Multi-agent based multi-document fact consistency conflict resolution method and system

Abstract

The application discloses a kind of multi-document fact consistency conflict resolution method and system based on multi-agent, which comprises: obtaining document set from multiple information sources and generating document semantic representation;Based on the preset target fact, at least two fact expressions around the target fact are identified;For each fact expression, a fact search agent is constructed respectively, and evidence retrieval is independently executed with the fact expression as the query condition to generate the corresponding evidence document subset;Extract the source information, publication time and semantic similarity information of the evidence document subset and construct the corresponding relationship;Each fact expression and its evidence document subset is input into the judgment agent, and the credibility is evaluated based on the source authority, time consistency and semantic support degree, the relationship type between fact expressions is determined, and the final consistency conclusion and the corresponding evidence document set are output.The application improves the stability, interpretability and processing efficiency of multi-document fact consistency judgment.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and system for resolving multi-document fact consistency conflicts based on multiple agents, applicable to application scenarios involving consistency analysis and conflict determination of text data from multiple information sources. Background Technology

[0002] Multi-document factual consistency research aims to identify, verify, and integrate factual information from multiple heterogeneous sources to ensure that system outputs are consistent with real-world semantics. This issue is of great significance in scenarios such as news generation, intelligence analysis, scientific research reviews, and legal text understanding.

[0003] In document-based fact verification and information integration scenarios, different information sources often provide textual content with differences in time, expression, or conclusion regarding the same target fact. Existing technologies typically employ a single model or rule-based approach to uniformly analyze multiple documents. However, such methods are prone to instability in reasoning or conflicting conclusions when faced with heterogeneous sources, inconsistent publication times, or significant differences in textual expression.

[0004] Furthermore, while some multi-agent solutions introduce multiple processing units, they fail to differentiate and bind different factual expressions, leading to mutual interference between different factual judgments and affecting the reliability of the final consensus conclusion. Therefore, it is necessary to provide a technical solution that can use factual expressions as processing units, retrieve evidence in parallel, and centrally determine consistency relationships. Summary of the Invention

[0005] This invention provides a method and system for resolving multi-document fact consistency conflicts based on multi-agent systems. It can improve the stability and consistency of multi-document fact judgment at the computer system level by distinguishing and processing different fact expressions and performing evidence retrieval and evaluation separately.

[0006] The technical solution for implementing the present invention is as follows: A multi-agent, multi-document fact consistency conflict resolution method, executed by a computer device, includes: Obtain a collection of documents from multiple information sources, where each document is associated with at least a document source identifier and publication time information; The document set is segmented into sentences, and a semantic representation of the documents is generated using a semantic encoding model; Based on the preset target facts, at least two factual expressions surrounding the target facts are identified from the semantic representation of the document; For each factual statement, a factual search agent is constructed to independently perform evidence retrieval using that factual statement as the query condition, generating a corresponding subset of evidence documents; Each factual statement and its corresponding subset of evidence documents are input into the evaluation agent, which evaluates the credibility of each factual statement based at least on the authority of the document source, time consistency, and semantic support. Based on the credibility assessment results of each fact statement, the relationship type between fact statements is determined, and a final consistency conclusion and corresponding evidence document set for the target fact are output.

[0007] Furthermore, the identification of factual expressions from the semantic representation of the document is achieved through semantic clustering or matching processing, wherein the cosine similarity between vector representations is used as a semantic similarity metric.

[0008] Furthermore, when the fact-searching agent performs evidence retrieval, it scores the relevance between the retrieved documents and the factual statements. The score is a weighted comprehensive evaluation based on temporal similarity, source authority, and semantic similarity.

[0009] Furthermore, when the judging agent evaluates the credibility of each factual expression, it calculates the credibility score by weighting and comprehensively calculating authority, temporal consistency, and semantic support.

[0010] Furthermore, the fact-searching agents operate independently of each other, performing evidence retrieval in parallel based on the bound factual expressions, in order to reduce mutual interference between different factual judgments.

[0011] Furthermore, the judging agent is a centralized agent used to comprehensively evaluate each factual expression and its corresponding subset of evidence documents in order to improve the stability of consistency judgments.

