Multi-agent-based fact checking method and system, storage medium and electronic equipment

Abstract

The invention relates to the technical field of network data security, discloses a fact checking method and system based on multiple agents, a storage medium and electronic equipment, and aims at solving the problem that the understanding ability of complex declarations is insufficient in the prior art. Decomposing the complex original declaration into a plurality of sub-declarations which cannot be subdivided; in order to solve the problem that external knowledge is lacked to assist in verifying facts in the prior art, a retrieval agent taking Qwen2.5-72B as a basic model is constructed, and fact verification evidences are retrieved from a fact verification knowledge base or the Internet for sub-declarations; in order to solve the problem that in the prior art, a fact checking result lacking logic verification is directly output, Qwen2.5-72B trained through a human preference learning algorithm is constructed to serve as a decision-making agent, whether all evidences ei and answers ai in a sub-declaration sequence logically support an original declaration or not is evaluated, and a final fact checking conclusion is output.

Description

Technical Field

[0001] The present invention relates to fields such as network data security technology, and specifically to a multi-agent based fact verification method, system, storage medium and electronic equipment. Background Art

[0002] With the explosive growth of social media and digital information, the prevalence of disinformation has become a global challenge. False information, misleading content, and malicious rumors spread rapidly and quickly, posing a serious threat to social stability, public health, and business trust. Fact-checking has emerged to address this challenge. It refers to the process of verifying the authenticity of claims through a range of methods and techniques. Existing technologies are categorized into traditional manual fact-checking methods and automated fact-checking methods. While accurate, traditional manual fact-checking methods are inefficient and struggle to cope with the real-time verification demands of massive amounts of information. Therefore, research on automated fact-checking methods is of great practical significance for curbing the spread of disinformation and maintaining a clean information ecosystem. Automated fact-checking methods can be divided into three categories based on their technological evolution: 1. Rule-matching-based fact-checking methods: These rely on predefined rules and keyword matching, and are suitable for simple fact verification, but lack flexibility and have difficulty handling complex semantics. 2. Machine learning-based fact-checking methods: These utilize classification models and natural language processing techniques to improve accuracy through feature engineering, but they are highly dependent on labeled data and have limited generalization capabilities. 3. Large language model-based fact-checking methods: Large language models are trained with large amounts of data and possess a rich knowledge base. They are capable of identifying and verifying factual information in multiple fields, and are gradually becoming a mainstream research direction. However, relying on a single model for factual reasoning may lead to bias or lack of explainability.

[0003] Although existing technologies have improved the accuracy of fact-checking to a certain extent, the following key issues still exist: 1) Insufficient ability to understand complex statements: When dealing with statements that require multi-hop logical reasoning, they are easily troubled by the complexity of the information, which leads to deviations in the final verification conclusions.

[0004] 2) Lack of external knowledge to assist in fact verification: The lack of ability to retrieve external factual evidence has a number of adverse effects on fact verification tasks.

[0005] 3) Fact-checking results lack logical verification: Most fact-checking methods directly output fact-checking conclusions, but lack verification of whether the logical reasoning chain is correct, which reduces the reliability of the results. Summary of the Invention

[0006] The purpose of the present invention is to solve the deficiencies of the prior art and provide a multi-agent based fact-checking method, system, storage medium and electronic device. The fact-checking method addresses the problem that the prior art is unable to understand complex statements. The present invention constructs a sub-statement extraction agent to decompose the complex original statement into multiple indivisible sub-statements. Addressing the problem that the prior art lacks external knowledge to assist in verifying facts, the present invention constructs a retrieval agent to retrieve fact-verification evidence for the sub-statements from a fact-checking knowledge base or the Internet. Addressing the problem that the prior art directly outputs fact-verification results that lack logical verification, the present invention constructs a decision-making agent to evaluate all evidence in the sub-statement sequence. i With answer a i Whether the original statement is logically supported, and the final fact-checking conclusion is output. The system is used to implement the fact-checking method; the storage medium is used to store the computer program that implements the fact-checking method; and the electronic device is a hardware supporting device for the storage medium.

[0007] The present invention is implemented through the following technical solution: a multi-agent based fact-checking method, comprising the following specific steps: 1) The sub-claim extraction agent based on Qwen2.5-72B decomposes the complex original statement into multiple sub-claims, serializes all sub-claims, and generates a logical expression L(a1, a2, ..., a) that can deduce the authenticity of the original statement based on the fact verification conclusions of all sub-claims. i ,…a n ); where the sub-declaration is defined as C i , containing the description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; 2) Using the retrieval agent based on Qwen2.5-72B as the base model, each sub-statement C i Retrieve verification evidence i , and generate the answer a i , according to the logical expression L(a1,a2,…,a i ,…a n ) The answer to the sub-statement a i Combine and generate the initial original statement verification result A; 3) Using Qwen2.5-72B trained by the human preference learning algorithm as a decision agent, evaluate all evidence e in the sub-statement sequence i With answer a i Whether the original statement is logically supported and the final fact-checking conclusion is output.

[0008] To further better implement the multi-agent-based fact-checking method of the present invention, the following configuration is particularly adopted: Step 1) includes the following specific steps: 1.1) Input Understanding and Statement Decomposition: Through the use of a few-sample prompt words, the sub-statement extraction agent analyzes the entities, relations, and logical structure of the original statement and decomposes the original statement into multiple sub-statements. In other words, this step uses the few-sample prompt method to enable the sub-statement extraction agent to better understand the semantics and structure of the original statement and accurately extract key information such as entities (subject, predicate, and object), relations, and logical structure. Based on the extracted key information, the sub-statement extraction agent decomposes the original statement C into multiple sub-statements C i Each sub-declaration C i Each corresponds to a specific verification task. The few-sample prompt words used in this step are as follows: You are a fact-checking assistant who needs to break down an original claim into indivisible atomic sub-claims. Each atomic sub-claim must: 1. contain only a single fact or relation; 2. preserve the entity and logical semantics of the original claim; and 3. be able to fully reconstruct the original claim when the sub-claims are combined.

[0009] Example: Input claim: "** is a Chinese internet company that provides cloud computing services." Output sub-claims: 1. ** is a Chinese internet company. 2. ** provides cloud computing services. For example, the original statement: "Houttuynia cordata contains aristolochic acid, which is not a Class 1 carcinogen" can be broken down into the following sub-statements: C1: "Houttuynia cordata contains aristolochic acid"; C2: "Aristolochic acid is not a Class 1 carcinogen".

