False news detection method and system based on cooperative reasoning and adaptive backtracking

Abstract

The application relates to the technical field of fake news detection, and proposes a fake news detection method and system based on cooperative reasoning and self-adaptive backtracking. A plurality of intelligent agents are connected in series to work cooperatively, key fact elements are extracted in turn, evidence is searched, contradiction between the key fact elements and the evidence is detected, and fake news discrimination and verification are output; when the judgment evidence coverage and the contradiction point are insufficient, a backtracking mechanism is executed; the event set, the evidence search result and the contradiction set obtained through contradiction detection are comprehensively used to calculate the evidence consistency confidence and output the fake news discrimination result, and a structured interpretable evidence chain is generated. The application realizes transparent traceability and verifiable results of the fake news detection process through the cooperation of the plurality of intelligent agents connected in series and the generation of the structured interpretable evidence chain. Through the self-adaptive backtracking mechanism based on the coverage and the judgment confidence, the evidence or the contradiction is automatically reprocessed when the evidence or the contradiction is insufficient, so that the robustness and the reliability are improved.

Description

Technical Field

[0001] This invention relates to the field of fake news detection technology, specifically to a fake news detection method and system based on collaborative reasoning and adaptive backtracking. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid development of information dissemination, fake news has spread widely through social media, news platforms, and other channels, becoming a pressing problem that needs to be addressed. Especially with the increasing prevalence of multimedia and multimodal content, traditional methods relying on manual identification or rule-based filtering are insufficient to meet the demands of identifying fake content in large-scale, complex contexts. Therefore, developing automated, intelligent, and interpretable fake news detection methods using technologies such as natural language processing, large-scale model reasoning, and evidence retrieval has become an important research direction in the intersection of artificial intelligence and information security.

[0004] However, existing fake news detection technologies still have many shortcomings in engineering practice, making it difficult to support accurate identification and interpretation in highly complex and multimodal environments. On the one hand, most mainstream methods use a single model to perform end-to-end reasoning or process data in simple functional segments, such as a linear process of "information extraction—evidence retrieval—reasoning judgment," lacking deep collaboration and linkage between modules. This results in a "black box" in the reasoning process, poor transparency, and difficulty in providing verifiable judgment criteria. On the other hand, existing technologies generally neglect robust design in cases of incomplete processing results or insufficient evidence coverage. They often draw conclusions based on only partial evidence, failing to adapt to uncertain or ambiguous scenarios and easily leading to misjudgments. Furthermore, in analyzing contradictions between news content and external evidence, traditional methods often use coarse-grained judgments, such as the "contradiction / neutrality / implication" labels in natural language reasoning models. These methods fail to distinguish different types of conflict relationships, such as time, space, facts, and textual semantics, limiting the depth of identification and interpretability of detection systems for complex semantic contradictions. Faced with increasingly complex cross-modal misinformation in real-world scenarios, these limitations significantly constrain the performance of existing methods in terms of accuracy, transparency, and adaptability. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes a fake news detection method and system based on collaborative reasoning and adaptive backtracking. By enabling multi-agent collaboration and generating a structured, interpretable chain of evidence, the fake news detection process achieves transparency, traceability, and verifiability of results. Through an adaptive backtracking mechanism based on coverage and decision confidence, automatic reprocessing is implemented when evidence or contradictions are insufficient, thereby improving robustness and reliability.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of this invention provides a method for detecting fake news based on collaborative reasoning and adaptive backtracking, comprising the following steps: The acquired multimodal news input data is subjected to deep semantic parsing, alignment and fusion, and structured processing in sequence to obtain an event set; Construct a collaborative reasoning model involving multiple agents. Through the interconnected and collaborative work of multiple agents, key fact elements are extracted, evidence is retrieved, contradictions between key fact elements and evidence are detected, and fake news is identified and verified. When the evidence coverage is deemed insufficient, a backtracking mechanism is executed to return to the previous agent and re-examine the evidence retrieval; when there are insufficient points of contradiction, a backtracking mechanism is executed to return to the previous agent and re-examine the contradiction detection between key factual elements and evidence. By combining the event set, evidence retrieval results, and contradiction set obtained from contradiction detection, the consistency confidence of evidence is calculated and the fake news judgment result is output. At the same time, a structured and interpretable evidence chain is generated as the final result of fake news detection.

[0007] A second aspect of this invention provides a fake news detection system based on collaborative reasoning and adaptive backtracking, comprising: The event set construction module is configured to perform deep semantic parsing, alignment and fusion, and structured processing on the acquired multimodal news input data to obtain the event set. The multi-agent collaboration module is configured to build a collaborative reasoning model that includes multiple agents. Through the collaborative work of multiple agents, key fact elements are extracted, evidence is retrieved, contradictions between key fact elements and evidence are detected, and fake news is identified and verified. The backtracking module is configured to execute a backtracking mechanism when the evidence coverage is insufficient, returning to the previous agent to re-retrieve the evidence; and to execute a backtracking mechanism when there are insufficient contradictions, returning to the previous agent to re-detect contradictions between key factual elements and evidence. The output module is configured to integrate the event set, evidence retrieval results, and contradiction set obtained from contradiction detection, calculate the evidence consistency confidence level, output the fake news discrimination result, and generate a structured and interpretable evidence chain as the final result of fake news detection.