[0012] A multi-agent, multi-document fact consistency conflict resolution system includes: The document acquisition and preprocessing module is used to acquire a collection of documents from multiple information sources and generate semantic representations of the documents; The factual representation recognition module is used to identify multiple factual representations surrounding a target fact; The fact search agent module includes multiple independent fact search agents, each of which is bound to a fact expression to generate a corresponding subset of evidence documents; The evidence structuring module is used to extract source information, publication time information, and semantic similarity information from subsets of evidence documents; The evaluation agent module is used to assess the credibility of factual statements based on multiple evaluation factors. The consistency determination and output module is used to output the final consistency conclusion and the corresponding set of evidence documents.

[0013] Beneficial effects: 1. This invention treats different factual expressions as independent processing units, constructs a fact search agent for each factual expression, and performs evidence retrieval in parallel, thereby avoiding mutual interference between different factual judgments and significantly improving the processing stability of the system in multi-document fact verification scenarios.

[0014] 2. This invention employs a centralized evaluation agent to comprehensively assess each factual expression and its corresponding subset of evidence documents. It determines credibility based on multiple dimensions such as source authority, temporal consistency, and semantic support, effectively reducing the impact of the instability of single-model reasoning on the final conclusion.

[0015] 3. This invention performs structured processing on a subset of evidence documents, extracts source information, publication time, and semantic similarity information, and constructs corresponding relationships, making the evaluation process interpretable and traceable, which facilitates subsequent analysis and verification.

[0016] 4. This invention adopts a multi-agent parallel processing architecture, in which each fact search agent runs independently, enabling it to handle retrieval tasks of multiple fact expressions simultaneously, thereby improving the overall processing efficiency of multi-document fact consistency analysis. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a specific embodiment of the present invention. Detailed Implementation

[0018] A multi-agent, multi-document fact consistency conflict resolution method, executed by a computer device, such as... Figure 1 As shown, it includes the following steps: S1. Document Acquisition and Preprocessing Steps: Acquire a collection of documents from multiple information sources, where each document is associated with at least a document source identifier and publication time information; perform sentence segmentation on the document collection, and convert the segmented text into a document semantic representation using a semantic encoding model.

[0019] S2. Fact Expression Recognition Step: Based on a preset target fact, perform semantic clustering or matching processing on the document semantic representation to identify at least two fact expressions surrounding the target fact from the document set. Each fact expression corresponds to a set of semantically similar text fragments pointing to the same factual judgment. To measure the semantic similarity between statements, this step uses the cosine similarity formula between vector representations:

[0020] In the formula, A and B are vector representations of the two statements; This indicates the calculation of the dot product of the two. and

[0021] These represent the magnitudes of the two vectors, respectively.

[0022] S3. Agent Verification Steps for Fact Expression Binding: For each fact expression, a fact search agent is constructed. Each fact search agent independently performs evidence retrieval processing using the corresponding fact expression as the query condition, generating a subset of evidence documents corresponding to that fact expression. The relevance of each retrieved document can be evaluated based on its similarity to the fact expression. This step involves weighted summation of different factors (such as time and source authority). The evidence relevance scoring formula used is as follows:

[0023] In the formula, R(d,e) is the relevance score between document d and fact expression e; , , These represent the weights for time, source authority, and semantic similarity, respectively.

[0024] S4. Structural processing steps for evidence sets: For each subset of evidence documents corresponding to a factual statement, extract the document source information, publication time information, and semantic similarity information between the evidence and the factual statement, and construct the correspondence between the factual statement and its subset of evidence documents.

[0025] S5. Evaluation Steps of the Judging Agent: Input each factual statement and its corresponding subset of evidence documents into the judging agent. The judging agent evaluates the credibility of each factual statement based on at least the following evaluation factors: document source authority factor, temporal consistency factor between evidence and factual statement, and semantic support factor between evidence. The credibility score of each factual statement can be calculated by weighted averaging of credibility factors (e.g., authority, consistency, and semantic support).

[0026] In the formula, is the credibility score of the factual expression f; w1, w2, and w3 are the weights for authority, temporal consistency, and semantic support, respectively.

