Multi-agent driven e-commerce dispute resolution method and system

Abstract

The application relates to the technical field of electronic commerce and discloses a multi-agent driven e-commerce dispute adjudication method and system, which comprises the following steps: obtaining multi-modal evidence of a dispute case to be adjudicated, and initializing a review panel and a corresponding social network; individual adjudication thinking chains of the review panel members are used to perform structured reasoning on the dispute case, so that individual adjudication results and reasons are generated; multi-round discussions of the review panel are simulated, in each round, each review panel member adjusts his / her own adjudication according to the adjudication reasons of the neighbors in the last round and the global review summary in the social network; when the proportion of the review panel members supporting a certain adjudication result reaches a consensus threshold or the maximum number of discussion rounds is reached, the discussion is terminated, the final winning side is determined, and the global review summary in the last round is output as the adjudication reason; the application can greatly shorten the adjudication response cycle while ensuring the decision quality, and can ensure the high unification of the adjudication efficiency and precision.

Description

Technical Field

[0001] This invention relates to the field of e-commerce technology, specifically to a multi-agent-driven method and system for adjudicating e-commerce disputes. Background Technology

[0002] The booming development of the digital economy has led to an exponential increase in online transaction disputes. Traditional dispute resolution mechanisms, such as litigation and human customer service, are insufficient to handle the massive demand for dispute resolution due to limited capacity, lengthy processes, and high costs. To address this, e-commerce platforms have widely adopted a jury system: both parties provide multimodal evidence, including chat logs, images, and videos, to support their claims; multiple judges evaluate the cases based on the content and their personal understanding, with the party receiving the most votes winning. However, this mechanism relies heavily on a large number of volunteer judges, resulting in drawbacks such as long recruitment periods (usually several days), low decision-making efficiency, inconsistent professional levels among judges, and susceptibility to subjective biases. It also suffers from poor scalability, becoming a bottleneck restricting the healthy development and service experience of e-commerce platforms. In this context, developing an intelligent dispute resolution system for e-commerce has become an urgent need for maintaining stability in the digital society.

[0003] The development of an intelligent dispute resolution system for e-commerce not only provides innovative experimental paradigms and evaluation benchmarks for interdisciplinary research in computational social science, artificial intelligence, and e-commerce, but also serves as a crucial infrastructure for e-commerce platform governance, digital market supervision, and even the construction of online credit systems. This system has profound application value in the following aspects: 1) Platform governance and rule optimization: The intelligent resolution system can efficiently handle massive numbers of dispute cases. Its resolutions and reasoning can provide real-time, objective data insights for platforms to optimize transaction rules and clarify liability definitions, assisting in the design of fairer and more enforceable community norms, and reducing disputes at the source. 2) Building a trustworthy transaction ecosystem: By providing stable, fair, and explainable automated resolution services, the system helps enhance users' trust in platform dispute resolution mechanisms, boost consumer confidence, and thus promote the establishment of a healthier and more trustworthy digital transaction ecosystem. 3) Social computing and behavioral research: In simulating the cognition, interaction, and consensus-building process of diverse reviewers, the system generates a large amount of structured data on collective decision-making, bias formation, and normative evolution, providing valuable experimental environments and empirical materials for research on fairness, trust, and cooperation in sociology, economics, and cognitive science. 4) Regulatory support: The system can simulate the outcome of dispute resolution under different rules or scenarios, thereby predicting potential risks and providing decision support for achieving more accurate and adaptable digital market governance strategies.

[0004] To achieve intelligent dispute resolution in e-commerce, it is necessary to simulate adjudicators with diverse backgrounds who predict judgments based on multimodal, multi-round chains of evidence from both parties. This typically requires basing decisions on dynamic, informal transaction practices. However, existing dispute resolution tasks focus on the legal field, usually simulating legal professionals predicting judgments based on structured textual evidence. These tasks typically rely on legal provisions as the basis for judgment. Therefore, they differ fundamentally from e-commerce dispute resolution tasks in three dimensions: decision-making mechanism, evidence form, and basis for judgment. To address this, this invention defines and proposes for the first time the new task of "e-commerce dispute resolution," as follows... Figure 1 As shown in the figure. This task aims to simulate the fine-grained perception of multimodal evidence by heterogeneous reviewers, reproduce the decision-making dynamics of the review panel, and realize the automated and professional adjudication of e-commerce transaction disputes.

[0005] In recent years, multi-agent systems have demonstrated their advantages in optimizing complex situations from multiple perspectives by assigning specific roles to each agent and designing interaction, debate, or collaboration mechanisms, providing core support for the intelligent adjudication of disputes. For example, AgentCourt simulates roles such as judges and lawyers, constructing adversarial legal reasoning processes and achieving accurate judgments on structured legal documents. However, such methods are difficult to directly apply to e-commerce dispute scenarios, mainly facing the following two major challenges: First, e-commerce disputes rely on multimodal and fragmented evidence, and the disputing parties engage in multiple rounds of questioning, rebuttal, and explanation. This leads to key clues being hidden in lengthy interactions, failing to clearly provide the causal logic of the dispute. For example, Figure 1 Disputes over charging of floor scrubbers in China require in-depth analysis of video content by models to identify details such as the battery level being only 2% and the indicator light flashing. Existing methods are mostly based on textual reasoning; even with large multimodal models, their one-off, passive perception approach cannot extract fine-grained evidence from redundant content, let alone clarify the causal logic of the disputes. This leads to the empirical phenomenon that existing multimodal models (MLLMs) have lower accuracy than large language models (LLMs) in e-commerce adjudication tasks. Secondly, e-commerce adjudication lacks clear normative basis like legal provisions and relies heavily on dynamic transaction practices. For example, Figure 1 In a scenario where the seller proves the goods were in good condition before shipment, while the buyer insists they discovered the abnormality upon receipt, the jury ultimately supports the buyer based on established trade practices. In such flexible scenarios, existing methods rely on potential biases in the training corpus, tending to favor the seller and failing to maintain the fairness, consistency, and interpretability of the rulings, thus failing to leverage the advantages of group consensus simulation. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a multi-agent-driven method and system for adjudicating e-commerce disputes, aiming to achieve intelligent and fair adjudication of such disputes. At the individual level, the e-commerce dispute adjudication task is broken down into a structured reasoning process encompassing understanding, perception, analysis, and adjudication through an individual adjudication thought chain. This allows reviewers to perceive evidence details with fine granularity, clarify the causal relationship between evidence and the dispute, thereby improving the accuracy and interpretability of the adjudication. At the review panel level, this invention simulates multiple rounds of discussion among multiple reviewers and summarizes the results of each round into a global review summary to correct individual adjudication biases. In a further technical solution, this invention also constructs a precedent library during the review panel's multiple rounds of discussion, serving as explicit adjudication norms to enhance the fairness of e-commerce dispute adjudication. Finally, the review panel adjudicates cases by vote, with the party receiving the most votes from the buyer and seller winning. Furthermore, the global review summary serves as the reasoning for the adjudication, enhancing the persuasiveness of the case's decision.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a multi-agent-driven method for adjudicating e-commerce disputes, comprising:

[0009] Acquire multimodal evidence of dispute cases to be adjudicated, and initialize a jury composed of multiple jury agents with heterogeneous profiles and their corresponding social networks;

[0010] Each reviewer agent performs structured reasoning on dispute cases through an individual adjudication thought chain, generating individual adjudication results and reasons for the adjudication; the individual adjudication thought chain includes extracting the points of contention from the statements of both parties, iteratively locating key clues from multimodal evidence based on the points of contention, conducting adversarial causal analysis on the key clues of both parties, and making a final adjudication by comprehensively considering the points of contention, the key clues of both parties, and the results of the adversarial causal analysis;

[0011] The simulated review panel held multiple rounds of discussions. In each round, each reviewer agent adjusted its own decision based on the previous round's rulings from its neighbors in the social network, the global review summary, and the explicit ruling rules.

[0012] When the proportion of reviewer agents supporting a certain ruling reaches the consensus threshold or the maximum number of discussion rounds is reached, the discussion is terminated, the final ruling of the review panel in the last round of discussion is calculated to determine the final winner, and the global review summary of the last round is output as the ruling reason.

[0013] In one embodiment, the multimodal evidence includes: transaction metadata, evidence items submitted multiple times by the buyer, and evidence items submitted multiple times by the seller; the transaction metadata includes product information and chat logs; the evidence items include text, images, and videos; the text in all buyer evidence items constitutes the buyer's text statement, and the images and videos in the buyer evidence items are used to corroborate the buyer's text statement; the text in all seller evidence items constitutes the seller's text statement, and the images and videos in the seller evidence items are used to corroborate the seller's text statement.

[0014] In one embodiment, extracting the points of contention from the statements of both parties specifically includes:

[0015] The reviewer agent first reviews the transaction metadata. Buyer's written statement Seller's written statement Perform structured analysis; extract the points of contention. This includes a summary of the conflicts between the two parties. The buyer's core demands and the seller's core demands :

[0016] ;

[0017] in, Extract key words for the points of contention for the reviewer agent.

[0018] In one embodiment, the iterative location of key clues from multimodal evidence based on points of contention specifically includes:

[0019] The reviewer agent uses the point of contention as the query, actively selects the evidence items most likely to contain key clues from multimodal evidence, and locates the key clues based on the selected evidence items. This process is repeated until the termination condition is met, specifically including:

[0020] For the buyer, the reviewer agent is based on the points of contention. and buyer's written statement Choose the first Buyer's Evidence for the Wheel To verify the authenticity of the buyer's written statements:

[0021] ;

[0022] Prompt words to guide the reviewer agent in selecting evidence items; An index to the evidence presented by the selected buyer; Indicates as of the date The text set of all selected buyer evidence items;

[0023] The reviewer agent selects the buyer's evidence items. Extract key buyer clues directly related to the dispute. And provide key clues from the buyer to support or weaken the buyer's written statements as evidence of the buyer's arguments. :

[0024] ;

[0025] Indicates key clues from the buyer Buyer's justification Judgment Result The output set is composed of The prompts are used to help the reviewer's intelligent agent perceive key clues; the termination condition for extracting key clues from the buyer is reaching the maximum number of iterations or determining the result. The current key clues are sufficient to verify the buyer's textual statements;

[0026] For the seller, the reviewer agent acts based on the points of contention. and seller's textual statement collection Choose the first Seller's Evidence for the Wheel To verify the authenticity of the seller's written statements:

[0027] ;

[0028] Indicates as of the date The text set of all selected seller evidence items;

[0029] The reviewer agent selects the seller's evidence items. Extract key seller clues directly related to the dispute. and provide the seller's key clues to support or weaken the seller's written statements as evidence. :

[0030] ;

[0031] Indicates key clues from the seller Seller's justification Composition of the judgment result The output set; the termination condition for extracting key seller clues is reaching the maximum number of iterations, or determining the result. The current key clues from the seller are sufficient to verify the seller's written statements;

[0032] Output of the key clue location stage in the individual decision-making thought chain This summarizes all the key clues and arguments selected after the final iteration rounds from both the buyer's and seller's perspectives, and represents them as follows: ; For the final iteration round number of the buyer, This represents the final iteration round number for the seller.

[0033] In one embodiment, the adversarial causal analysis of key clues between the disputing parties specifically includes:

[0034] The reviewer agent identifies contradictions in the statements of the selected buyer and seller based on key clues from both parties. Generate a buyer structured analysis report Seller Structured Analysis Report :

[0035] ;

[0036] The prompt word indicates that the review agent is conducting adversarial causal analysis using a multimodal large model; This represents the output of the individual's decision-making thought process in the adversarial causal analysis stage, by... , as well as composition.