[0010] 1.2) Sub-statement serialization: The sub-statement extraction agent is made to serialize all sub-statements through the prompt word method, and the sub-statement series S={C1,C2,…,C i ,…,C n}, this step uses the following prompt words to make the sub-statement extraction agent generate a sub-statement sequence S={C1,C2,…,C i ,…,C n}: “Please organize the decomposed sub-statements into an ordered sub-statement sequence S={C1,C2,…,C i ,…,C n}, each sub-declaration contains a description of c i 、Evidence i and answer a i The question is the specific content of the sub-claim, the evidence is used to store the evidence used to verify the factuality of the sub-claim, and the answer is used to store the factual verification conclusion of the sub-claim (true / false). Except for the description, the other variable values ​​are temporarily empty. Through this step, the sub-declaration sequence S can be expressed as: S=[{"Description":c1,"Evidence":e1,"Answer":a1},{"Description":c2,"Evidence":e2,"Answer":a2},{"Description":c3,Evidence":e3,"Answer":a3}] Each subdeclaration C i Contains description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; 1.3) Logical expression generation: The sub-statement extracts the agent definition through a few-sample prompt word method and outputs the logical expression L(a1, a2, ..., a i ,…a n ), to deduce the authenticity of the original statement based on the sub-statement fact verification conclusion. That is, in order to combine the sub-statement fact verification conclusions into the final original statement fact verification conclusion, the sub-statement extraction agent needs to generate a logical expression L(a1, a2, ..., a i ,…a n ). This expression defines how to deduce the truth of the original claim based on the sub-claim’s factual verification conclusion. This step uses a few-shot prompt method, with the following prompt words: “Please define how to deduce the truth of the original statement based on the sub-statement’s factual verification conclusions, based on the logical structure between the sub-statements and the original statement. For example, if the factual verification conclusions of sub-statement 1 are true and sub-statement 2 are true, then the original statement is true, and the output logical expression L(a1, a2) = a1∧a2 is given.” Through the above steps, the sub-statement extraction agent decomposes the complex statement C into easy-to-process sub-statements C. i , organized into an ordered sub-declaration sequence S={C1,C2,…,C i ,…,C n}, and generates a logical expression L(a1, a2, ..., a i ,…a n ). Provide a clear structure and logical framework for subsequent retrieval and decision-making agents.

[0011] To further better implement the multi-agent-based fact-checking method of the present invention, the following configuration is particularly adopted: Step 2) includes the following steps: 2.1) Evidence Retrieval: The retrieval agent verifies the evidence of the factuality of the sub-claim by searching the fact-checking knowledge base or searching from the Internet using retrieval tools. i; That is, the retrieval agent receives the sub-statement and extracts the sub-statement sequence S={C1,C2,…,C i ,…,C n}, for each sub-declaration C i , the retrieval agent needs to retrieve the verification c i Factual evidence i To obtain external retrieval knowledge, the retrieval agent uses two evidence retrieval methods: retrieval based on fact-checking knowledge base and retrieval from the Internet using retrieval tools.

[0012] 2.2) Sub-claim answer generation: Using prompt words to enable the retrieval agent to compare the evidence e i Describe the sub-claim and judge its truthfulness and give the answer a i After retrieving relevant evidence, this step uses prompt words to enable the retrieval agent to judge the authenticity of the sub-claim by comparing the evidence with the sub-claim description and give an answer. The prompt words are as follows: “Please refer to the evidence i Predicate statement c i The authenticity of the answer (true / false) is given, please store the corresponding answer in a i Among them." For example, for sub-claim description c1: "Houttuynia cordata contains aristolochic acid," evidence e1 is: "Houttuynia cordata does contain aristolochic acid, a metabolite of aristolochic acid, and exists independently and widely in nature..." The answer a1 generated by the search agent is true. For sub-claim description c2: "Aristolochic acid is not a Class 1 carcinogen," evidence e2 is: "Aristolochic acid is not carcinogenic. The ingredient classified as a Class 1 carcinogen by the International Agency for Research on Cancer is aristolochic acid..." The answer a2 generated by the search agent is true.

[0013] 2.3) Initial original statement fact verification conclusion generation: Through the prompt word method, the retrieval agent extracts the logical expression L(a1, a2, ..., a1) generated by the agent according to the sub-statement. i ,…a n ), the answer to the sub-statement a i The combination generates the initial original statement verification result A. That is, this step uses the prompt word method to enable the retrieval agent to extract the logical expression L (a1, a2, ..., a i ,…a n ), combine the answers to the sub-claims into the initial original claim verification result A. The prompt words are as follows: Please extract the logical expression L(a1, a2, ..., a i ,…a n ), the answer to the sub-statement ai Combined into the initial original statement verification result A. " For example, the logical expression is L(a1, a2) = a1Λa2, which means that if the fact verification conclusion of sub-claim 1 is true and the fact verification conclusion of sub-claim 2 is true, then the original claim is true.

[0014] The retrieval agent combines the answers to the sub-claims into the initial original claim verification result A(true)=a1(true)Λa2(true).

[0015] Through the above steps, the retrieval agent provides detailed verification evidence and answers for each sub-claim, and provides the initial original claim verification result A.

[0016] To further implement the multi-agent-based fact-checking method of the present invention, the following configuration is particularly adopted: in step 2.1), the evidence e verifying the factuality of the sub-claim is retrieved from the Internet based on the search of the fact-checking knowledge base or using a search tool. i , comprising the following steps: 2.1.1) Establishing a Fact-Verification Knowledge Base: We collected over 20,000 pieces of factual data verified by professionals from the China Internet Joint Rumor-Refuting Platform and authoritative news websites. These data were then integrated into a fact-verification knowledge base after removing non-standard characters using regular expressions. 2.1.2) Build the Knowledge Base Index: To support subsequent retrieval technology implementation, this step first constructs a dual index consisting of a FAISS semantic index and an Elasticsearch inverted index for each piece of data in the fact-checking knowledge base. The FAISS semantic index is configured with HNSW parameters (M=32, efConstruction=200) to support approximate nearest neighbor search. The Elasticsearch inverted index is configured with the BM25 search algorithm (k1=1.2, b=0.75) to facilitate keyword search. 2.1.3) Calculate the retrieval relevance score: Calculate the retrieval relevance score for each fact-checking knowledge base data after step 2.1.2), and only return the retrieval results with the highest retrieval relevance score above the threshold of 0.8; that is, for each data in the fact-checking knowledge base, calculate two types of scores in parallel: Specifically, for the input sub-claim description c i First parallel calculation: a. Semantic retrieval uses cosine similarity calculation to check each data in the knowledge base and the sub-statement description c i Semantic similarity score between: s sem =cos(SBERT(c i ),v i), where SBERT converts the sub-statement description ci into a semantic vector, v i is the vector representation of fact data in the knowledge base; b. Keyword retrieval uses the BM25 retrieval algorithm to obtain the score s of the match between the fact data and the query keyword lex Next, for each piece of data in the fact-checking knowledge base, calculate the retrieval relevance score. This is done using the following formula: Score=β•s sem +(1-β)•s lex ; Among them, β is the weight that controls the semantic similarity (cosine similarity) and keyword matching score (BM25 score), and its value is 0.7; s sem For each piece of data and sub-claim description c in the fact-checking knowledge base i After the above retrieval relevance score calculation, the retrieval result with the highest retrieval relevance score higher than the threshold of 0.8 is finally returned as evidence to verify the factuality of the sub-statement. i ,If not there, the output is that the search of the fact checking knowledge base has failed; 2.1.4) If the search based on the fact-checking knowledge base fails, the search agent will output that the search for the fact-checking knowledge base failed, and call the search engine (such as the Bing search plug-in) to search from the Internet and return the most relevant first search result as evidence to verify the factuality of the sub-claim. i Finally, the retrieval agent stores the retrieved evidence in each sub-statement C i e i middle.

