debate texts with case background information to generate clear, credible and traceable adjudication reasons.

[0046] Specifically, in the interpretable adjudication stage based on graph attention networks, in order to structurally model the temporal dependencies, adversarial relationships and logical connections between the content of the debate, and at the same time realize the traceability of the adjudication process and the interpretability of the decision basis, this invention constructs an adjudication module based on graph attention networks and interactive attention, which uniformly maps the debate sequence and case background information into a topological structure and completes accurate reasoning.

[0047] Debate Graph Construction and Formal Definition. Addressing the discourse sequences and nonlinear logical interactions generated in multi-agent adversarial debates, this invention maps debate texts and case backgrounds into debate graphs. Directed edges are used to capture the topological relationships between different stages and roles. The node set... It contains a total of 7 functional nodes. to These are the debate nodes, corresponding to the arguments delivered by both sides in the dispute. This serves as a case background information node, used to aggregate the dispute's theme and the claims of both parties, acting as a global logical anchor to ensure that reasoning does not deviate from the core of the case. (Edge set) These are categorized into three types: intra-role temporal edges, used to represent the evolutionary relationship of arguments within the same agent; cross-role adversarial edges, used to explicitly model the rebuttal relationship between the two sides; and fully connected anchor edges, which bidirectionally connect case background information nodes with all debate nodes, allowing global information to participate in every message transmission. The rebuttal relationship refers to the interactive connection between the lawyer agents of the buyer and seller sides during multiple rounds of debate, where one side questions, refutes, and cross-examines the other side's statements, evidence, and logical arguments. It is a logical relationship of opposing and mutually refuting viewpoints. For example, there is a clear rebuttal relationship between two lawyer agents in the cross-rebuttal phase.

[0048] Role-aware node feature integration. To enable the adjudication model to perceive the differences between debate roles and stages, this invention introduces a role-aware encoder to achieve deep alignment between semantics and identity. First, semantic features are extracted: , ; in, Text embeddings representing debate.

[0049] Next, semantic alignment of roles is performed: .

[0050] in, The latent vector represents the identity and stage attributes of each node. This latent vector is a learnable parameter that is randomly generated during model initialization and continuously optimized during overall model training to learn the identity and stage characteristics of each node. This is the complete character representation vector. Finally, node representations are obtained through feature concatenation and fusion. : .

[0051] Logical deduction based on multi-layer graph attention networks. This invention uses a multi-layer graph attention network as the core inference engine to complete message passing and feature aggregation between nodes: .

[0052] in, For nodes In the Features of the layer For activation function, For nodes The set of neighboring nodes, It is a learnable linear transformation matrix. The attention coefficient is a core parameter of Graph Attention Networks (GAT), used to measure the performance of neighboring nodes. For the current node The logical importance of attention. Based on the evolution of attention weights, this invention achieves decision interpretability and extracts the logical importance of attention. The final attention coefficient between case nodes and debate nodes after the layered graph attention network (GAT) is used to filter high-impact debate nodes and generate ruling reasons: .

[0053] Interactive attention fusion and final decision-making. Obtaining the evolved logical representation. Then, the system projects the features onto a unified dimension: , .

[0054] in, The features of the case background information nodes are updated after passing through a multi-layer graph attention network. Then, using... For the query vector, with the global feature matrix The projected feature submatrices corresponding to the 6 debate nodes obtained from the extraction are used as keys and values ​​to perform interactive attention calculation: ; ; The graph is then pooled and concatenated with focus interaction features to obtain the final decision representation: ; .

[0055] The final decision result is output by the classifier: .

[0056] Example: This embodiment fully simulates the entire process of environment initialization, proactive evidence perception, adversarial debate reasoning, and interpretable adjudication based on graph attention networks in a multimodal e-commerce dispute case. It mainly includes four core parts: task request and environment initialization, proactive iterative evidence perception, multi-agent adversarial debate and adjudication precedent constraints, and interpretable adjudication output based on graph attention networks.