[0008] A third aspect of the present invention provides a computer program product, including a computer program / instruction that, when executed by a processor, implements the steps of the fake news detection method based on collaborative reasoning and adaptive backtracking described in the first aspect of the present invention.

[0009] The fourth aspect of the present invention provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the steps in the fake news detection method based on collaborative reasoning and adaptive backtracking described in the first aspect of the present invention.

[0010] The fifth aspect of the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps in the fake news detection method based on cooperative reasoning and adaptive backtracking described in the first aspect of the present invention.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows: To address the technical problems of existing technologies that typically employ end-to-end single models or linear segmented processes, resulting in a black box-like reasoning process, poor transparency, and difficulty in providing verifiable evidence, this invention constructs a collaborative reasoning model involving multiple agents. Multiple agents with independent reasoning capabilities work collaboratively in a series to sequentially complete key fact element extraction, evidence retrieval, contradiction detection between key fact elements and evidence, and fake news identification and verification. This clarifies the input-output relationship of each processing stage and allows the reasoning process to be presented in stages. Simultaneously, this invention generates a structured and interpretable evidence chain while outputting the fake news identification result. It structurally correlates and presents the event set, key fact elements, evidence retrieval results, contradiction set obtained from contradiction detection, and the final judgment result, making the reasoning path and judgment basis traceable, verifiable, and retrievable. This improves the interpretability and retrievability of fake news detection results and reduces the review difficulties and risk of misjudgment caused by simply outputting conclusions without supporting evidence.

[0012] To address the technical problems of existing technologies that generally neglect robust design in scenarios with insufficient evidence coverage or uncertain judgments, and are prone to drawing conclusions based on partial evidence, leading to misjudgments, this invention introduces an adaptive backtracking mechanism based on coverage and judgment confidence: when the evidence coverage is insufficient, backtracking to evidence retrieval for re-retrieval; when there are insufficient contradictions, backtracking to contradiction detection for re-detection. In the overall process, the evidence consistency confidence is calculated by combining the event set, evidence retrieval results, and contradiction set to support the judgment output. This achieves multi-stage control and comprehensive evaluation of evidence coverage and reasoning sufficiency, thereby improving stability and reliability in complex news scenarios and enhancing the system's adaptability and overall robustness in the face of incomplete evidence.

[0013] The advantages of the present invention, as well as its additional advantages, will be described in detail in the following specific embodiments. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute a limitation thereof.

[0015] Figure 1 This is a flowchart of the fake news detection method according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the information extraction intelligent agent module provided in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the retrieval agent module provided in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the contradiction analysis intelligent agent module provided in Embodiment 1 of the present invention; Figure 5 This is a block diagram of a multimodal fake news detection system based on multi-agent collaborative reasoning provided in Embodiment 2 of the present invention. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] It should be noted that the terminology used herein is for describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof. It should be noted that, without conflict, the various embodiments and features within those embodiments can be combined with each other. The embodiments will now be described in detail with reference to the accompanying drawings.

[0019] Example 1 In one or more of the technical solutions disclosed in the embodiments, such as Figures 1 to 4 As shown, the fake news detection method based on collaborative reasoning and adaptive backtracking includes the following steps: Step 1: Perform deep semantic parsing, alignment and fusion, and structured processing on the acquired multimodal news input data to obtain an event set; the multimodal news input data can include news text and news images; Step 2: Construct a collaborative reasoning model involving multiple agents. Through the interconnected and collaborative work of multiple agents, key fact elements are extracted, evidence is retrieved, contradictions between key fact elements and evidence are detected, and fake news is identified and verified. Step 3: If the evidence coverage is insufficient, execute the backtracking mechanism to return to the previous agent and re-examine the evidence retrieval; if the contradiction points are insufficient, execute the backtracking mechanism to return to the previous agent and re-examine the contradiction between key factual elements and evidence. Step 4: Combine the event set, evidence retrieval results, and contradiction set obtained from contradiction detection to calculate the evidence consistency confidence and output the fake news discrimination result. At the same time, generate a structured and interpretable evidence chain as the final result of fake news detection.

[0020] In the above embodiments of this example, addressing the technical problems of existing technologies that typically employ end-to-end single models or linear segmented processes, resulting in a black box-like reasoning process, poor transparency, and difficulty in providing verifiable evidence, this example constructs a collaborative reasoning model involving multiple agents. Multiple agents with independent reasoning capabilities work collaboratively in a serial manner to sequentially complete key fact element extraction, evidence retrieval, contradiction detection between key fact elements and evidence, and fake news identification and verification. This clarifies the input-output relationship of each processing stage and allows the reasoning process to be presented in stages. Simultaneously, this example generates a structured and interpretable evidence chain while outputting the fake news identification result. It structurally associates and presents the event set, key fact elements, evidence retrieval results, contradiction set obtained from contradiction detection, and the final judgment result, making the reasoning path and judgment basis traceable, verifiable, and retrievable. This improves the interpretability and retrievability of the fake news detection results and reduces the review difficulties and risk of misjudgment caused by only outputting conclusions without supporting evidence.

[0021] To address the technical problems of existing technologies generally neglecting robust design in scenarios with insufficient evidence coverage or uncertain judgments, and easily drawing conclusions based on partial evidence, thus leading to misjudgments, this embodiment introduces an adaptive backtracking mechanism based on coverage and judgment confidence: when the evidence coverage is insufficient, backtracking to evidence retrieval for re-retrieval; when there are insufficient contradictions, backtracking to contradiction detection for re-detection. In the overall process, the evidence consistency confidence is calculated by combining the event set, evidence retrieval results, and contradiction set to support the judgment output. This achieves multi-stage control and comprehensive evaluation of evidence coverage and reasoning sufficiency, thereby improving stability and reliability in complex news scenarios and enhancing the system's adaptability and overall robustness in the face of incomplete evidence.