[0017] For example, for sub-claim description c1: "Zhe Er Radix contains aristolochic acid," evidence e1 might be: "Zhe Er Radix does contain aristolochic acid, which is a metabolite of aristolochic acid and exists widely and independently in nature..." For sub-claim description c2: "Aristolochic acid is not a Class 1 carcinogen," evidence e2 might be: "Aristolochic acid is not carcinogenic. The component classified as a Class 1 carcinogen by the International Agency for Research on Cancer is aristolochic acid..."

[0018] In order to better implement the multi-agent based fact checking method of the present invention, the following configuration is particularly adopted: when calculating the retrieval relevance score, the following formula is used: Score = β·s sem +(1-β)•s lex ; Among them, β is the weight that controls the semantic similarity (cosine similarity) and keyword matching score (BM25 score), and its value is 0.7; s sem For each piece of data and sub-claim description c in the fact-checking knowledge base i The semantic similarity score between lexThe score of the fact data matching the query keyword.

[0019] To further better implement the multi-agent-based fact-checking method of the present invention, the following configuration is particularly adopted: Step 3) includes the following specific steps: 3.1) Optimizing Qwen2.5-72B using the human preference learning algorithm: A human preference dataset is constructed as training data, and then the human preference learning algorithm is used to optimize Qwen2.5-72B to obtain a decision-making agent. In order to perform more accurate logical evaluation, this step uses the human preference learning algorithm DPO to train Qwen2.5-72B. Specifically, it includes two steps: constructing a human preference dataset for training and training Qwen2.5-72B using the human preference learning algorithm.

[0020] 3.1.1) Construct a human preference dataset for training. The human preference dataset contains positive and negative samples. Through the human preference learning algorithm, Qwen2.5-72B will learn how to generate positive samples of human preference and reject negative samples. This step first uses the prompt word method to make Qwen2.5-72B evaluate all evidences in the sub-statement sequence without any optimization conditions. i With answer a i Whether the original statement is logically supported, this process uses more than 1000 input data. The original statements of these input data come from the HOVER dataset, and after steps 1) and 2) a complete sub-statement sequence S={C1,C2,…,C i ,…,C n} as input for constructing a human preference dataset. Examples of prompt words are as follows: “Given the following subclaims, describe the evidence and answer: Sub-claim 1: Houttuynia cordata contains aristolochic acid → Evidence: Houttuynia cordata does contain aristolochic acid, which is a metabolite of aristolochic acid and exists independently in nature → Answer: True Sub-claim 2: Aristolochic acid is not a Class 1 carcinogen → Evidence: Aristolochic acid is not carcinogenic. The ingredient classified as a Class 1 carcinogen by the International Agency for Research on Cancer is aristolochic acid → Answer: True Please judge: Do these evidence and answers logically support the original statement that "Zhe Er Radix contains aristolochic acid lactone, which is not a Class 1 carcinogen"? First, output the logical evaluation conclusion: Y = support / disagree, and then output your thought process. This process generated over 1,000 logical evaluation conclusions and thought processes. Fact-checking experts selected and annotated 800 examples with clear logical evaluation processes and correct conclusions as human preference data. Next, using these 800 data points as input, the Qwen2.5-72B was instructed to intentionally insert incorrect logical thought processes and alter the logical evaluation conclusions using prompts as negative samples. This generated 800 pairs of human preference data for the experiment, with the training and validation sets split at an 8:2 ratio.

[0021] 3.1.2) Then, we use the human preference learning algorithm to train Qwen2.5-72B. The loss function is as follows: ;in, x For input, y w is a positive sample, y l is a negative sample, π θ represents the current strategy, π ref is the reference strategy, i.e. the original strategy, γ is the temperature coefficient (set to 0.3), Represents the data distribution The sample in ( x , y w , y l ) performs expectation calculation; Represents the sigmoid function, which is used to map the input to the interval [0, 1]. The purpose of this loss function is to optimize the model parameters θ by minimizing the difference between the model prediction probability and the reference model prediction probability.

[0022] 3.2) Logical evaluation: The decision agent is made to perform logical evaluation by prompting words, evaluating all evidences in the sub-statement sequence. i With answer a i Whether the original statement is logically supported; that is, Qwen2.5-72B optimized by the human preference learning algorithm in step 3.1) is used as the final decision-making agent to perform a logical evaluation and evaluate all the evidence e in the sub-statement sequence. i With answer a i Whether the original statement is logically supported, this step uses prompt words to enable the decision-making agent to perform a logical evaluation task and output the logical evaluation conclusion Y and the decision-making agent's thinking process; the prompt words are the same as the prompt words used to construct the human preference dataset in step 3.1).

[0023] 3.3) Fact-checking conclusion output: if the original statement is logically supported, the authenticity of the original statement is indeed A, and the evidence and conclusion of the sub-declaration are summarized as the fact-checking conclusion of the original statement and output by the decision-making agent; if the original statement is not logically supported, the logical evaluation fails and the thinking process of the decision-making agent is output. That is, this step determines whether to accept the fact verification result A of the original statement according to whether the logical evaluation conclusion Y supports or does not support, if Y is support, it proves that all evidence e i and answer a i in the sub-declaration sequence indeed logically supports the original statement, and the authenticity of the original statement is indeed A, and the evidence and conclusion of the sub-declaration are summarized as the fact-checking conclusion of the original statement and output, if not, output the logical evaluation failure and the thinking process of the decision-making agent, prompt as follows: “Please determine whether to accept the fact verification result A of the original statement according to whether the logical evaluation conclusion Y supports or does not support, if Y is support, it proves that all evidence e i and answer a i in the sub-declaration sequence indeed logically supports the original statement, and the authenticity of the original statement is indeed A, and the evidence and conclusion of the sub-declaration (that is, S={C1,C2,…,C i ,…,C n}) e i and a i are summarized as the fact-checking conclusion of the original statement and output, if not, output the logical evaluation failure and your thinking process”.

[0024] Through the above process, the decision-making agent verifies whether all evidence e i and answer a i in the sub-declaration sequence logically supports the original statement, and gives the final fact-checking conclusion.

[0025] Further, in order to better realize the fact-checking method based on multiple agents described in the present application, the following setting mode is particularly used: in step 3.1), the human preference data set is constructed as training data, which is: the logically coherent correct evaluation thought and conclusion samples generated by Qwen2.5-72B are used as positive samples; the error evaluation thought and conclusion samples generated by intentionally inserting error logic into the positive samples by Qwen2.5-72B are used as negative samples; the positive samples and the negative samples jointly constitute the training data; When the human preference learning algorithm is used to optimize Qwen2.5-72B, the loss function is: ; wherein, x is the input, y w is the positive sample, y l is the negative sample, and θrepresents the current strategy, π ref is the reference strategy, i.e. the original strategy, γ is the temperature coefficient (set to 0.3), Represents the data distribution The sample in ( x , y w , y l ) performs expectation calculation; Represents the sigmoid function, which is used to map the input to the interval [0, 1]. The purpose of this loss function is to optimize the model parameters θ by minimizing the difference between the model prediction probability and the reference model prediction probability.