[0037] In one embodiment, the final ruling based on the comprehensive analysis of the points of contention, key clues from both parties, and adversarial causal analysis results specifically includes: comprehensively considering the points of contention... Output of the key clue location stage Output of the adversarial causal analysis stage The k-th reviewer agent Assess the reasonableness of both parties' core demands to arrive at a final ruling. And generate traceable reasons for the ruling. :

[0038] ;

[0039] The prompt word is to guide the review agent to adjudicate disputes using a multimodal large model. This represents the output of the individual's decision-making thought process at the final decision-making stage, by... as well as composition.

[0040] In one embodiment, the simulated review panel engages in multiple rounds of discussion. In each round, each reviewer agent adjusts its own decision based on the previous round's rulings from its neighbors in the social network, a global review summary, and explicit ruling specifications. Specifically, this includes:

[0041] Given a dispute case, the k-th reviewer agent In the During the round of discussions, through social networks Interaction between the two parties generates their own adjudication results. Reasons for the ruling for:

[0042] ;

[0043] in, The individual decision-making thought chain was invoked and combined with... Corresponding portrait Memory module Make a ruling; for On social networks The set of reasons given by the reviewer agent in the previous round of decision received by the receiving entity, the first... Summary of global review in rounds of discussion This aggregates the decision-making reasons of each reviewer agent from the previous round of discussions. :

[0044] ; ;

[0045] Let be the set of directed edges in a social network. Represents the k-th reviewer agent. Focus on the j-th reviewer agent ; This indicates a prompt that suggests the reasoning behind the decisions of all reviewer agents in the previous round of discussion within a multimodal large model aggregation. N represents the total number of reviewer agents. This represents the set union operator.

[0046] In one embodiment, the multi-round deliberation phase of the jury also includes a process for developing fair adjudication standards to explicitly regulate case adjudication, specifically including:

[0047] Record the rulings of real historical cases Extracted into review guidelines To form a review precedent library Ruling Record This includes details of disputed cases and the reasoning behind the jury's rulings; and utilizes a database of precedents to customize the memory modules of the jury agents. :

[0048] For pending dispute cases, semantic encoders are used. By characterizing the transaction metadata and the core demands of both buyers and sellers in dispute cases, key features of the dispute cases awaiting adjudication are obtained; cosine similarity is used to calculate key features and compare them with the review precedent library. Review Guidelines Based on the similarity between them, the most relevant historical rulings were retrieved. and corresponding review guidelines ;in, It includes multiple review rules related to historical ruling records. Use cosine similarity to measure each review rule. With the reviewer's intelligent agent portrait The similarity is compared, and the one with the highest similarity is selected. Each set of review rules is stored in the memory module as the memory of the reviewer's intelligent agent.

[0049] ;

[0050] Indicates the review guidelines being retrieved. In the middle, each review rule Character profile of the reviewer's intelligent agent A similarity ranking function is used to select the most similar individuals. Review rules; This represents the operation of using cosine similarity to measure the similarity between each review rule and the profile of the reviewer agent.

[0051] In one embodiment, the statistical review panel discusses the final round of review panel decisions to determine the final winner, and outputs a summary of the final round of global reviews as the reasoning for the decision, specifically including:

[0052] The discussion will end after round T, and the number of votes cast by the jury in support of the buyer in round T will be tallied. and the number of votes in favor of the seller :

[0053] ; ;

[0054] N represents the total number of reviewer agents; Represents the k-th reviewer agent. In the During the round of discussions, through social networks The self-determination generated by the interaction between the two parties; Represents the reviewer's intelligent agent In the After rounds of discussions, the ruling supported the buyer; if , indicating the reviewer's intelligent agent In the After rounds of discussion, the ruling supported the seller. This indicates an indicator function; the value is 1 if the condition within the parentheses is met, and 0 if the condition within the parentheses is not met.

[0055] ;

[0056] The reviewer agent supports the party that wins the case more often. This indicates the jury's final decision. This indicates that the buyer has won. The seller wins, and a summary of the global review from the T-round discussion will be provided. As the reason for the final ruling.

[0057] In a second aspect, the present invention provides a computer system including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method of any embodiment of the first aspect.

[0058] Compared with the prior art, the beneficial technical effects of the present invention are:

[0059] This invention fills a gap in e-commerce dispute adjudication: the "e-commerce dispute adjudication" task aims to simulate the decision-making process of a review panel by understanding the multimodal temporal evidence presented by both parties in a fine-grained manner, thereby achieving robust, fair, and intelligent dispute adjudication in specific e-commerce scenarios, replacing the time-consuming human review panel. The complexity of the task reveals its development potential and research value in areas such as dispute adjudication, fine-grained content perception, user intent reasoning, and multi-agent simulation. Experimental data show that when handling real multimodal dispute tasks, this invention improves the accuracy of judgments by 9.48% and 9.38% respectively compared to the best-performing benchmark Large Language Models (LLMs) and Multimodal Models (MLLMs); and compared to existing court simulators, it also achieves a significant increase in accuracy of 6.19%.

[0060] This invention innovatively proposes the Individual Adjudication Thinking Chain (IV-CoT), which decomposes the task of adjudicating e-commerce disputes into a four-stage structured reasoning process consisting of dispute focus analysis, key clue location, adversarial causal analysis, and interpretable judgment. This invention represents a technological leap from "full content encoding" to "key clue focusing," accurately clarifying the causal logical relationship between "evidence clues" and "dispute focus." This structured reasoning paradigm not only significantly improves the system's ability to process fragmented and adversarial evidence in complex e-commerce scenarios but also ensures that the final adjudication is based on rigorous logical deduction rather than simple statistical prediction, endowing the intelligent adjudication process with extremely high professional depth and logical persuasiveness.