[0022] Step 1, the method for generating the event set, includes the following steps: Step 11: Perform deep semantic analysis on the news text and news images in the multimodal news input data, and realize semantic alignment and fusion between text and images through a cross-modal attention mechanism to generate multimodal semantic descriptions; Step 12: Perform event structuring on the multimodal semantic description to obtain an event set; Furthermore, in step 11 of this embodiment, a multimodal semantic parsing module is constructed to perform deep semantic parsing on news text and news images respectively, and text-image semantic alignment and fusion are achieved through a cross-modal attention mechanism; Specifically, for the news text portion, the news text is segmented and input into the pre-trained language model Qwen2.5-7B. The semantic representation of the text is obtained using the hidden states output by its Transformer backbone network. Let the news text sequence be... Where n is the text length, and the text semantics are represented as: ; in, This is the semantic representation matrix of the text.

[0023] For the news image portion, a visual encoder using a vision-language pre-trained model extracts image features. The vision-language pre-trained model can be a CLIP model. Let the news image be... The image was segmented into Image blocks The semantic representation of the image is as follows: ; in, This is the semantic representation matrix of the image.

[0024] In the cross-modal fusion stage, a cross-attention mechanism is used to achieve semantic alignment between text and images. The formula for calculating attention weights is as follows: ; ; in, , For learnable queries and key projection matrices, Indicates the scaling factor; The fused multimodal semantic representation is as follows: ; in, This indicates a feature concatenation operation. The projection matrix is ​​a value.

[0025] Furthermore, in step 12 of this embodiment, by constructing an event structuring processing module, the multimodal semantic description is decomposed into a structured event representation, which facilitates the subsequent extraction of factual elements and retrieval of evidence by the intelligent agent.

[0026] Event structuring processing utilizes Qwen2.5-7B to fuse semantic representations. Convert to a structured collection of events. Each event is represented as a six-tuple: ; in, As the main body of the event, As the object of the event, For time information, For location information, For action description, A claim or viewpoint regarding an event.

[0027] The event set is represented as: ; in, The number of events identified; Event extraction is guided to execute via a prompt in the Qwen2.5-7B project: ; in, The prompt or instruction template indicates the event extraction process and is used to constrain the model to "extract which fields, output in what format, and follow what rules".

[0028] Step 2: Construct a collaborative reasoning model involving multiple agents, including a serially connected information extraction agent, retrieval agent, contradiction analysis agent, and conclusion verification agent. Each agent executes the following process sequentially: Step 21: Extract key factual elements from the event set using an information extraction agent; Step 22: Through the retrieval agent, a retrieval query is generated based on key fact elements, and retrieval evidence is obtained from the knowledge base, forming an evidence set and evidence context; Evidence set: A structured data set, namely the Top-K highest-scoring evidence fragments output by the retrieval agent through the RAG hybrid retrieval process.

[0029] Evidence context: A textual context formed by organizing and piecing together a set of evidence, used as input by downstream agents during LLM inference. The evidence context is constructed from the evidence set in a specific way.

[0030] Step 23: Using a contradiction analysis agent, detect contradictions between key factual elements and the context of evidence, and output a set of contradictions; Step 24: Verify the agent through the conclusion, integrate the event set, evidence context and contradiction set, calculate the confidence level of evidence consistency and output the fake news judgment result, and generate a structured and interpretable evidence chain as the final output.

[0031] This embodiment constructs a complete closed loop for fake news detection. Starting with multimodal data of input news, it uses collaborative reasoning by serial intelligent agents to ultimately generate an interpretable chain of evidence and a fake news determination result. This method provides a highly transparent reasoning process, enhancing user trust in the reasoning results and flexibly adapting to the needs of different scenarios. Through modular design, the system's maintainability and scalability are improved, ensuring the broad applicability of the technical solution in practical applications.

[0032] One feasible implementation is an information extraction agent that can use a LoRA-tuned Qwen2.5-7B model as the extraction model based on an event set of a specific dataset in the news domain, to accurately extract key factual elements from news events and extract key factual elements from multimodal news. Optionally, key factual elements may include people, time, place, events, etc. Specifically, the fine-tuning process is based on the low-rank adapter (LoRA) strategy, which is optimized for the news domain to improve the model's ability to extract key information from news texts and images.

[0033] This embodiment utilizes LoRA fine-tuning technology to optimize the parameters of the Qwen2.5-7B large language model and construct an information extraction agent module. Through low-rank adaptive (LoRA) technology, while keeping the main parameters of the pre-trained model frozen, a trainable low-rank decomposition matrix is ​​introduced to achieve efficient news domain-adaptive fine-tuning, enabling the model to acquire accurate semantic understanding capabilities for news fact element extraction tasks.