[0026] A multi-agent-based fact-checking system, used to implement the multi-agent-based fact-checking method, comprises: The sub-claim extraction agent uses Qwen2.5-72B as the basic model, decomposes the complex original claim into multiple sub-claims, serializes all sub-claims, and generates a logical expression L(a1, a2, ..., a i ,…a n ); where the sub-declaration is defined as C i , containing the description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; Retrieve the agent, using Qwen2.5-72B as the base model, and declare C for each child i Retrieve verification evidence i And generate the answer a i , according to the logical expression L(a1,a2,…,a i ,…a n ) The answer to the sub-statement a i Combine and generate the initial original statement verification result A; The decision agent is trained on Qwen2.5-72B by using the human preference learning algorithm to evaluate all evidences in the sub-statement sequence. i With answer a i Whether the original statement is logically supported and the final fact-checking conclusion is output.

[0027] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the multi-agent-based fact-checking method.

[0028] An electronic device includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the multi-agent based fact-checking method.

[0029] Compared with the prior art, the present invention has the following advantages and beneficial effects: The sub-statement extraction agent employed in this paper is based on the Qwen2.5-72B large model. Using a small sample size, it can break down complex statements into atomic sub-statements, alleviating the limited semantic understanding capabilities of traditional methods when directly processing complex statements. Compared to existing technologies, this method not only improves the accuracy and semantic integrity of sub-statement extraction but also generates logical expressions, significantly enhancing the efficiency and interpretability of statement extraction.

[0030] The retrieval agent of this invention utilizes a dual retrieval mechanism, combining FAISS semantic indexing and Elasticsearch inverted indexing, to achieve efficient and accurate evidence matching. Compared to single retrieval methods, this method prioritizes authoritative knowledge bases and expands evidence coverage by combining a hybrid retrieval algorithm with an internet search backup mechanism.

[0031] This invention uses a decision-making agent optimized by a human preference learning algorithm to perform rigorous logical consistency assessments on sub-claim verification results. Compared to traditional methods that directly output fact-checking conclusions, this method significantly reduces the risk of misjudgment. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a framework diagram of the method described in the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described in detail below with reference to the examples, but the embodiments of the present invention are not limited thereto.

[0034] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention for which protection is sought, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0035] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature identified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0036] Glossary: Fact-checking: This refers to the systematic verification of the authenticity of claims. It typically involves comparing authoritative data sources, logical reasoning, and expert evaluation to reduce the spread of misinformation.

[0037] Intelligent agent: refers to an independent functional module built based on a large model. Each intelligent agent has the ability to handle specific tasks and achieves specialized processing through structured input and output.

[0038] Qwen2.5-72B: This is a large-scale language model developed by Alibaba. It has 72 billion parameters and supports multilingual understanding and generation tasks.

[0039] Few-sample tip: Few-sample learning means that only a small number of typical examples are needed, and the model can understand the new classification logic through these "example samples", avoiding the pain point of traditional machine learning methods that require re-labeling large amounts of data.

[0040] Retrieval-Augmented: Retrieval-Augmented technology combines real-time retrieval results from external knowledge bases with the internal knowledge of large models.

[0041] FAISS (Facebook AI Similarity Search): FAISS is an open-source vector similarity search library from Meta, optimized for fast retrieval of high-dimensional vectors.

[0042] Elasticsearch: Elasticsearch is a distributed search engine based on Lucene. In this solution, it is responsible for handling keyword-driven precise retrieval requirements.

[0043] BM25 retrieval algorithm: BM25 is one of the most classic ranking algorithms in the field of information retrieval. It scores by calculating the statistical correlation between query terms and documents.

[0044] HNSW (Hierarchical Navigable Small World): HNSW is an efficient approximate nearest neighbor search algorithm based on a hierarchical navigable small world graph and is also one of the core indexing algorithms of the FAISS library.

[0045] Human Preference Learning Algorithm: DPO (Direct Preference Optimization) is an optimization algorithm that adjusts model parameters by directly optimizing the model's performance on paired preference data, making the model output more consistent with human preferences and expectations. In practice, the DPO algorithm can directly encourage the model to generate more preferred responses while penalizing less preferred responses, without the need to explicitly train a reward model. This makes DPO a direct and stable optimization method.

[0046] HOVER dataset: HOVER (Human-Verified Fact Verification) is a benchmark dataset for fact-checking. It contains complex claims, their subclaims, supporting evidence, and human-verified labels. This dataset is commonly used to train and evaluate the multi-hop reasoning capabilities of AI models.

[0047] Example 1: The present invention designs a fact-checking method based on multiple agents. The fact-checking method addresses the problem that the existing technology is unable to understand complex statements. The present invention constructs a sub-statement extraction agent to decompose the complex original statement into multiple indivisible sub-statements. In response to the problem that the existing technology lacks external knowledge to assist in verifying facts, the present invention constructs a retrieval agent to retrieve fact-verification evidence for the sub-statements from a fact-checking knowledge base or the Internet. In response to the problem that the existing technology directly outputs fact-verification results that lack logical verification, the present invention constructs a decision-making agent to evaluate all evidence in the sub-statement sequence. i With answer a i Whether the original statement is logically supported and the final fact-checking conclusion is output. Figure 1 As shown, the specific steps include: 1) The sub-claim extraction agent based on Qwen2.5-72B decomposes the complex original statement into multiple sub-claims, serializes all sub-claims, and generates a logical expression L(a1, a2, ..., a) that can deduce the authenticity of the original statement based on the fact verification conclusions of all sub-claims. i ,…a n ); where the sub-declaration is defined as C i , containing the description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; 2) Using the retrieval agent based on Qwen2.5-72B as the base model, each sub-statement C i Retrieve verification evidence i , and generate the answer a i, according to the logical expression L(a1,a2,…,a i ,…a n ) The answer to the sub-statement a i Combine and generate the initial original statement verification result A; 3) Using Qwen2.5-72B trained by the human preference learning algorithm as a decision agent, evaluate all evidence e in the sub-statement sequence i With answer a i Whether the original statement is logically supported and the final fact-checking conclusion is output.

[0048] Example 2: This embodiment is further optimized based on the above embodiment, and the same parts as the above technical solutions are not repeated here. Figure 1 As shown, in order to better implement the multi-agent based fact checking method of the present invention, the following setting is particularly adopted: the step 1) includes the following specific steps: 1.1) Input Understanding and Statement Decomposition: Through the use of a few-sample prompt words, the sub-statement extraction agent analyzes the entities, relations, and logical structure of the original statement and decomposes the original statement into multiple sub-statements. In other words, this step uses the few-sample prompt method to enable the sub-statement extraction agent to better understand the semantics and structure of the original statement and accurately extract key information such as entities (subject, predicate, and object), relations, and logical structure. Based on the extracted key information, the sub-statement extraction agent decomposes the original statement C into multiple sub-statements C i Each sub-declaration C i Each corresponds to a specific verification task.