[0057] 1. Dispute Adjudication Request and Environment Initialization. Upon receiving an e-commerce dispute adjudication request, this invention first performs unified initialization of the input case and operating environment. It acquires all multimodal data of the case to be adjudicated, including case background information, textual evidence, image evidence, and video evidence, and performs standardized preprocessing on the image and video evidence. Simultaneously, it loads the multimodal large model and graph attention network parameters, constructs and indexes the adjudication precedent library, and completes the semantic encoder initialization. Furthermore, the system configures the roles and behavioral norms of the buyer's and seller's lawyer intelligent agents, sets debate process constraints and inference parameters, and provides a stable and unified operating environment for subsequent evidence perception, adversarial debate, and interpretable adjudication.

[0058] 2. Proactive Iterative Evidence Perception Based on Evidence Perception Mechanism. This section aims to accurately extract key clues from redundant and heterogeneous multimodal evidence, alleviate the pressure on the context window of a large model, and establish a causal relationship between evidence and the focus of the dispute. The lawyer's agent first analyzes transaction information and statements from both parties to extract the core points of contention in the case, using these as anchors for evidence retrieval. Then, a proactive iterative evidence collection process is initiated: In the key evidence location stage, the agent actively filters multimodal evidence that best reflects the core of the contradiction within the evidence set, rather than reading all information at once, focusing on the points of contention; in the fine-grained evidence perception stage, selected images and videos are analyzed region by region and frame by frame to extract key clues such as product defects, status characteristics, timestamps, and logistics information, forming arguments that support or weaken corresponding claims. This process is executed independently and symmetrically from the buyer's and seller's perspectives, iterating multiple times until sufficient clues are available or the maximum number of iterations is reached, ultimately outputting a structured evidence fact matrix, transforming redundant raw data into high signal-to-noise ratio structured facts that can be directly used in debate.

[0059] 3. Multi-Agent Adversarial Debate and Precedent Constraints. To deeply explore disputes and correct subjective biases of agents, this step conducts a three-stage adversarial debate based on structured evidence and introduces a precedent database to provide normative guidance. First, the system constructs an adversarial framework composed of buyer's lawyer agents and seller's lawyer agents, executing the debate sequentially according to the process of opening statements, cross-rebuttals, and closing arguments: In the opening statement stage, both parties clarify their positions and state the facts based on their own evidence and claims; in the cross-rebuttal stage, both parties cross-examine and refute the opponent's arguments, evidentiary loopholes, and logical contradictions; in the closing arguments stage, both parties eliminate invalid arguments, converge valid facts, and submit a ruling request. At the same time, based on the principle of "following precedent," historical precedents highly relevant to the current case are matched from the precedent database through semantic encoding and similarity retrieval. Structured ruling guidelines are injected into the debate process as external explicit constraints, guiding agents to align with platform consensus and transaction norms, and correcting positional biases and cognitive biases. Ultimately, the system generates a complete, logically rigorous, and standardized sequence of six rounds of debate texts, providing high-quality reasoning for the subsequent adjudication module.

[0060] 4. Explainable Adjudication Output Based on Graph Attention Network. To model the temporal dependencies, adversarial relationships, and logical connections between debate stages and achieve a fair and traceable final adjudication, this step employs a fusion mechanism of graph attention network and interactive attention to complete reasoning and judgment. First, the content of the six rounds of debate and the case background information are mapped into a debate graph containing seven functional nodes. Intra-role temporal edges, cross-role adversarial edges, and global anchor edges are set to explicitly construct the debate logical topology. Second, the semantic features of the debate are fused with the agent's identity and stage features through a role-aware encoder to enhance the distinguishability of node representations. Subsequently, a multi-layer graph attention network is used to complete message passing and feature updates between nodes, automatically learning the importance weights of each debate stage to achieve global logical reasoning. Next, using the core claims of the case as query vectors, a focus-driven interactive attention mechanism is used to focus on key debate content, completing the fusion of global pooled features and focus features. Finally, the fused features are fed into a classifier to output the adjudication probability. The winning party is determined based on the probability, and key arguments are extracted based on debate nodes with high attention weights to generate clear, credible, and traceable adjudication reasons, completing the entire intelligent adjudication process.