[0061] This invention ensures fairness and consistency in adjudication results in dynamic transaction environments: E-commerce dispute adjudication relies on dynamic transaction rules and lacks clearly defined trial rules similar to legal provisions. The invention constructs a panel consensus adjudication module based on precedent in its technical solution. First, it innovatively introduces the principle of "following precedent," providing explicit normative guidance for informal and dynamic transaction dispute scenarios through a database of adjudication precedents. Coupled with a panel discussion mechanism, this invention simulates a multi-round opinion discussion and exchange process with characteristics of social network dynamics, effectively correcting the inherent biases of a single model due to training data or specific preferences by utilizing collective wisdom. This multi-agent collaborative model ensures that the final adjudication result is based on collective consensus after deliberation from multiple perspectives, thereby establishing fair and consistent adjudication standards in complex and flexible dispute scenarios and effectively avoiding the adjudication risks caused by individual review biases.

[0062] In summary, this invention addresses the massive demand for e-commerce disputes by significantly shortening the adjudication response cycle while ensuring decision-making quality, thus achieving a high degree of consistency between adjudication efficiency and accuracy. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the intelligent adjudication task for e-commerce disputes.

[0064] Figure 2 This is a schematic diagram of the method flow of the present invention.

[0065] Figure 3 This is a schematic diagram of the framework structure proposed in this invention. Detailed Implementation

[0066] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0067] like Figure 2 As shown, this invention provides a multi-agent-driven e-commerce dispute adjudication method, comprising the following steps:

[0068] S1: Obtain multimodal evidence of the dispute case to be adjudicated, and initialize a review panel consisting of multiple reviewer agents with heterogeneous profiles and the corresponding social network.

[0069] S2, each reviewer agent performs structured reasoning on the dispute case through an individual adjudication thought chain, generating an individual adjudication result and reasoning for the adjudication; the individual adjudication thought chain includes extracting the points of contention from the statements of both parties, iteratively locating key clues from multimodal evidence based on the points of contention, conducting adversarial causal analysis on the key clues of both parties, and making a final adjudication by comprehensively considering the points of contention, the key clues of both parties, and the adversarial causal analysis results;

[0070] S3 simulates multiple rounds of discussion by the review panel. In each round, each reviewer agent adjusts its own decision based on the previous round's decision reasons from its neighbors in the social network, the global review summary, and the explicit decision rules.

[0071] S4. When the proportion of reviewer agents supporting a certain ruling reaches the consensus threshold or the maximum number of discussion rounds is reached, the discussion is terminated, the final ruling of the review panel in the last round of discussion is calculated to determine the final winner, and the global review summary of the last round is output as the ruling reason.

[0072] The technical solution of the present invention will be described in detail below in several parts.

[0073] 1. Task request and environment initialization.

[0074] During the initialization phase of system startup, the system first acquires full multimodal evidence data for the dispute cases to be adjudicated, including transaction logs submitted by both parties, their claims, and corresponding image and video evidence. Simultaneously, the intelligent adjudication system initializes the user profiles of the review panel members and the social network topology of the review panel. Specifically, the system pre-defines user profiles with heterogeneous background knowledge for the reviewer agents and initializes the review panel's social network, defining how reviewer agents interact and influence each other's decisions; for example, how reviewer agents reference and influence the decisions of other reviewer agents.

[0075] In one embodiment, multimodal evidence in a dispute case to be adjudicated. It contains transaction metadata. For example, product information, chat history, etc., also include Each buyer's evidence item ,as well as Each seller's evidence item ; For the index of the i-th buyer's evidence item, This is the index of the j-th seller's evidence item; each evidence item contains text, images, and videos.

[0076] .

[0077] This invention is modeled as a multi-agent system, represented as a directed social network. :

[0078] ;

[0079] in, Indicates by A judging panel composed of intelligent agents of various judges, with directed edges Represents the k-th reviewer agent. Focus on the j-th reviewer agent .

[0080] In one embodiment, all N reviewer agents in this invention can be constructed using the Gemini-2.5-Flash-Lite-Nothinking multimodal large model.

[0081] 2. Each reviewer agent in the review panel generates a ruling based on the case information.

[0082] The adjudication process is essentially a decision-making process that uses textual claims as an index to continuously locate key clues among redundant evidence in order to verify the statements of both parties. However, existing dispute adjudication methods typically rely on the content-awareness capabilities of multimodal models (MLLMs) and input all dispute information at once. This coarse-grained understanding makes it difficult for the reviewer agent to achieve complex causal reasoning of "evidence-dispute focus" in the e-commerce field. To address this, this invention proposes Individual Adjudication Thinking Chain (IV-CoT), which decomposes the e-commerce dispute adjudication task into a four-stage structured reasoning process, enabling the reviewer agent to deepen its understanding from "full coarse encoding" to "key evidence focusing," and from "appearance description" to "causal inference." Specifically, each stage is as follows: 1) Extracting the dispute focus: The reviewer agent first parses the core dispute focus from the textual statements of the buyer and seller, identifying the key claims and positions of both parties in the dispute. 2) Locating key clues: Based on the dispute focus, this invention designs a "selection-perception" proactive iterative evidence collection mechanism, transforming the one-time multimodal content understanding commonly used in existing methods into an iterative key clue locating process based on dispute claims. 3) Adversarial Analysis: By comparing the evidence from both sides, the reviewer agent identifies contradictions in the statements and makes causal inferences, further clarifying the causal relationship between the evidence and the points of contention, revealing the accuracy and contradictions in the statements made by both sides. 4) Final Ruling: Combining the above analysis, the reviewer agent makes a fair and interpretable ruling, providing clear reasons for the ruling, ensuring transparency and credibility in the final judgment. Through these four stages, the Individual Ruling Thought Chain (IV-CoT) helps the reviewer agent deeply understand the core of the dispute and extract effective clues from fragmented multimodal evidence, accurately clarifying the causal relationship between the evidence and the points of contention, thereby achieving a more intelligent and reliable adjudication process.