[0049] 1.2) Sub-statement serialization: The sub-statement extraction agent is made to serialize all sub-statements through the prompt word method, and the sub-statement series S={C1,C2,…,C i ,…,C n}: Each sub-declaration C i Contains description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; 1.3) Logical expression generation: The sub-statement extracts the agent definition through a few-sample prompt word method and outputs the logical expression L(a1, a2, ..., a i ,…a n ), to deduce the authenticity of the original statement based on the sub-statement fact verification conclusion. That is, in order to combine the sub-statement fact verification conclusions into the final original statement fact verification conclusion, the sub-statement extraction agent needs to generate a logical expression L(a1, a2, ..., a i ,…a n). This expression defines how the truth of the original claim is inferred from the factual verification conclusion of the sub-claim.

[0050] Through the above steps, the sub-statement extraction agent decomposes the complex statement C into easy-to-process sub-statements C i , organized into an ordered sub-declaration sequence S={C1,C2,…,C i ,…,C n} and generates a logical expression L(a1, a2, ..., a i ,…a n ). It provides a clear structure and logical framework for subsequent retrieval and decision-making agents.

[0051] Example 3: This embodiment is further optimized based on any of the above embodiments, and the same parts as the above technical solutions are not repeated here. Figure 1 As shown, in order to better implement the multi-agent based fact checking method of the present invention, the following setting is particularly adopted: the step 2) includes the following steps: 2.1) Evidence Retrieval: The retrieval agent verifies the evidence of the factuality of the sub-claim by searching the fact-checking knowledge base or searching from the Internet using retrieval tools. i ; That is, the retrieval agent receives the sub-statement and extracts the sub-statement sequence S={C1,C2,…,C i ,…,C n}, for each sub-declaration C i , the retrieval agent needs to retrieve the verification c i Factual evidence i To obtain external retrieval knowledge, the retrieval agent uses two evidence retrieval methods: retrieval based on fact-checking knowledge base and retrieval from the Internet using retrieval tools.

[0052] 2.2) Sub-claim answer generation: Using prompt words to enable the retrieval agent to compare the evidence e i Describe the sub-claim and judge its truthfulness and give the answer a i That is, after retrieving relevant evidence, this step uses prompt words to enable the retrieval agent to judge the authenticity of the sub-claim by comparing the evidence with the sub-claim description and give an answer.

[0053] 2.3) Initial original statement fact verification conclusion generation: Through the prompt word method, the retrieval agent extracts the logical expression L(a1, a2, ..., a1) generated by the agent according to the sub-statement. i ,…a n ), the answer to the sub-statement a iThe combination generates the initial original statement verification result A. That is, this step uses the prompt word method to enable the retrieval agent to extract the logical expression L (a1, a2, ..., a i ,…a n ), combine the answers to the sub-claims into the initial original claim verification result A.

[0054] Through the above steps, the retrieval agent provides detailed verification evidence and answers for each sub-claim, and provides the initial original claim verification result A.

[0055] Example 4: This embodiment is further optimized based on any of the above embodiments, and the same parts as the above technical solutions are not repeated here. Figure 1 As shown, in order to better implement the multi-agent-based fact-checking method of the present invention, the following configuration is particularly adopted: in the step 2.1), the evidence e of the factuality of the sub-claim is retrieved from the Internet based on the retrieval of the fact-checking knowledge base or by using a retrieval tool. i , comprising the following steps: 2.1.1) Establishing a Fact-Verification Knowledge Base: We collected over 20,000 pieces of factual data verified by professionals from the China Internet Joint Rumor-Refuting Platform and authoritative news websites. These data were then integrated into a fact-verification knowledge base after removing non-standard characters using regular expressions. 2.1.2) Build the Knowledge Base Index: To support subsequent retrieval technology implementation, this step first constructs a dual index consisting of a FAISS semantic index and an Elasticsearch inverted index for each piece of data in the fact-checking knowledge base. The FAISS semantic index is configured with HNSW parameters (M=32, efConstruction=200) to support approximate nearest neighbor search. The Elasticsearch inverted index is configured with the BM25 search algorithm (k1=1.2, b=0.75) to facilitate keyword search. 2.1.3) Calculate the retrieval relevance score: Calculate the retrieval relevance score for each fact-checking knowledge base data after step 2.1.2), and only return the retrieval results with the highest retrieval relevance score above the threshold of 0.8 as evidence to verify the factuality of the sub-claim. i ; 2.1.4) If the search based on the fact-checking knowledge base fails, the search agent will output that the search for the fact-checking knowledge base failed, and call a search engine (such as the Bing search plug-in) to search from the Internet and return the most relevant first search result as evidence to verify the factuality of the sub-claim. i Finally, the retrieval agent stores the retrieved evidence in each sub-statement Ci e i middle.

[0056] Example 5: This embodiment is further optimized based on any of the above embodiments, and the same parts as the above technical solutions are not repeated here. Figure 1 As shown, in order to better implement the multi-agent based fact checking method of the present invention, the following setting is particularly adopted: the step 3) includes the following specific steps: 3.1) Optimizing Qwen2.5-72B with a Human Preference Learning Algorithm: A human preference dataset was constructed as training data, and then a human preference learning algorithm was used to optimize Qwen2.5-72B to obtain a decision-making agent. 3.2) Logical evaluation: The decision agent is made to perform logical evaluation by prompting words, evaluating all evidences in the sub-statement sequence. i With answer a i Whether the original claim is logically supported; 3.3) Fact-checking conclusion output: If the original claim is logically supported and the authenticity of the original claim is indeed A, the sub-claim evidence and conclusion are summarized into the fact-checking conclusion of the original claim and output through the decision-making agent; Through the above process, the decision agent checks all evidences e in the sub-statement sequence. i With answer a i Whether the original claim is logically supported and a final fact-checking conclusion is given.

[0057] Example 6: This embodiment is further optimized based on any of the above embodiments, and the same parts as the above technical solutions are not repeated here. Figure 1 As shown, to further better implement the multi-agent-based fact-checking method described in the present invention, the following configuration is particularly adopted: in step 3.1), constructing a human preference dataset as training data specifically comprises: logically coherent and correct evaluation ideas and conclusion samples generated by Qwen2.5-72B as positive samples; using Qwen2.5-72B to intentionally insert erroneous logic into the positive samples to generate erroneous evaluation ideas and conclusion samples as negative samples; the positive and negative samples together constitute the training data; When using the human preference learning algorithm to optimize Qwen2.5-72B, the loss function is: Among them, x For input, y w is a positive sample, y l is a negative sample, π θ represents the current strategy, πref is the reference strategy, i.e. the original strategy, γ is the temperature coefficient (set to 0.3), Represents the data distribution The sample in ( x , y w , y l ) performs expectation calculation; Represents the sigmoid function, which is used to map the input to the interval [0, 1]. The purpose of this loss function is to optimize the model parameters θ by minimizing the difference between the model prediction probability and the reference model prediction probability.