[0061] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0062] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0063] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0064] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0065] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A social simulation-based method for adjudicating e-commerce disputes, characterized in that, include: Acquire multimodal data of the case to be adjudicated and initialize the environment; the multimodal data includes transaction information, the claims of the disputing parties, and the evidence set corresponding to the claims; the evidence includes text, images, and videos; A proactive iterative evidence collection mechanism is adopted, taking the demands as the guide, iteratively executing the location and perception of key evidence in the evidence, and independently generating a structured evidence fact matrix for both parties in the dispute; A library of precedents is introduced as a normative constraint, and an adjudication guide is created. Based on a structured evidence fact matrix, a three-part multi-agent adversarial debate, including opening statements, cross-rebuttals, and closing arguments, is executed, generating a normatively calibrated sequence of debate texts. The debate text sequence and case background information consisting of transaction information and the claims of both parties are mapped into a debate graph that includes temporal sequence, adversarial relationship and global anchor relationship. Logical reasoning is performed through graph attention network to output a verifiable and explainable ruling result and reasoning.

2. The method for adjudicating e-commerce disputes based on social simulation according to claim 1, characterized in that, The aforementioned proactive iterative evidence collection mechanism, guided by the claims, iteratively locates and perceives key evidence within the evidence, independently generating a structured fact matrix of evidence for both parties in the dispute, specifically including: The proactive iterative evidence collection mechanism operates independently and in parallel from the perspectives of both parties in the dispute, continuing until sufficient clues are available or the maximum number of iterations is reached. Each iteration includes: The key evidence location step is used to select from the evidence set the evidence most relevant to the current point of contention and the claims of both parties in the dispute. The fine-grained evidence perception step is used to directly analyze the images of selected evidence or to analyze the videos of selected evidence frame by frame to extract key clues and form evidence of argumentation. After the iteration is completed, a structured evidence fact matrix for both parties in the dispute is output.

3. The method for adjudicating e-commerce disputes based on social simulation according to claim 2, characterized in that, The key evidence location step is used to select the evidence most relevant to the current point of contention and the claims of both parties from the evidence set, specifically including: For the buyer, in the first During the round of iteration, based on the case background information The evidence includes a collection of texts from the buyer's evidence, from which the most relevant piece of evidence is selected to verify the truthfulness of the buyer's statements. ； in, For the index of the selected evidence, For the first Buyer evidence selected in rounds of iterations. Indicates the first Index of round iteration, Indicates as of the date The wheel is a collection of texts consisting of all the evidence presented by the selected buyers. This refers to the text set consisting of all the text in the buyer's evidence. Indicates that during the first The text set of buyer evidence that was not selected during the round of iteration; The prompts indicate how to extract the points of contention based on case background information and evidence text, and how to select evidence. For the seller, in the first During the round of iteration, based on the case background information The evidence includes a collection of texts from the seller's evidence. From this collection, select the piece of evidence that best reflects the point of contention to verify the seller's statement. ； For the first Seller evidence selected in rounds of iteration; Indicates as of the date A round is a collection of texts consisting of all selected seller evidence. A collection of texts representing all the seller's evidence. Indicates that during the first The set of texts of seller evidence that were not selected during the round of iteration.

4. The method for adjudicating e-commerce disputes based on social simulation according to claim 3, characterized in that, The fine-grained evidence perception step is used to directly analyze the images of selected evidence or to analyze the videos of selected evidence frame by frame to extract key clues and form evidence of argumentation. Specifically, it includes: If the selected evidence is an image, the image is directly analyzed; if the selected evidence is a video, the video is analyzed frame by frame. For the buyer, it is crucial to extract key leads directly related to the dispute. And provide the key clue that can support or weaken the buyer's argument. : ； in, Indicates the first In the round of iterations, key clues from the buyer Buyer's justification The output set is composed of Clue words for intelligent agents to perceive key cues in a fine-grained manner; For the seller, it is crucial to extract key leads directly related to the dispute. And provide the seller's arguments that support or weaken the seller's statement based on this key clue. : ； in, Indicates the first In the round of iterations, key clues from the seller Seller's justification The output set; After the iteration is completed, a structured evidence fact matrix for both parties in the dispute is output, specifically including: By summarizing the evidence perceptions from both sides of the dispute, a structured matrix of evidence facts is generated. ; This represents the evidence perception result obtained by the buyer after completing all iteration rounds. It integrates the key clues and arguments of the buyer from each round of fine-grained evidence perception output from the buyer's perspective; This represents the evidence perception result obtained by the seller after completing all iteration rounds. It integrates the key clues and arguments of the seller from each round of fine-grained evidence perception output from the seller's perspective; This represents the maximum number of iteration rounds.