[0034] The LoRA-based information extraction agent was obtained by fine-tuning the Qwen2.5-7B large language model in the news domain. The information extraction agent includes a semantic encoding layer, a LoRA adaptation layer, and an element extraction layer. Furthermore, the process by which the information extraction agent extracts key factual elements includes the following steps: Step 211: Using the event text as input, and leveraging the pre-trained language understanding capabilities of Qwen2.5-7B, obtain the initial semantic representation of the event; Step 212: The LoRA adaptation layer fine-tunes the parameters of the key linear layer of Qwen2.5-7B through low-rank matrix factorization, and processes the initial semantic representation based on the adjusted linear layer to obtain the event semantic representation specific to the news domain. Step 213: The element extraction layer performs entity recognition and attribute extraction on the semantic representation of the event to obtain a set of key fact elements.

[0035] Specifically, the semantic encoding layer performs word segmentation and semantic understanding processing on the target event text, and uses the language understanding capabilities of the Qwen2.5-7B pre-trained model to extract the semantic features of the event.

[0036] Based on the event text input, an input representation is constructed using the vocabulary and positional encoding of Qwen2.5-7B. An initial semantic representation of the event is obtained by pre-training the semantic understanding capabilities of Qwen2.5-7B. : ; in, This represents the forward computation process of the Transformer backbone network in Qwen2.5-7B. This is the initial event semantic representation matrix. To hide the dimension.

[0037] Furthermore, in step 21, the LoRA adaptation layer is constructed based on low-rank matrix factorization. The semantic representation is optimized by introducing a low-rank adaptation matrix into the key linear layer of Qwen2.5-7B to inject news domain-specific information. In the low-rank adaptation stage, i.e., the fine-tuning training stage of the Qwen2.5-7B model, for the linear layer weight matrix in Qwen2.5-7B... The LoRA fine-tuning method uses two low-rank matrices. and Perform incremental parameter updates using the following formula: ; in, It is a low-rank dimension; The core advantage of LoRA adaptation lies in its parameter efficiency. Assuming the original weight matrix... need There are 10 parameters, while LoRA introduces only 10 parameters. One parameter. When At this point, the number of trainable parameters is significantly reduced. The parameter efficiency ratio is: ; In addition to the basic semantic information obtained from the pre-trained weights, news-domain-specific information obtained through LoRA adaptation is also considered. The output of the LoRA adaptation layer is represented as: ; in, The semantic representation of events obtained after LoRA adaptation This is a scaling factor used to control the intensity of the LoRA adaptation effect. Typically set to This is to balance the contributions of pre-trained knowledge with domain-adaptive information.

[0038] In step 212, which is the feature fusion stage, news-domain-specific information from pre-trained knowledge and LoRA adaptation is fused to obtain the final event representation that incorporates multi-level semantics. The semantic representation of an event is recursively represented as follows: ; in, Layer normalization is used to ensure the stability of feature fusion.

[0039] The formula for layer normalization is: ; in, and These are the mean and variance, respectively. and For learnable scaling and offset parameters, This is a small constant set to prevent division by zero.

[0040] After semantic encoding and LoRA adaptation, an event semantic representation that integrates pre-trained knowledge and news-specific information is obtained. : ; in, To obtain the event semantic representation matrix, This is the activation function.

[0041] In step 213, the element extraction layer identifies and extracts key factual elements from the event semantic representation, outputting a structured set of factual elements. This results in the final event semantic representation. The input is fed into the feature extraction layer, where key fact elements are obtained through sequence labeling and slot filling. : ; in, For entities, As an attribute, For attribute values, As the source event identifier, The number of factual elements extracted.

[0042] Optionally, sequence labeling employs a Conditional Random Field (CRF) layer to identify entity boundaries. The CRF selects the optimal label by calculating the conditional probability of the entire label sequence; for event semantic representation sequences... tag sequence The conditional probability is: ; for: ; in, The label transition score matrix; For the first Each position is a tag The emission score is used to describe the strength of the association between the location and the tag.

[0043] Normalization factor: ; in, Indicates that in a given input sequence Any candidate label sequence.

[0044] In this embodiment, CRF uses the Viterbi algorithm to find the globally optimal label sequence, that is, by comprehensively considering the current position score and the probability of label transition before and after, the label sequence with the highest overall probability is selected as the final set of key fact elements; During the training of the information extraction agent in step 212, only the LoRA adaptation matrices A and B and the parameters of the feature extraction layer need to be updated, while the principal parameters of Qwen2.5-7B... Keep the sequence frozen. Consistent with CRF sequence labeling, the training objective of the LoRA fine-tuning Qwen2.5-7B model uses the negative log-likelihood loss of CRF: ; Where N is the number of training samples. For sample index.

[0045] To further improve model performance, label smoothing regularization is introduced: ; During training, the overall loss function is: ; in, , For smoothing coefficients, For the number of tags, For CRF at location The label is The marginal probability.

[0046] The LoRA-based information extraction agent maintains the powerful language understanding capabilities of Qwen2.5-7B while efficiently adapting to news fact element extraction tasks with a small number of trainable parameters, achieving efficient parameter fine-tuning and excellent information extraction performance. By using a large language model fine-tuned with LoRA as the information extraction agent, and performing specific optimizations for the news domain, it can efficiently and accurately extract fact elements from text and images. Simultaneously, the retrieval agent based on RAG hybrid retrieval technology can select appropriate retrieval strategies for different types of news questions, ensuring the efficiency and accuracy of the retrieval process. This technical solution achieves customized processing for different question types through intelligent allocation of computing resources, effectively improving system processing efficiency and reducing computational overhead.