[0058] Example 7: The subclaim extraction agent uses Qwen2.5-72B as its base model. Its core task is to decompose a complex original statement into multiple indivisible subclaims and generate a logical expression L that derives the truth of the original statement based on the subclaims' factual verification conclusions. The following is a specific implementation of the subclaim extraction agent: 1.1) Input Understanding and Statement Decomposition: This step uses a few-shot prompting approach to enable the sub-statement extraction agent to better understand the semantics and structure of the original statement and accurately extract key information such as entities (subject, predicate, and object), relations, and logical structures. Based on the extracted key information, the sub-statement extraction agent decomposes the original statement C into multiple sub-statements C i Each sub-declaration C i Each corresponds to a specific verification task. The few-sample prompt words used in this step are as follows: You are a fact-checking assistant who needs to break down an original claim into indivisible atomic sub-claims. Each atomic sub-claim must: 1. contain only a single fact or relation; 2. preserve the entity and logical semantics of the original claim; and 3. be able to fully reconstruct the original claim when the sub-claims are combined.

[0059] Example: Input claim: "** is a Chinese internet company that provides cloud computing services." Output sub-claims: 1. ** is a Chinese internet company. 2. ** provides cloud computing services. For example, the original statement: "Houttuynia cordata contains aristolochic acid, which is not a Class 1 carcinogen" can be broken down into the following sub-statements: C1: "Houttuynia cordata contains aristolochic acid"; C2: "Aristolochic acid is not a Class 1 carcinogen".

[0060] 1.2) Sub-statement serialization: In this step, the sub-statement extraction agent generates a sub-statement sequence S={C1,C2,…,C i ,…,C n}: “Please organize the decomposed sub-statements into an ordered sub-statement sequence S={C1,C2,…,C i ,…,C n}, each sub-declaration contains a description of c i 、Evidence i and answer a i The question is the specific content of the sub-claim, the evidence is used to store the evidence used to verify the factuality of the sub-claim, and the answer is used to store the factual verification conclusion of the sub-claim (true / false). Except for the description, the other variable values ​​are temporarily empty. Through this step, the sub-declaration sequence S can be expressed as: S=[{"Description":c1,"Evidence":e1,"Answer":a1},{"Description":c2,"Evidence":e2,"Answer":a2},{"Description":c3,Evidence":e3,"Answer":a3}] Each subdeclaration C i Contains description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; 1.3) Logical expression generation: In order to combine the sub-claims’ fact verification conclusions into the final original claim’s fact verification conclusion, the sub-claim extraction agent needs to generate a logical expression L(a1, a2, … a i ,…a n ). This expression defines how to deduce the truth of the original claim based on the sub-claim’s factual verification conclusion. This step uses a few-shot prompt method, with the following prompt words: "Please define how to deduce the truth of the original claim based on the factual verification conclusions of the subclaims, based on the logical structure between the subclaims and the original claim. For example, if the factual verification conclusions of subclaim 1 and subclaim 2 are true, then the original claim is true. Output the logical expression L(a1, a2) = a1 ∧ a2." Through the above steps, the sub-statement extraction agent decomposes the complex statement C into easy-to-process sub-statements C i , organized into an ordered sub-declaration sequence S={C1,C2,…,C i ,…,C n} and generates a logical expression L(a1, a2, ..., a i ,…a n ). It provides a clear structure and logical framework for subsequent retrieval and decision-making agents.

[0061] Example 8: The retrieval agent uses Qwen2.5-72B as its base model. Its core task is to retrieve factual verification evidence for sub-claims and generate corresponding sub-claim factual verification answers based on the retrieval results. These answers are combined into the initial original original claim factual verification conclusion based on the logical expression L generated by the sub-claim intelligent extraction agent. The following is a specific implementation of the retrieval agent: 2.1) Evidence Retrieval: The retrieval agent receives the sub-statement sequence S={C1,C2,…,C i ,…,C n}, for each sub-declaration C i , the retrieval agent needs to retrieve the verification c i Factual evidence i To acquire external search knowledge, the search agent uses two evidence retrieval methods: searching based on a fact-checking knowledge base and searching from the Internet using search tools. The implementation steps include the following: 2.1.1) Establishing a Fact-Verification Knowledge Base: We collected over 20,000 pieces of factual data verified by professionals from the China Internet Joint Rumor-Refuting Platform and authoritative news websites. These data were then integrated into a fact-verification knowledge base after removing non-standard characters using regular expressions. 2.1.2) Build the Knowledge Base Index: To support subsequent retrieval technology implementation, this step first constructs a dual index consisting of a FAISS semantic index and an Elasticsearch inverted index for each piece of data in the fact-checking knowledge base. The FAISS semantic index is configured with HNSW parameters (M=32, efConstruction=200) to support approximate nearest neighbor search. The Elasticsearch inverted index is configured with the BM25 search algorithm (k1=1.2, b=0.75) to facilitate keyword search. 2.1.3) Calculate the retrieval relevance score: For each piece of data in the fact-checking knowledge base, two types of scores are calculated in parallel: Specifically, for the input sub-claim description c i First parallel calculation: a. Semantic retrieval uses cosine similarity calculation to check each data in the knowledge base and the sub-statement description c i Semantic similarity score between: s sem =cos(SBERT(c i ),v i ), where SBERT describes the sub-statement c i Converted into semantic vector, v i is the vector representation of fact data in the knowledge base; b. Keyword retrieval uses the BM25 retrieval algorithm to obtain the score s of the match between the fact data and the query keyword lexNext, for each piece of data in the fact-checking knowledge base, calculate the retrieval relevance score. This is done using the following formula: Score=β•s sem +(1-β)•s lex ; Among them, β is the weight that controls the semantic similarity (cosine similarity) and keyword matching score (BM25 score), and its value is 0.7; s sem For each piece of data and sub-claim description c in the fact-checking knowledge base i After the above retrieval relevance score calculation, the retrieval result with the highest retrieval relevance score higher than the threshold of 0.8 is finally returned as evidence to verify the factuality of the sub-statement. i ,If not there, the output is that the search of the fact checking knowledge base has failed; 2.1.4) If the search based on the fact-checking knowledge base fails, the search agent will output that the search for the fact-checking knowledge base failed, and call the Bing search plug-in to search from the Internet and return the most relevant first search result as evidence to verify the factuality of the sub-claim. i Finally, the retrieval agent stores the retrieved evidence in each sub-statement C i e i middle.

[0062] For example, for sub-claim description c1: "Zhe Er Radix contains aristolochic acid," evidence e1 might be: "Zhe Er Radix does contain aristolochic acid, which is a metabolite of aristolochic acid and exists widely and independently in nature..." For sub-claim description c2: "Aristolochic acid is not a Class 1 carcinogen," evidence e2 might be: "Aristolochic acid is not carcinogenic. The component classified as a Class 1 carcinogen by the International Agency for Research on Cancer is aristolochic acid..."