5. The method for adjudicating e-commerce disputes based on social simulation according to claim 1, characterized in that, The introduction of a library of adjudication precedents as normative constraints and the creation of adjudication guidelines specifically include: Trial Precedent Library , For adjudicated cases, To extract structured adjudication rules from adjudicated cases; through a semantic encoder. Treating adjudicated cases The text and the claims of both parties in the dispute are characterized using a retrieval function. Calculate similarity and return A number of similar cases serve as guidelines for adjudication. : 。 6. The method for adjudicating e-commerce disputes based on social simulation according to claim 1, characterized in that, The structured evidence fact matrix-based, multi-agent adversarial debate, comprising an opening statement, cross-refutation, and closing statement, specifically includes: The three-stage multi-agent adversarial debate is executed sequentially by the buyer's lawyer agent and the seller's lawyer agent: During the opening statement phase, each side declares its arguments based on its perceived evidence, generating the content of its opening statement; In the cross-rebuttal phase, the logical flaws and evidentiary validity of the opposing side's opening statement are refuted and examined. During the closing arguments phase, both parties filter out invalid arguments, retain valid evidence that has been verified through cross-examination, and submit their requests for a ruling.

7. The method for adjudicating e-commerce disputes based on social simulation according to claim 6, characterized in that, During the opening statement phase, each side shall state its arguments based on its perception of the evidence presented, specifically including: First Buyer's Lawyer Intelligent Agent With the first seller's lawyer's intelligent agent Based on their respective perceptions of evidence, each speaker takes turns to present their arguments, extracting supporting factual evidence from a structured evidence matrix to generate their opening statement: ； ； This refers to the opening statement of the first buyer's lawyer's AI agent. This refers to the opening statement of the first seller's lawyer's AI agent. This indicates that the debate generation prompts enable the lawyer agent to generate the debate text based on the given prompts. For case background information, This represents the buyer's perception of complete evidence within the structured evidence fact matrix. This represents the seller's perceived completeness of evidence in the structured evidence fact matrix. Adjudication Guidelines; The cross-rebuttal phase involves refuting and examining the logical flaws and evidentiary validity of the opposing side's opening statements, specifically including: ； ； This represents the cross-rebuttal content of the second buyer's lawyer's AI agent. This represents the cross-rebuttal content of the second seller's lawyer's intelligent agent. For the second buyer's lawyer intelligent agent, For the second seller's lawyer intelligent agent; Both parties filtered out invalid arguments, retained valid evidence verified through cross-examination, and submitted a request for a ruling, specifically including: ； ； This refers to the closing statement of the third-party buyer's legal counsel. This refers to the closing statement of the third-party seller's lawyer agent. This refers to the third-party buyer's lawyer's intelligent agent. This represents the third-party seller's lawyer's intelligent agent; Will Combined into a debate text sequence .

8. The method for adjudicating e-commerce disputes based on social simulation according to claim 7, characterized in that, The mapping of the debate text sequence and case background information consisting of transaction information and the claims of both parties to the dispute into a debate graph that includes temporal sequence, adversarial relationship, and global anchor point relationship specifically includes: Mapping debate text sequences with case background information to create debate graphs. Node set It contains 7 nodes in total. to These are debate nodes, corresponding sequentially to the debate outputs of the two sides in a multi-agent adversarial debate. ; This is a case background information node, used to aggregate transaction information and the demands of both parties in the dispute; edge set It is divided into three categories: intra-role temporal edges, which are used to represent the evolutionary relationship of the arguments of lawyer agents with the same position; cross-role adversarial edges, which are used to explicitly model the rebuttal relationship between the two parties in a dispute; and fully connected anchor edges, which are bidirectionally connected to the case background information nodes and all debate nodes.