[0083] In one embodiment, the stage of extracting the point of contention in the individual decision-making thought process includes: the reviewer agent. First, analyze the transaction metadata. The textual statements of both parties were analyzed in a structured manner to identify the key issues in dispute. This includes a summary of the conflicts between the two parties. The buyer's core demands and the seller's core demands :

[0084] ;

[0085] in, Extract key words indicating points of contention for the reviewer agent. This indicates the buyer's written statement. The text of the i-th buyer's evidence item; , The text of the j-th seller's evidence item.

[0086] In one embodiment, the key clue locating stage of the individual adjudication thought chain includes: To establish a clear correspondence between key clues and the focus of the dispute, this invention designs a proactive iterative evidence collection mechanism of "selection-perception," transforming the existing method of understanding all multimodal content at once into an iterative key clue locating process based on the dispute's claims. To prevent prior interference from mixed-source evidence, this process is conducted independently by both parties to the dispute, ensuring a fair assessment of the reasonableness of both parties' claims.

[0087] Buyer's Taking the round of evidence collection as an example:

[0088] 1) First, the reviewer agent proactively locates the evidence items most likely to contain key clues within a large-scale evidence pool, based on the textual clues of the dispute. This is done according to the focus of the dispute. and buyer's written statement The reviewer agent actively selects one item of buyer evidence. To verify the authenticity of the buyer's written statements:

[0089] ;

[0090] in, Prompt keywords for selecting key evidence in multimodal large-scale models. For the index of the selected evidence, Indicates as of the date The text set of all selected buyer evidence items.

[0091] 2) The reviewer agent reviews the selected buyer's evidence items. Perform fine-grained perception and multimodal large model prompts. Extract key buyer clues directly related to the dispute from this. And provide key clues from the buyer to support or weaken the buyer's written statements as evidence of the buyer's arguments. :

[0092] ;

[0093] The "selection-perception" proactive iterative evidence collection mechanism is executed iteratively, and its termination condition is reaching the maximum number of iterations. or the judgment result The current key clues from the buyer are sufficient to verify the buyer's written statements. Furthermore, the mechanism for locating the seller's evidence and key clues is logically symmetrical with that of the buyer.

[0094] In the stage of locating key clues in an individual's thought process, the final output is... This summarizes all the key clues and arguments selected after the final iteration rounds from both the buyer's and seller's perspectives, and represents them as follows: ; For the final iteration round number of the buyer, This provides a reliable clue for the seller's final iteration rounds and subsequent stages.

[0095] In one embodiment, the adversarial analysis phase of the individual adjudication thought chain includes: in order to detect false statements by both parties, the reasoning objective of this phase shifts from "clue matching" to "dispute causal analysis." After obtaining key adversarial clues from both sides, the reviewer agent delves into the detailed context of the case, identifying contradictions in the evidence presented by both parties. This report summarizes the strength and weaknesses of the evidence presented by both parties, and presents a structured analysis report for the buyer. Seller Structured Analysis Report By using adversarial analysis to deduce the root causes of disputes, in-depth evidence is provided for adjudication.

[0096] .

[0097] The prompt word indicates that the review agent is conducting adversarial causal analysis using a multimodal large model; This represents the output of an individual's thought process chain in the adversarial causal analysis phase, by... , as well as composition.

[0098] In one embodiment, the final decision-making stage of the individual decision-making thought chain includes: synthesizing the structured outputs of the previous three stages, the reviewer agent... Assess the reasonableness of both parties' claims and reach a final ruling. And generate traceable reasons for the ruling. The following are the judging opinions from the subsequent mock judging module:

[0099] .

[0100] The prompt word is to guide the review agent to adjudicate disputes using a multimodal large model. This represents the output of the individual's thought process chain at the final decision-making stage, by... as well as composition.

[0101] 3. Jury simulation.

[0102] To establish standardized and fair adjudication criteria, this invention, based on the precedent principle of case law, retrieves historical real-world cases from a pre-defined review precedent database that highly match the characteristics of the current dispute to be adjudicated. The extracted structured review norms are then injected as explicit constraints into the memory modules of each heterogeneous reviewer agent, thereby establishing a unified and traceable normative benchmark for subsequent decisions. Subsequently, by constructing a social network topology, reviewer agents with different cognitive backgrounds are organized to conduct multiple rounds of opinion exchange. Each reviewer agent, based on independent evaluation of evidence, refers in real-time to the adjudication reasons of its social neighbors and adjusts its own cognition according to the global review summary, thus simulating the process of opinion dissemination and correction in social dynamics. Finally, a group decision aggregation and summary process is executed. The system deeply summarizes the individual adjudication tendencies and logical reasons of all reviewer agents, ultimately generating an adjudication summary reflecting the collective will. This summary integrates diverse adjudication perspectives to correct cognitive biases of the reviewer agents. After each round of simulated review panel discussion, the opinions of each reviewer agent in the review panel become visible to their neighboring reviewer agents, ensuring data flow and state updates during the simulation process. This allows their neighboring reviewer agents to proceed to the next round of decision-making based on social relationships, global review summaries, and their current decision.

[0103] In one embodiment, to mitigate individual reviewer bias during the jury discussion phase, the present invention designs a system based on... The review panel mechanism, involving rounds of discussion, simulates the interactive decision-making among reviewer agents. Specifically, given a dispute case, the reviewer agents... In the During the round of discussions, through social networks Interaction, generating its own judgment Reasons for the ruling :

[0104] ;

[0105] Among them, the reviewer intelligent agent This involves applying the individual adjudication thought process described earlier (i.e., the stages of extracting the focus of the dispute, locating key clues, analyzing the adversarial process, and making the final adjudication), and combining this with a character profile. Memory module A ruling will be made. Furthermore... For the reviewer's intelligent agent On social networks The global review summary is a collection of the previous rulings made by the reviewer agents they are interested in. This aggregates the decision-making reasons of each reviewer agent from the previous round. :

[0106] .