[0047] One feasible implementation involves a retrieval agent employing a RAG (Retrieval-Augmented Generation) hybrid retrieval method. This method generates a retrieval query based on key information obtained from an information extraction agent and retrieves relevant evidence and context from an external knowledge base. The hybrid retrieval strategy matches the query vector with document vectors in the knowledge base based on similarity to retrieve the most relevant evidence fragments. The retrieval agent utilizes RAG-based hybrid retrieval technology, combining the advantages of dense vector retrieval and sparse retrieval.

[0048] The constructed retrieval agent includes a query generation layer, a parallel retrieval layer, and a result fusion layer. In step 22, a retrieval query is generated based on key fact elements, and the evidence set and evidence context are retrieved from the knowledge base. The RAG hybrid retrieval method is adopted, including the following steps: Step 221: The query generation layer converts key fact elements into retrieval queries, generates dense vector representations and sparse feature representations, and merges them into a comprehensive retrieval query representation. Step 222: In the known external knowledge base, the parallel retrieval layer performs dense and sparse retrieval (BM25) through the vector database based on the obtained comprehensive representation of the retrieval query to obtain candidate evidence fragments; Step 223: The result fusion layer sorts and reorders the candidate evidence fragments by relevance and outputs the most relevant set of evidence.

[0049] Specifically, in step 221, the query generation layer performs feature extraction and query construction on key fact elements. For each key fact element... Retrieval queries are generated using the Qwen2.5-7B model. : ; in, This serves as a prompt template to guide the model in generating highly targeted search queries.

[0050] Generated query After post-processing, keywords are extracted and synonyms are expanded to form an expanded query set: ; in, Indicates query The set of synonyms obtained by expanding the keywords in the Chinese dictionary; Representation and Query A semantically related extended term set, including hyponyms, related entities, and domain-related terms.

[0051] Furthermore, a dense vector representation of the query is constructed using a semantic encoding model. ,in For dense vector dimensions, the dense vector representation is obtained through the query encoder of the dual-encoder architecture, as shown in the following formula: ; Furthermore, sparse feature representations are constructed using word segmentation and keyword extraction algorithms. ,in The size of the vocabulary; sparse representation is calculated based on TF-IDF weights: ; in, For words In the query word frequency in Inverse document frequency, Total number of documents For containing words The number of documents.

[0052] Comprehensive representation of search query for: ; In step 222, the parallel retrieval layer is constructed based on a vector database and sparse retrieval. The process of obtaining candidate evidence fragments by performing dense and sparse retrieval (BM25) through the vector database according to the obtained comprehensive representation of the retrieval query includes the following steps: Step 2221, Dense Search: Comprehensive representation of the search query Each query is encoded as a query-dense vector. Approximate Nearest Neighbor (ANN) retrieval is performed in the vector database corresponding to the external knowledge base to recall K1 candidate fragments, resulting in a dense retrieval candidate set. And the dense similarity score for each candidate document; Vector databases are retrieval libraries or indexing systems that encode documents or evidence fragments from external knowledge bases into vectors and store them, then use query vectors to perform similarity searches and quickly return the most relevant Top-K fragments. The external knowledge base provides the original document content, while the vector database stores the vector representation and index of the original document (or its segmented fragments).

[0053] Dense retrieval utilizes query dense vectors Calculate similarity with document vectors from authoritative external knowledge bases: ; An approximate nearest neighbor search algorithm is used to calculate cosine similarity to obtain each candidate document. Dense Scoring : ; Step 2222, Sparse Search: Represent the search query comprehensively. Each query is converted into a sparse representation. By using BM25 for keyword matching, K2 candidate documents are retrieved, resulting in a sparse search candidate set. And the sparse score of each candidate document; ; The formula for calculating sparsity score is: ; in, This indicates the word in query q. In the document word frequency in For document length, This represents the average document length. and For BM25 parameters.

[0054] in, and These represent the number of candidate fragments returned by dense retrieval and sparse retrieval, respectively.

[0055] Step 2223, Preliminary Fusion: Merge the dense retrieval candidate set and the sparse retrieval candidate set to obtain a candidate pool after deduplication. For each candidate document in the candidate pool... A weighted fusion method is used to obtain the mixed score. A second candidate pool is constructed based on the mixed scores, selecting those higher than the set score; ; in, and For the weight parameters, satisfying The weight parameters are optimized based on the characteristics of fake news detection scenarios, with a bias towards semantic retrieval to capture deep semantic relationships.

[0056] Step 223, Reordering: Combine the search query into a single representation. Each query q is compared with candidate documents in the second candidate pool. After concatenation, the input cross-encoder re-ranking model uses attention interaction to model query-document relevance, obtaining a fine-grained ranking score for each candidate segment. : ; Cross encoders model query-document relevance through complete attention interactions: ; in, For splicing sequences The labeled hidden state obtained through forward computation of the model, For the category head weight, It is the sigmoid activation function.

[0057] Output the final evidence set: Based on the re-ranking score, select the Top-K most relevant evidence fragments with the highest ranking scores as the final retrieved evidence set: ; One feasible implementation is a contradiction analysis agent, which is configured to: analyze key factual elements and evidence output by an information extraction agent module and a retrieval agent module, construct a multi-dimensional contradiction classification system, identify spatiotemporal contradictions, factual contradictions, and cross-modal inconsistencies between text and images, and score contradictions based on the degree of contradiction conflict and the credibility of evidence sources.