[0063] 2.2) Sub-claim answer generation: After retrieving relevant evidence, this step uses prompt words to enable the search agent to determine the authenticity of the sub-claim by comparing the evidence with the sub-claim description and provide an answer. The prompt words are as follows: “Please refer to the evidence i Predicate statement c i The authenticity of the answer (true / false) is given, please store the corresponding answer in a i Among them." For example, for sub-claim description c1: "Houttuynia cordata contains aristolochic acid," evidence e1 is: "Houttuynia cordata does contain aristolochic acid, a metabolite of aristolochic acid, and exists independently and widely in nature..." The answer a1 generated by the search agent is true. For sub-claim description c2: "Aristolochic acid is not a Class 1 carcinogen," evidence e2 is: "Aristolochic acid is not carcinogenic. The ingredient classified as a Class 1 carcinogen by the International Agency for Research on Cancer is aristolochic acid..." The answer a2 generated by the search agent is true.

[0064] 2.3) Initial original statement fact verification conclusion generation: This step uses the prompt word method to enable the retrieval agent to extract the logical expression L(a1, a2, ..., a1) generated by the agent based on the sub-statement. i ,…a n ), combine the answers to the sub-claims into the initial original claim verification result A. The prompt words are as follows: Please extract the logical expression L(a1, a2, ..., a i ,…a n ), the answer to the sub-statement a i Combined into the initial original statement verification result A." For example, the logical expression is L(a1, a2) = a1Λa2, which means that if the fact verification conclusion of sub-claim 1 is true and the fact verification conclusion of sub-claim 2 is true, then the original claim is true.

[0065] The retrieval agent combines the answers to the sub-claims into the initial original claim verification result A(true)=a1(true)Λa2(true).

[0066] Through the above steps, the retrieval agent provides detailed verification evidence and answers for each sub-claim, and provides the initial original claim verification result A.

[0067] Example 9: The decision agent is based on the Qwen2.5-72B model and is optimized through the human preference learning algorithm. Its core function is to evaluate all the evidence in the sub-statement sequence. i With answer a i Whether the original statement is logically supported, and the decision-making agent outputs the final fact-checking conclusion. The following is the specific implementation of the decision-making agent: 3.1) Optimizing Qwen2.5-72B using the human preference learning algorithm: To perform more accurate logic evaluation, this step uses the human preference learning algorithm (DPO) to train Qwen2.5-72B. Specifically, this step involves two steps: constructing a human preference dataset for training and training Qwen2.5-72B using the human preference learning algorithm.

[0068] 3.1.1) Construct a human preference dataset for training. The human preference dataset contains positive and negative samples. Through the human preference learning algorithm, Qwen2.5-72B will learn how to generate positive samples of human preference and reject negative samples. This step first uses the prompt word method to make Qwen2.5-72B evaluate all evidences in the sub-statement sequence without any optimization conditions. i With answer a i Whether the original statement is logically supported, this process uses more than 1000 input data, the original statements of these input data are from the HOVER dataset, and through the steps of Example 7 and Example 8, a complete sub-statement sequence S = {C1, C2, ..., C i ,…,C n} as input for constructing a human preference dataset. Examples of prompt words are as follows: “Given the following subclaims, describe the evidence and answer: Sub-claim 1: Houttuynia cordata contains aristolochic acid → Evidence: Houttuynia cordata does contain aristolochic acid, which is a metabolite of aristolochic acid and exists independently in nature → Answer: True Sub-claim 2: Aristolochic acid is not a Class 1 carcinogen → Evidence: Aristolochic acid is not carcinogenic. The ingredient classified as a Class 1 carcinogen by the International Agency for Research on Cancer is aristolochic acid → Answer: True Please judge: Do these evidence and answers logically support the original statement that "Zhe Er Radix contains aristolochic acid lactone, which is not a Class 1 carcinogen"? First, output the logical evaluation conclusion: Y = support / disagree, and then output your thought process. This process generated over 1,000 logical evaluation conclusions and thought processes. Fact-checking experts selected and annotated 800 examples with clear logical evaluation processes and correct conclusions as human preference data. Next, using these 800 data points as input, the Qwen2.5-72B was instructed to intentionally insert incorrect logical thought processes and alter the logical evaluation conclusions using prompts as negative samples. This generated 800 pairs of human preference data for the experiment, with the training and validation sets split at an 8:2 ratio.

[0069] 3.1.2) Then, we use the human preference learning algorithm to train Qwen2.5-72B. The loss function is as follows: ;in, x For input, y w is a positive sample, y l is a negative sample, π θ represents the current strategy, π refis the reference strategy, i.e. the original strategy, γ is the temperature coefficient (set to 0.3), Represents the data distribution The sample in ( x , y w , y l ) performs expectation calculation; Represents the sigmoid function, which is used to map the input to the interval [0, 1]. The purpose of this loss function is to optimize the model parameters θ by minimizing the difference between the model prediction probability and the reference model prediction probability.

[0070] 3.2) Logical evaluation: After step 3.1) of the human preference learning algorithm optimization, Qwen2.5-72B is used as the final decision agent to perform logical evaluation, evaluating all evidence e in the sub-statement sequence. i With answer a i Whether the original statement is logically supported, this step uses prompt words to enable the decision-making agent to perform a logical evaluation task and output the logical evaluation conclusion Y and the decision-making agent's thinking process; the prompt words are the same as the prompt words used to construct the human preference dataset in step 3.1).

[0071] 3.3) Fact Verification Conclusion Output: This step determines whether to accept the fact verification result A of the original statement based on the logical evaluation conclusion Y. If Y is supported, then all evidence e in the sub-statement sequence is proved. i With answer a i If the original statement is indeed logically supported and the authenticity of the original statement is indeed A, then the sub-statement evidence and conclusions are summarized as the fact-checking conclusion output of the original statement. Otherwise, the logical evaluation failure and the decision-making agent's thinking process are output, as shown below: "Please decide whether to accept the fact verification result A based on the logical evaluation of conclusion Y to see whether it supports or not. If Y is supported, then prove that all evidence e in the sub-statement sequence is correct. i With answer a i If the original statement is indeed logically supported and the truth of the original statement is indeed A, then the sub-statement evidence and the conclusion (i.e. S={C1,C2,…,C i ,…,C n}) in e i with a i Summarize the fact-checking conclusion output for the original statement, if otherwise the output logical evaluation fails with your thought process".

[0072] Through the above process, the decision agent checks all evidences e in the sub-statement sequence. i With answer a iWhether the original claim is logically supported and a final fact-checking conclusion is given.

[0073] Example 10: A multi-agent-based fact-checking system, used to implement the multi-agent-based fact-checking method, comprises: The sub-claim extraction agent uses Qwen2.5-72B as the basic model, decomposes the complex original claim into multiple sub-claims, serializes all sub-claims, and generates a logical expression L(a1, a2, ..., a i ,…a n ); where the sub-declaration is defined as C i , containing the description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; The retrieval agent uses Qwen2.5-72B as the base model and retrieves verification evidence e for each sub-claim Cᵢ i And generate the answer a i , according to the logical expression L(a1,a2,…,a i ,…a n ) The answer to the sub-statement a i Combine and generate the initial original statement verification result A; The decision agent is trained on Qwen2.5-72B by using the human preference learning algorithm to evaluate all evidences in the sub-statement sequence. i With answer a i Whether the original statement is logically supported and the final fact-checking conclusion is output.