[0027] S6. Consistency Relationship Determination and Output Steps: Based on the credibility assessment results of each fact statement, determine the relationship type between the fact statements. The relationship type includes at least a fact conflict relationship or a consistent support relationship. Output the final consistency conclusion for the target fact and the set of evidence documents corresponding to the final consistency conclusion.

[0028] System technical solution A multi-agent, multi-document fact consistency conflict resolution system includes: The document acquisition and preprocessing module is used to acquire a collection of documents from multiple information sources and generate semantic representations of the documents; The factual representation recognition module is used to identify multiple factual representations surrounding a target fact; The fact search agent module includes multiple independent fact search agents, each of which is bound to a fact expression to generate a corresponding subset of evidence documents; The evidence structuring module is used to extract source information, publication time information, and semantic similarity information from subsets of evidence documents; The evaluation agent module is used to assess the credibility of factual statements based on multiple evaluation factors. The consistency determination and output module is used to output the final consistency conclusion.

[0029] Example 1: Determining the Consistency of Policy Implementation Time The document set is obtained from policy interpretation documents published by multiple information sources and then processed by sentence segmentation and semantic encoding. Taking "the official implementation time of a certain policy" as the target fact, at least two factual expressions are identified through semantic matching, each corresponding to a different implementation time statement.

[0030] For each factual statement, a fact-searching agent is constructed to independently retrieve a subset of relevant evidence documents from the document set and extract the publication time and source information of each evidence document. This subset of evidence documents is then input into an evaluation agent, which assesses the credibility of each factual statement based on the authority of the source and the temporal consistency between the publication time and the factual statement.

[0031] Finally, the evaluating agent determines the factual expressions with higher credibility as the final consensus conclusion and outputs a set of evidence documents supporting this conclusion.

[0032] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for resolving multi-document fact consistency conflicts based on multi-agent systems, executed by a computer device, characterized in that, include: Obtain a collection of documents from multiple information sources, where each document is associated with at least a document source identifier and publication time information; The document set is segmented into sentences, and a semantic representation of the documents is generated using a semantic encoding model; Based on the preset target facts, at least two factual expressions surrounding the target facts are identified from the semantic representation of the document; For each factual statement, a factual search agent is constructed to independently perform evidence retrieval using that factual statement as the query condition, generating a corresponding subset of evidence documents; Each factual statement and its corresponding subset of evidence documents are input into the evaluation agent, which evaluates the credibility of each factual statement based at least on the authority of the document source, time consistency, and semantic support. Based on the credibility assessment results of each fact statement, the relationship type between fact statements is determined, and a final consistency conclusion and corresponding evidence document set for the target fact are output.

2. The method according to claim 1, characterized in that, The identification of factual expressions from document semantic representations is achieved through semantic clustering or matching, where the cosine similarity between vector representations is used as a semantic similarity metric.

3. The method according to claim 1 or 2, characterized in that, When the fact-searching agent performs evidence retrieval, it scores the relevance between the retrieved documents and the factual statements. The score is a weighted comprehensive evaluation based on time similarity, source authority, and semantic similarity.

4. The method according to claim 1, characterized in that, When the evaluation agent assesses the credibility of each factual statement, it calculates the credibility score by weighting and comprehensively considering authority, temporal consistency, and semantic support.

5. The method according to claim 1, characterized in that, The fact-searching agents operate independently of each other, performing evidence retrieval in parallel based on the bound factual expressions, in order to reduce mutual interference between different factual judgments.

6. The method according to claim 5, characterized in that, The evaluation agent is a centralized agent used to comprehensively evaluate each factual expression and its corresponding subset of evidence documents in order to improve the stability of consistency judgments.

7. A multi-agent, multi-document fact consistency conflict resolution system, characterized in that, include: The document acquisition and preprocessing module is used to acquire a collection of documents from multiple information sources and generate semantic representations of the documents; The factual representation recognition module is used to identify multiple factual representations surrounding a target fact; The fact search agent module includes multiple independent fact search agents, each of which is bound to a fact expression to generate a corresponding subset of evidence documents; The evidence structuring module is used to extract source information, publication time information, and semantic similarity information from subsets of evidence documents; The evaluation agent module is used to assess the credibility of factual statements based on multiple evaluation factors. The consistency determination and output module is used to output the final consistency conclusion and the corresponding set of evidence documents.

Images