[0058] Specifically, the contradiction analysis agent detects contradictions between key factual elements and the context of evidence, employs a fine-grained contradiction type classification method, and outputs a set of contradictions and a contradiction intensity score. The contradiction analysis agent constructs a multi-dimensional fine-grained contradiction classification system to achieve accurate positioning and in-depth analysis of contradictions between news content and evidence. The contradiction analysis agent includes a contradiction detection layer, a contradiction classification layer, and a contradiction scoring layer. Step 23, which uses a contradiction analysis agent to detect contradictory pairs between key factual elements and the evidentiary context, and outputs a set of contradictions, includes the following steps: Step 231, the contradiction detection layer, compares key factual elements with evidence one by one to identify potential contradictions; Step 232: Contradictory Classification Layer - The identified contradictions are categorized into a fine-grained contradiction type system. Step 233: Conflict scoring layer. Calculate the conflict intensity score for each conflict pair.

[0059] Specifically, in step 231, the contradiction detection layer uses Qwen2.5-7B to compare key factual elements with evidence one by one, generating contradiction detection results.

[0060] For the set of key factual elements and evidence set Conflict detection is performed using fully connected comparison: ; in, This represents the i-th key fact element. This represents the j-th retrieved piece of evidence in the evidence set; Conflict detection is achieved by guiding the Qwen2.5-7B model to perform Natural Language Inference (NLI) to identify conflicting pairs. : ; in, This is a template for contradiction detection prompts; In step 232, a fine-grained contradiction classification system with multiple classification dimensions is constructed, and the detected contradictory pairs are assigned to their corresponding categories, outputting contradiction type labels: ; In this embodiment, the optional fine-grained contradiction classification system includes the following contradiction types: (1) Time contradiction The time of the event described in the news is inconsistent with the time recorded in the evidence.

[0061] (2) Spatial contradictions The location of the event described in the news is inconsistent with the location recorded in the evidence.

[0062] (3) Substantive contradictions The information about the people, organizations, and other entities mentioned in the news is inconsistent with the evidence.

[0063] (4) Factual contradiction The objective facts asserted in the news report are inconsistent with the objective facts recorded in the evidence.

[0064] (5) Numerical contradictions The data and statistics in the news are inconsistent with those from authoritative sources.

[0065] (6) Cross-modal contradictions There is a semantic inconsistency between the news text description and the accompanying image.

[0066] (7) Conflict regarding the authority of the source The news source cited deviates from the actual authoritative source.

[0067] Contradictory classification is achieved using a multi-label classification model, allowing a contradiction to belong to multiple types simultaneously. Classification probabilities are determined through... calculate: ; in, This is a hidden representation of contradictory pairs. For classification weights.

[0068] In step 233, the contradiction intensity score is obtained by weighting the degree of contradiction and conflict, the credibility of the evidence source, and the relevance of the evidence. Specifically, the formula for calculating the intensity of conflict is as follows: ; in, The degree of conflict is represented by the confidence level of the conflict detection model: ; in, To assess the credibility of evidence sources, a score based on the authority of the source is used: ; in, The ranking is based on the authority, historical accuracy, and professionalism of the source.

[0069] in, For evidence relevance, semantic similarity is used as a metric: ; , , The weighting coefficients are and satisfy the following conditions: .

[0070] A set of contradictions is constructed based on contradiction pairs, contradiction type labels, and contradiction intensity scores. The set of contradictions is represented as follows: ; In the above implementation, addressing the technical problem that existing technologies often use coarse-grained labels in the analysis of contradictions between news content and external evidence, making it difficult to distinguish different conflict relationship types such as time, space, entity, numerical value, and cross-modal relationships, thus limiting the depth and interpretability of complex semantic contradiction identification, this embodiment sets up a layered processing mechanism of contradiction detection layer, contradiction classification layer, and contradiction scoring layer: The contradiction detection layer compares key factual elements with retrieved evidence one by one to locate potential contradiction pairs, and the contradiction classification layer classifies the contradiction pairs into a fine-grained contradiction type system and outputs corresponding contradiction type labels. At the same time, the contradiction scoring layer calculates the contradiction intensity score for each contradiction pair, so that the system can not only discover whether a contradiction exists, but also clarify where the contradiction occurs, what type it belongs to, and how strong it is, thereby improving the granularity and coverage of the identification of complex semantic conflict relationships, and providing verifiable contradiction type and intensity basis for subsequent evidence consistency confidence calculation and structured interpretable evidence chain output, enhancing the interpretability and verifiability of the detection results.

[0071] One feasible implementation involves a conclusion verification agent that integrates the event set, evidence context, and contradiction set to calculate the evidence consistency confidence level and output the fake news discrimination result. Simultaneously, it generates a structured and interpretable evidence chain as the final output.

[0072] Specifically, the conclusion verification agent calculates the confidence level of evidence consistency. The confidence level of evidence consistency is calculated by weighted aggregation of the credibility and contradiction strength of each piece of evidence: ; in, The corresponding contradiction intensity score is assigned. This represents the weight of the credibility of the evidence.