[0074] Example 11: A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the multi-agent-based fact-checking method.

[0075] Example 12: An electronic device includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor. When the processor executes the computer program, it implements the multi-agent based fact-checking method.

[0076] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the scope of protection of the present invention.

Claims

1. A multi-agent based fact-checking method, characterized by: The specific steps include: 1) The sub-claim extraction agent based on Qwen2.5-72B decomposes the complex original statement into multiple sub-claims, serializes all sub-claims, and generates a logical expression L(a1, a2, ..., a) that can deduce the authenticity of the original statement based on the fact verification conclusions of all sub-claims. i ,…a n ); where the sub-declaration is defined as C i , containing the description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; 2) Using the retrieval agent based on Qwen2.5-72B as the base model, each sub-statement C i Retrieve verification evidence i , and generate the answer a i , according to the logical expression L(a1,a2,…,a i ,…a n ) The answer to the sub-statement a i Combine and generate the initial original statement verification result A; 3) Using Qwen2.5-72B trained by the human preference learning algorithm as a decision agent, evaluate all evidence e in the sub-statement sequence i With answer a i Whether the original statement is logically supported and the final fact-checking conclusion is output.

2. The multi-agent based fact-checking method according to claim 1, characterized in that: The step 1) includes the following specific steps: 1.1) Using a few-shot prompt word approach, the sub-statement extraction agent analyzes the entities, relations, and logical structure of the original statement, decomposing the original statement into multiple sub-statements. 1.2) The sub-statement extraction agent is made to serialize all sub-statements by using prompt words, and the sub-statement series S={C1,C2,…,C i ,…,C n }; Each sub-declaration C i Contains description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; 1.3) Through the few-sample prompt word method, the sub-statement extracts the agent definition and outputs the logical expression L(a1, a2, ..., a i ,…a n ), to deduce the authenticity of the original statement based on the fact-verification conclusion of the sub-statement.

3. The multi-agent based fact-checking method according to claim 2, characterized in that: The step 2) includes the following steps: 2.1) The retrieval agent retrieves evidence to verify the factuality of the sub-claims by searching the fact-checking knowledge base or using retrieval tools from the Internet. i ; 2.2) Using prompt words to enable the retrieval agent to compare evidence e i Describe the sub-claim and judge its truthfulness and give the answer a i ; 2.3) Through the prompt word, the retrieval agent extracts the logical expression L(a1, a2, ..., a i ,…a n ), the answer to the sub-statement a i The combination generates the initial original claim verification result A.

4. The multi-agent based fact-checking method according to claim 3, characterized in that: In step 2.1), the evidence for verifying the factuality of the sub-claim is retrieved from the Internet based on the search of the fact-checking knowledge base or using a search tool. i , comprising the following steps: 2.1.1) Collect over 20,000 pieces of factual data verified by professionals, cleanse them of non-standard characters using regular expressions, and integrate them into a fact-checking knowledge base; 2.1.2) Build a FAISS semantic index and an Elasticsearch inverted index for each piece of data in the fact-checking knowledge base; 2.1.3) Calculate the retrieval relevance score for each fact-checking knowledge base data after step 2.1.2), and return only the retrieval results with the highest retrieval relevance score above the threshold of 0.8 as evidence to verify the factuality of the sub-claim. i ; 2.1.4) If the search based on the fact-checking knowledge base fails, the search agent will output that the search for the fact-checking knowledge base failed, and call the search engine to search from the Internet and return the first most relevant search result as evidence to verify the factuality of the sub-claim. i .

5. The multi-agent based fact-checking method according to claim 4, characterized in that: When calculating the retrieval relevance score, the following formula is used: Score = β•s sem +(1-β)•s lex ; Among them, β is the weight that controls the semantic similarity and keyword matching score, and its value is 0.7; s sem For each piece of data and sub-claim description c in the fact-checking knowledge base i The semantic similarity score between lex The score of the fact data matching the query keyword.

6. The multi-agent based fact-checking method according to claim 1, characterized in that: The step 3) includes the following specific steps: 3.1) Construct a human preference dataset as training data, and then use the human preference learning algorithm to optimize Qwen2.5-72B to obtain a decision-making agent; 3.2) Use prompt words to enable the decision agent to perform logical evaluation and evaluate all evidence e in the sub-statement sequence i With answer a i Whether the original claim is logically supported; 3.3) If the original statement is logically supported and the authenticity of the original statement is indeed A, the sub-statement evidence and conclusions are summarized as the fact-checking conclusion of the original statement and output through the decision-making agent; if the original statement is logically not supported, the logical evaluation failure and the decision-making agent's thinking process are output.

7. The multi-agent based fact-checking method according to claim 6, characterized in that: In step 3.1), the human preference dataset is constructed as training data by: using the logically coherent and correct evaluation ideas and conclusion samples generated by Qwen2.5-72B as positive samples; using Qwen2.5-72B to deliberately insert incorrect logic into the positive samples to generate incorrect evaluation ideas and conclusion samples as negative samples; the positive samples and negative samples together constitute the training data; when using the human preference learning algorithm to optimize Qwen2.5-72B, the loss function is: ； in, x For input, y w is a positive sample, y l is a negative sample, π θ represents the current strategy, π ref is the reference strategy, i.e. the original strategy, γ is the temperature coefficient, Represents the data distribution The sample in ( x , y w , y l ) performs expectation calculation; σ Represents the sigmoid function, which is used to map the input to the [0, 1] interval.

8. A multi-agent based fact-checking system, characterized by: A method for implementing a multi-agent based fact-checking method as claimed in any one of claims 1 to 7, comprising: The sub-claim extraction agent uses Qwen2.5-72B as the basic model, decomposes the complex original claim into multiple sub-claims, serializes all sub-claims, and generates a logical expression L(a1, a2, ..., a i ,…a n ); where the sub-declaration is defined as C i , containing the description c i , the initial value is empty evidence e i and the answer a with an initial value of null i ; Retrieve the agent, using Qwen2.5-72B as the base model, and declare C for each child i Retrieve verification evidence i And generate the answer a i , according to the logical expression L(a1,a2,…,a i ,…a n ) The answer to the sub-statement a i Combine and generate the initial original statement verification result A; The decision agent is trained on Qwen2.5-72B by using the human preference learning algorithm to evaluate all evidences in the sub-statement sequence. i With answer a i Whether the original statement is logically supported and the final fact-checking conclusion is output.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements a multi-agent based fact checking method as described in any one of claims 1 to 7.

10. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein when the processor executes the computer program, a multi-agent-based fact-checking method as described in any one of claims 1 to 7 is implemented.

Images