[0107] This refers to prompts that aggregate the reasoning behind the decisions of all reviewer agents in the previous round of discussion within a multimodal large model. This mechanism simulates the social dynamics of opinion propagation, allowing each reviewer agent to consider their social neighbors while thinking independently. Reasons for the ruling Adjust its own ruling. Meanwhile, a global review summary. It can guide individuals who deviate from the mainstream to re-examine their own judgments.

[0108] In one embodiment, a key issue during the jury discussion phase is how to establish fair adjudication standards. Inspired by the precedent principle in case law, e-commerce dispute adjudication should also follow recognized adjudication norms established in historical rulings. Therefore, this invention designs a review precedent library. This review precedent database records the rulings of real historical cases. (i.e., precedents) are distilled into explicit, structured review guidelines. To standardize the review results of the reviewer agent, this invention uses a review precedent library to customize the reviewer agent. memory module This provides an interpretable basis for decision-making. Specifically, it addresses multimodal evidence in cases pending adjudication. This invention utilizes a semantic encoder This involves characterizing the transaction metadata and core claims of both buyers and sellers in disputed cases, and calculating the key features of cases awaiting adjudication and comparing them with those of real review cases. Based on the similarity between them, the most relevant historical rulings were retrieved. and its corresponding review guidelines :

[0109] .

[0110] in, It includes multiple review rules related to historical ruling records. , This indicates the use of cosine similarity to calculate key features and a review precedent library. Review Guidelines Operations on inter-relationship similarity. Measuring these review rules. With the reviewer's intelligent agent portrait Similarity, and select from them The rules serve as the memory of the reviewer's intelligent agent, thereby regulating the reviewer's intelligent agent's fair judgment.

[0111] .

[0112] This represents the operation of using cosine similarity to measure the similarity between each review rule and the profile of the reviewer agent; Indicates the review guidelines being retrieved. In the middle, each review rule Character profile of the reviewer's intelligent agent A similarity ranking function is used to select the most similar individuals. Review rules.

[0113] 4. Jury decision.

[0114] The judging panel discussion phase continues iteratively until the predetermined time step is completed. To improve adjudication efficiency, the system monitors the distribution of opinions within the review panel in real time. When the proportion of reviewer agents supporting a particular ruling exceeds a preset consensus threshold, the system determines that a group consensus has been reached and executes an early termination strategy, ending the iteration prematurely. After the discussion ends, the system proceeds to the final ruling result statistics and output stage. The intelligent adjudication system determines the final winner of the case based on the majority rule by counting the total number of votes supporting the buyer and seller in the final round of the review panel. Simultaneously, the system extracts the global review summary from the final round as the final ruling reasoning.

[0115] In one embodiment, during the jury decision-making stage, the invention simulates... The simulation involves interactive discussions within the review panel. As the number of rounds increases, the panel discussion mechanism encourages the opinions of the reviewer agents to converge towards a robust group consensus. Finally, this invention tallies the number of votes cast by the final review panel in support of the buyer. and the number of votes in favor of the seller :

[0116] ;

[0117] The reviewer agent supports the majority of cases to achieve victory, while also providing a global review summary. As the reason for the final ruling.

[0118] .

[0119] Example:

[0120] This embodiment simulates the process of a review panel composed of 17 reviewer agents in making individual decisions, discussing group opinions, and finally adjudicating disputes. Figure 3The overall architecture diagram of this invention includes four parts: dispute adjudication request and environment initialization, adjudication by the reviewer intelligent agent based on Individual Adjudication Thinking Chain (IV-CoT), review panel simulation, and review panel consensus adjudication.

[0121] Step 1: Dispute Adjudication Request and Environment Initialization. Upon receiving an e-commerce dispute adjudication request, this invention first initializes the user profiles and social network topology of the review panel. The system assigns a basic profile to each reviewer agent in the review panel, including name, gender, age, and occupation. Age is truncated from a normal distribution to more accurately reflect population structure. Occupation and gender are allocated according to a preset ratio. Reviewer agents are assigned 3 to 5 potential interest areas to enhance the diversity of their behavior. Personality traits are encoded based on the Big Five personality model, encompassing openness, conscientiousness, extraversion, agreeableness, and emotional stability, influencing their viewpoint updates and behavioral performance. Agents also possess a personal memory bank storing past adjudication precedents, while an environmental memory bank records neighbor behavior and global review summaries to correct individual adjudication biases. Furthermore, the intelligent adjudication system initializes the social network topology using a directed graph. This indicates the relationship of interest among different judges within the judging panel, and the judges' intelligent agents at different time steps. Through social networks Receive the ruling from the reviewer agent it is concerned with.

[0122] Step 2: Reviewer Agent Adjudication Based on Individual Adjudication Thinking Chain (IV-CoT). Each reviewer agent in the review panel initiates an individual adjudication thinking chain processing flow, designed to deepen the understanding of the dispute case, based on the case information. This flow overcomes the shortcomings of existing technologies where large-scale multimodal models coarsely encode all information at once. Instead, it adopts a refined decision-making model that uses textual claims as an index to continuously locate key clues among redundant evidence, and then clarifies the causal relationship between the evidence and the dispute, thereby improving the accuracy and interpretability of the dispute adjudication. First, the reviewer agent executes the dispute focus extraction logic, parsing the core points of contention and refining the claims of both parties from their textual statements, thus establishing precise query anchors for subsequent perception actions. Subsequently, the system enters an active iterative evidence collection phase based on a "selection-perception" mechanism. The reviewer agent does not blindly read evidence, but rather, guided by the parsed focus, actively selects the multimodal evidence most likely to contain key clues from a large-scale evidence pool, and performs fine-grained perception to locate visual or logical key clues. This process is conducted independently for both the buyer and seller, ensuring the fairness and symmetry of the judgment. After obtaining key clues, the reviewer agent enters the adversarial analysis phase. Its reasoning objective evolves from simple clue matching to in-depth causal logic analysis. By comparing the evidence from both sides, it identifies logical contradictions and generates a structured analysis report to reveal the root cause of the dispute and summarize the strengths and weaknesses of both sides' evidence. Finally, combining the structured outputs from the preceding stages, the reviewer agent makes a highly interpretable preliminary individual ruling and generates detailed reasoning for the ruling as input for subsequent social interactions.