[0073] Subsequently, a comprehensive assessment was conducted based on the authority of the source and the quality of the evidence: ; in, For the first The credibility weight of each piece of evidence. and These represent the source authority indicator and the evidence quality assessment indicator, respectively, which can be obtained through a large language score. Evidence quality assessment indicators are calculated and normalized based on the completeness, timeliness, and relevance of the evidence: ; in, Present evidence The completeness score is obtained by calculating the ratio of the number of factual elements covered in the evidence text to the total number of key factual elements. Present evidence The timeliness score is obtained by normalizing the time difference after mapping it with an exponential decay function. Present evidence The relevance score is obtained by normalizing the fine-ranking score output by the cross encoder in the re-ranking stage of the retrieval agent. , , The weighting coefficients for completeness, timeliness, and relevance are respectively, satisfying the following conditions: .

[0074] Subsequently, the confidence level of evidence consistency was assessed. Make the final judgment: ; in, and These are the thresholds for distinguishing between real and fake news.

[0075] The discrimination result also includes a confidence score: ; The conclusion verification agent simultaneously generates a structured and interpretable chain of evidence, which includes the following components: ; in, It is a collection of events, containing all structured events identified in the news; It is a set of key factual elements, including verifiable factual elements extracted from the event; To retrieve the evidence set, including relevant evidence retrieved from authoritative external knowledge bases and their source information; It is a set of contradictions, including the detected contradiction points, contradiction types, and contradiction strengths; The judgment result includes true / false labels and a confidence score for consistency of evidence; To explain interpretable reasoning, the reasoning process is described in natural language and generated through the Qwen2.5-7B model.

[0076] Furthermore, in step 3, to implement the backtracking mechanism, if the evidence coverage is deemed insufficient, the backtracking mechanism is executed, returning to the previous agent to re-examine the evidence retrieval; if the contradiction points are insufficient, the backtracking mechanism is executed, returning to the previous agent to re-detect the contradictions between key factual elements and evidence. Specifically: Step 31: Set up a retrieval insufficiency backtracking mechanism between the retrieval agent and the contradiction analysis agent. When the evidence coverage is lower than the preset threshold, the retrieval agent is triggered to rewrite the retrieval query and perform a new retrieval. Step 32: Set up a backtracking mechanism for insufficient contradictions between the contradiction analysis agent and the conclusion verification agent. When the number of contradiction sets is lower than the threshold or the confidence level of evidence consistency is in the discrimination boundary range, trigger the contradiction analysis agent to re-execute the contradiction detection. In step 31, for the insufficient retrieval backtracking mechanism, the evidence coverage index is calculated. When the coverage index is lower than the preset threshold, the retrieval agent is triggered to rewrite the retrieval query and perform a new retrieval. Specifically, coverage indicators This is obtained by calculating the degree of semantic matching between key factual elements and retrieved evidence: ; in, For the number of key fact elements, For fact elements With search evidence The semantic similarity between them is calculated using a large language model driven by cue words: ; ; in, As the scoring criteria, The normalized mapping function maps the similarity score output by the model to... Interval.

[0077] Subsequently, when When this occurs, the backtracking mechanism is triggered. The backtracking mechanism starts the retrieval agent, regenerates the retrieval query, and performs supplementary retrieval: ; Furthermore, the strategy for rewriting retrieval queries by the retrieval agent may include: (1) For the uncovered factual elements Regenerate a more specific query; (2) Expand synonyms and related terms; (3) Adjust the query granularity, from abstract to concrete or from concrete to abstract; (4) Add constraints, such as time and location; Specifically, the query rewrite is executed via prompts and guidance in Qwen2.5-7B: ; in, Indicates a prompt word template; Furthermore, in step 32 of this embodiment, the contradiction insufficiency backtracking mechanism triggers the contradiction analysis agent to re-execute contradiction detection when the number of contradiction sets is lower than the threshold or the confidence level of evidence consistency is in the discrimination boundary range.

[0078] Specifically, the trigger condition for backtracking due to insufficient contradictions is: ; in, The minimum threshold for the number of contradictions; and To determine the threshold of the boundary interval, an uncertain region of confidence is defined; This represents the number of contradictory pairs in a single set of contradictions. When a backtracking process for insufficient contradictions is triggered, the contradiction analysis agent is prompted to re-execute the contradiction detection. The contradiction analysis agent uses the following strategy to re-perform the contradiction detection: (1) Refine the granularity of contradiction detection and perform fine-grained comparison at the sentence level for long documents; (2) Introduce additional sources of evidence to supplement the search for more authoritative evidence; (3) Use common sense reasoning to uncover hidden contradictions; (4) Verify consistency across factual elements and detect indirect contradictions.

[0079] In step 4, when the evidence coverage and contradictions meet the requirements, the conclusion verification agent calculates the consistency confidence based on the event set, evidence context, and contradiction set, and outputs the true / false judgment result accordingly. Simultaneously, a structured and interpretable chain of evidence is generated as the final output. This design improves the reliability and traceability of the conclusion.

[0080] Example 2 Based on Example 1, such as Figure 5 As shown, this embodiment provides a fake news detection system based on collaborative reasoning and adaptive backtracking, including: The event set construction module is configured to perform deep semantic parsing, alignment and fusion, and structured processing on the acquired multimodal news input data to obtain the event set. The multi-agent collaboration module is configured to build a collaborative reasoning model that includes multiple agents. Through the collaborative work of multiple agents, key fact elements are extracted, evidence is retrieved, contradictions between key fact elements and evidence are detected, and fake news is identified and verified. The backtracking module is configured to execute a backtracking mechanism when the evidence coverage is insufficient, returning to the previous agent to re-retrieve the evidence; and to execute a backtracking mechanism when there are insufficient contradictions, returning to the previous agent to re-detect contradictions between key factual elements and evidence. The output module is configured to integrate the event set, evidence retrieval results, and contradiction set obtained from contradiction detection, calculate the evidence consistency confidence level, output the fake news discrimination result, and generate a structured and interpretable evidence chain as the final result of fake news detection.