9. The method for adjudicating e-commerce disputes based on social simulation according to claim 1, characterized in that, The logical reasoning through graph attention networks, which outputs traceable and explainable adjudication results and reasons, specifically includes: In the debate graph, the debate outputs of each round by the two sides in a multi-agent adversarial debate are mapped as debate nodes; the case background information is mapped as case background information nodes. The debate nodes and case background information nodes are collectively referred to as nodes. Semantic features of each node in the debate graph are extracted and latent vectors representing node identity and stage attributes are fused to obtain node features. Using the case background information node as the global logical anchor point, message passing and feature aggregation are carried out between nodes through a multi-layer graph attention network. During the propagation process, the attention coefficient is used to measure the importance of each node to the final decision. Based on the final attention coefficient of each debate node to the case background information node in the graph attention network, high-influence debate nodes are selected to generate traceable decision reasons. The projected case background information node features are used as queries. Interactive attention calculation is performed on the features of each debate node. The final ruling representation is obtained based on the interactive attention fusion result. The ruling result is output based on the final ruling representation.

10. The method for adjudicating e-commerce disputes based on social simulation according to claim 9, characterized in that, The semantic features of each node in the debate graph are extracted and fused with latent vectors representing node identity and stage attributes to obtain node features, specifically including: Through semantic encoder Extracting semantic features of nodes: ， ； in, Indicates the first The debate text embedding of each debate node, , Text embedding indicating case background information; Background information for the case; For the first The debate output corresponding to each debate node; Through role awareness, perform semantic alignment of roles: ； in, The latent vector representing the identity and stage attributes of each node. For a complete character representation vector, ; Learnable role-aware linear transformation weight matrix; Node features are obtained through feature concatenation and fusion. : ； The process uses case background information nodes as global logical anchors, and employs a multi-layer graph attention network for message passing and feature aggregation among nodes. During propagation, attention coefficients are used to measure the importance of each node to the final ruling. Based on the final attention coefficients of each debate node to the case background information nodes in the graph attention network, high-influence debate nodes are selected, and traceable ruling reasons are generated. Specifically, this includes: A multi-layer graph attention network is used to complete message passing and feature aggregation between nodes: ； in, For the i-th node in the... Features of the layer For activation function, Let be the set of neighboring nodes of the i-th node. For the first The learnable linear transformation matrix of the layer; This is the attention coefficient used to measure the logical importance of the j-th neighboring node to the current i-th node; extract the... The final attention coefficient between the case background information nodes and the debate nodes after the layered graph attention network is used to filter high-impact debate nodes and generate the ruling reasoning. : ； These are prompt words used to generate the reasons for the ruling. Select an operator for sorting the attention coefficient set from largest to smallest and return the node numbers corresponding to the top n attention coefficients; Indicates the process After the layered graph attention network operation, the case background information nodes With the Debate Nodes The final attention coefficient between them; This indicates the number of high-impact debate nodes selected. The process involves using the projected case background information node features as queries, performing interactive attention calculations on the features of each debate node, obtaining the final ruling representation based on the interactive attention fusion result, and outputting the ruling result based on the final ruling representation. Specifically, this includes: The feature matrices of the 7 nodes output by the multi-layer graph attention network are denoted as the logical representation. ,Will Projected to a unified dimension: ， ； in, This represents the global feature matrix after projection of all nodes. This represents the global projection weight matrix, which consists of learnable parameters. This represents the feature vector after individual projection of the case background information node. This represents the case node-specific projection weight matrix, which consists of independent learnable parameters. The features of the case background information nodes after being updated by a multi-layer graph attention network; then... For the query vector, with the global feature matrix The projected feature submatrices corresponding to the 6 debate nodes obtained from the extraction are used as keys and values ​​to perform interactive attention calculation: ； Indicates from the global feature matrix The projected feature submatrices corresponding to the 6 debate nodes are extracted from the data. The query vector representing interactive attention. Keys representing interactive attention The value representing the interaction attention; ； This represents the output features after interactive attention fusion. Indicates the feature projection dimension. This is a matrix transpose operation. This represents the Softmax function; After that Pooling and focus interaction features are concatenated to obtain the final decision representation: ； Represents global aggregated features. This represents the graph pooling function. The final decision obtained after feature cascading is represented. This indicates a feature concatenation operation; The final decision result is output by the classifier. : ； It is a multilayer perceptron.

Images