[0123] Step 3: Jury Simulation. To establish fair adjudication standards in e-commerce scenarios lacking formal, explicit adjudication norms, the system first introduces a normative guidance mechanism based on the principle of "following precedent" during the jury simulation process. The system uses a semantic encoder to represent the characteristics of the current case and retrieves the most relevant historical real cases and their corresponding structured review guidelines from a pre-set review precedent library. By calculating the matching degree between the precedent rules in the review guidelines and the profile of the reviewer agent, the system customizes an explicit memory module for each reviewer agent, thereby transforming the precedent rules in historical cases into individual decision constraints. Based on this, the jury enters an interactive decision-making period based on multi-round discussions. Each reviewer agent receives the previous round's adjudication reasons from its followed neighbors in real time on the social network and refers to the global review summary aggregated from all the adjudication reasons of the previous jury members. This mechanism simulates the dynamics of opinion propagation in real society, enabling reviewer agents to maintain independent thinking while adjusting their own positions according to the logic of their social neighbors and the macro-inclination of the group, correcting individual cognitive biases (the specific adjudication process of each reviewer agent is determined by the individual adjudication thought chain).

[0124] Step 4: Panel Consensus Decision. After each simulation round, the system updates the opinions of the reviewer agents based on the social network, ensuring that the opinions of the reviewer agents are known to other reviewer agents who follow them. Simultaneously, the system aggregates the decision reasons of each reviewer agent into a global review summary. As the number of simulation rounds increases, the panel mechanism encourages the opinions of the reviewer agents to converge, at which point the system executes consensus determination and final output logic. To ensure the quality of the decision while improving processing efficiency, the system monitors the distribution of opinions within the panel in real time. Once the proportion of reviewer agents supporting a certain decision direction exceeds a preset consensus threshold... If the value is 0.8, the early stop strategy is immediately triggered, terminating the review panel discussion ahead of schedule; otherwise, iteration continues until the maximum time step is reached. After the discussion ends, the system counts the total number of votes in favor of the buyer and seller in the final review panel, determines the final winner based on the majority rule, and presents the final global review summary, which embodies the logic of multiple rounds of review panel discussions, as the reason for the final ruling to enhance the social credibility of the judgment.

[0125] 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.

[0126] 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.

[0127] In one embodiment, a computer system is provided, which may be a server. The computer system includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data used in the methods described above. The network interface communicates with external terminals via a network connection. The computer program is executed by the processor to implement the methods described above.

[0128] 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.

[0129] 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.

[0130] 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 multi-agent-driven method for adjudicating e-commerce disputes, characterized in that, include: Acquire multimodal evidence of dispute cases to be adjudicated, and initialize a jury composed of multiple jury agents with heterogeneous profiles and their corresponding social networks; Each reviewer agent performs structured reasoning on dispute cases through an individual adjudication thought chain, generating individual adjudication results and reasons for the adjudication; the individual adjudication thought chain includes extracting the points of contention from the statements of both parties, iteratively locating key clues from multimodal evidence based on the points of contention, conducting adversarial causal analysis on the key clues of both parties, and making a final adjudication by comprehensively considering the points of contention, the key clues of both parties, and the results of the adversarial causal analysis; The simulated review panel held multiple rounds of discussions. In each round, each reviewer agent adjusted its own decision based on the previous round's rulings from its neighbors in the social network, the global review summary, and the explicit ruling rules. When the proportion of reviewer agents supporting a certain ruling reaches the consensus threshold or the maximum number of discussion rounds, the discussion is terminated, the final ruling of the review panel in the last round of discussion is calculated to determine the final winner, and the global review summary of the last round is output as the ruling reason. The multimodal evidence includes: transaction metadata, evidence items submitted multiple times by the buyer, and evidence items submitted multiple times by the seller; the transaction metadata includes product information and chat logs; the evidence items include text, images, and videos; the text in all buyer evidence items constitutes the buyer's text statement, and the images and videos in the buyer evidence items are used to corroborate the buyer's text statement; the text in all seller evidence items constitutes the seller's text statement, and the images and videos in the seller evidence items are used to corroborate the seller's text statement. The extraction of points of contention from the statements of both parties specifically includes: The reviewer agent first reviews the transaction metadata. Buyer's written statement Seller's written statement Perform structured analysis; extract the points of contention. This includes a summary of the conflicts between the two parties. The buyer's core demands and the seller's core demands : ； in, Extract key words indicating points of contention for the reviewer agent; The iterative location of key clues from multimodal evidence based on points of contention specifically includes: The reviewer agent uses the point of contention as the query, actively selects the evidence items most likely to contain key clues from multimodal evidence, and locates the key clues based on the selected evidence items. This process is repeated until the termination condition is met, specifically including: For the buyer, the reviewer agent is based on the points of contention. and buyer's written statement Choose the first Buyer's Evidence for the Wheel To verify the authenticity of the buyer's written statements: ； Prompt words to guide the reviewer agent in selecting evidence items; An index to the evidence presented by the selected buyer; Indicates as of the date The text set of all selected buyer evidence items; The reviewer agent selects the buyer's evidence items. Extract key buyer clues directly related to the dispute. And provide key clues from the buyer to support or weaken the buyer's written statements as evidence of the buyer's arguments. : ； Indicates key clues from the buyer Buyer's justification Judgment Result The output set is composed of The prompts are used to help the reviewer's intelligent agent perceive key clues; the termination condition for extracting key clues from the buyer is reaching the maximum number of iterations or determining the result. The current key clues are sufficient to verify the buyer's textual statements; For the seller, the reviewer agent acts based on the points of contention. and seller's textual statement collection Choose the first Seller's Evidence for the Wheel To verify the authenticity of the seller's written statements: ； Indicates as of the date The text set of all selected seller evidence items; The reviewer agent selects the seller's evidence items. Extract key seller clues directly related to the dispute. and provide the seller's key clues to support or weaken the seller's written statements as evidence. : ； Indicates key clues from the seller Seller's justification Composition of the judgment result The output set; the termination condition for extracting key seller clues is reaching the maximum number of iterations, or determining the result. The current key clues from the seller are sufficient to verify the seller's written statements; Output of the key clue location stage in the individual decision-making thought chain This summarizes all the key clues and arguments selected after the final iteration rounds from both the buyer's and seller's perspectives, and represents them as follows: ; For the final iteration round number of the buyer, This represents the final iteration round number for the seller. The review panel's multi-round discussion phase also includes the process of developing fair adjudication standards to explicitly regulate case adjudication, specifically including: Record the rulings of real historical cases Extracted into review guidelines To form a review precedent library Ruling Record This includes details of disputed cases and the reasoning behind the jury's rulings; and utilizes a database of precedents to customize the memory modules of the jury agents. : For pending dispute cases, semantic encoders are used. By characterizing the transaction metadata and the core demands of both buyers and sellers in dispute cases, key features of the dispute cases awaiting adjudication are obtained; cosine similarity is used to calculate key features and compare them with the review precedent library. Review Guidelines Based on the similarity between them, the most relevant historical rulings were retrieved. and corresponding review guidelines ;in, It includes multiple review rules related to historical ruling records. Use cosine similarity to measure each review rule. With the reviewer's intelligent agent portrait The similarity is compared, and the one with the highest similarity is selected. Each set of review rules is stored in the memory module as the memory of the reviewer's intelligent agent. ； Indicates the review guidelines being retrieved. In the middle, each review rule Character profile of the reviewer's intelligent agent A similarity ranking function is used to select the most similar individuals. Review rules; This represents the operation of using cosine similarity to measure the similarity between each review rule and the profile of the reviewer agent.

2. The multi-agent-driven e-commerce dispute adjudication method according to claim 1, characterized in that, The aforementioned adversarial causal analysis of key clues from both sides of the dispute specifically includes: The reviewer agent identifies contradictions in the statements of the selected buyer and seller based on key clues from both parties. Generate a buyer structured analysis report Seller Structured Analysis Report : ； The prompt word indicates that the review agent is conducting adversarial causal analysis using a multimodal large model; This represents the output of the individual's decision-making thought process in the adversarial causal analysis stage, by... , as well as composition.

3. The multi-agent-driven e-commerce dispute adjudication method according to claim 2, characterized in that, The final ruling is made based on a comprehensive analysis of the points of contention, key clues from both parties, and adversarial causal relationships. Specifically, this includes: comprehensively considering the points of contention... Output of the key clue location stage Output of the adversarial causal analysis stage The k-th reviewer agent Assess the reasonableness of both parties' core demands to arrive at a final ruling. And generate traceable reasons for the ruling. : ； The prompt word is to guide the review agent to adjudicate disputes using a multimodal large model. This represents the output of the individual's decision-making thought process at the final decision-making stage, by... as well as composition.

4. The multi-agent driven e-commerce dispute adjudication method according to claim 3, characterized in that, The simulated review panel engages in multiple rounds of discussion. In each round, each reviewer agent adjusts its own decision based on the previous round's rulings from its neighbors in the social network, the global review summary, and explicit ruling rules. Specifically, this includes: Given a dispute case, the k-th reviewer agent In the During the round of discussions, through social networks Interaction between the two parties generates their own adjudication results. Reasons for the ruling for: ； in, The individual decision-making thought chain was invoked and combined with... Corresponding portrait Memory module Make a ruling; for On social networks The set of reasons given by the reviewer agent in the previous round of decision received by the receiving entity, the first... Summary of global review in rounds of discussion This aggregates the decision-making reasons of each reviewer agent from the previous round of discussions. : ； ； Let be the set of directed edges in a social network. Represents the k-th reviewer agent. Focus on the j-th reviewer agent ; This indicates a prompt that suggests the reasoning behind the decisions of all reviewer agents in the previous round of discussion within a multimodal large model aggregation. N represents the total number of reviewer agents. This represents the set union operator.

5. The multi-agent-driven e-commerce dispute adjudication method according to claim 1, characterized in that, The statistical review panel discusses the final round of review panel decisions to determine the final winner, and outputs a summary of the final round of global reviews as the reasoning for the decision, specifically including: The discussion will end after round T, and the number of votes cast by the jury in support of the buyer in round T will be tallied. and the number of votes in favor of the seller : ； ； N represents the total number of reviewer agents; Represents the k-th reviewer agent. In the During the round of discussions, through social networks The self-determination generated by the interaction between the two parties; Represents the reviewer's intelligent agent In the After rounds of discussions, the ruling supported the buyer; if , indicating the reviewer's intelligent agent In the After rounds of discussion, the ruling supported the seller. This indicates an indicator function; the value is 1 if the condition within the parentheses is met, and 0 if the condition within the parentheses is not met. ； The reviewer agent supports the party that wins the case more often. This indicates the jury's final decision. This indicates that the buyer has won. The seller has won, and a summary of the global review from the Tth round of discussions will be provided. As the reason for the final ruling.

6. A computer system comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

Images