[0081] It should be noted that each module in this embodiment corresponds one-to-one with each step in embodiment 1, and their specific implementation process is the same, so it will not be repeated here.

[0082] Example 3 Based on Embodiment 1, this embodiment provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the fake news detection method based on collaborative reasoning and adaptive backtracking described in Embodiment 1.

[0083] Example 4 Based on Embodiment 1, this embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the processor executes the computer instructions, it completes the steps in the fake news detection method based on collaborative reasoning and adaptive backtracking described in Embodiment 1.

[0084] Example 5 Based on Embodiment 1, this embodiment provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they complete the steps in the fake news detection method based on collaborative reasoning and adaptive backtracking described in Embodiment 1.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. 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.

[0086] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting fake news based on collaborative reasoning and adaptive backtracking, characterized in that, Includes the following steps: The acquired multimodal news input data is subjected to deep semantic parsing, alignment and fusion, and structured processing in sequence to obtain an event set; Construct a collaborative reasoning model involving multiple agents. Through the interconnected and collaborative work of multiple agents, key fact elements are extracted, evidence is retrieved, contradictions between key fact elements and evidence are detected, and fake news is identified and verified. When the evidence coverage is deemed insufficient, a backtracking mechanism is executed to return to the previous agent and re-examine the evidence retrieval; when there are insufficient points of contradiction, a backtracking mechanism is executed to return to the previous agent and re-examine the contradictions between key factual elements and evidence. By combining the event set, evidence retrieval results, and contradiction set obtained from contradiction detection, the consistency confidence of evidence is calculated and the fake news discrimination result is output. At the same time, a structured and interpretable evidence chain is generated as the final result of fake news detection.

2. The fake news detection method based on collaborative reasoning and adaptive backtracking as described in claim 1, characterized in that: The method for generating an event set includes the following steps: Deep semantic analysis is performed on news text and news images in multimodal news input data, and semantic alignment and fusion between text and images are achieved through a cross-modal attention mechanism to generate multimodal semantic descriptions; The event set is obtained by performing event structuring on the multimodal semantic description.

3. The fake news detection method based on collaborative reasoning and adaptive backtracking as described in claim 1, characterized in that: Construct a collaborative reasoning model involving multiple agents, including a series-connected information extraction agent, retrieval agent, contradiction analysis agent, and conclusion verification agent. Each agent executes the following process sequentially: By using an information extraction agent, key factual elements are extracted from a set of events. By using a retrieval agent, a retrieval query is generated based on key factual elements, and retrieval evidence is obtained from the knowledge base, forming an evidence set and evidence context. The contradiction analysis agent detects contradictory pairs between key factual elements and the context of evidence, and outputs a set of contradictions. The conclusion verification agent integrates the event set, evidence context, and contradiction set to calculate the evidence consistency confidence and output the fake news discrimination result. At the same time, it generates a structured and interpretable evidence chain as the final output.

4. The fake news detection method based on collaborative reasoning and adaptive backtracking as described in claim 3, characterized in that: The process of an information extraction agent extracting key factual elements includes the following steps: Using the event text as input, the initial semantic representation of the event is obtained by leveraging the pre-trained language understanding capabilities of Qwen2.5-7B. The key linear layer of Qwen2.5-7B is fine-tuned by low-rank matrix factorization. The initial semantic representation is then processed based on the adjusted linear layer to obtain a news domain-specific event semantic representation. Entity recognition and attribute extraction are performed on the semantic representation of events to obtain a set of key factual elements.

5. The fake news detection method based on collaborative reasoning and adaptive backtracking as described in claim 3, characterized in that: Based on key factual elements, a retrieval query is generated. Evidence sets and context are retrieved from the knowledge base. The RAG hybrid retrieval method is employed, including the following steps: The query generation layer transforms key factual elements into retrieval queries, generates dense vector representations and sparse feature representations, and merges them into a comprehensive retrieval query representation. In the known external knowledge base, the parallel retrieval layer performs dense and sparse retrieval through the vector database based on the comprehensive representation of the retrieval query obtained, and obtains candidate evidence fragments. The fusion layer performs relevance ranking and reordering of candidate evidence fragments, and outputs the most relevant set of evidence. Alternatively, a method that uses a contradiction analysis agent to detect contradictory pairs between key factual elements and the context of evidence, and outputs a set of contradictions, includes the following steps: Compare key factual elements with evidence one by one to identify potential contradictions; The identified contradictory pairs are categorized into a fine-grained contradiction type system, the detected contradictory pairs are categorized into their corresponding categories, and contradiction type labels are output. For each contradiction, a score for the contradiction intensity is calculated. Contradictory sets are constructed based on contradiction pairs, contradiction type labels, and contradiction intensity scores; Alternatively, the conclusion verification agent calculates the confidence level of evidence consistency. The confidence level of evidence consistency is calculated by weighted aggregation of the credibility and contradiction strength of each piece of evidence: ； in, The corresponding contradiction intensity score is assigned. This represents the weight of the credibility of the evidence.

6. The fake news detection method based on collaborative reasoning and adaptive backtracking as described in claim 3, characterized in that: When the evidence coverage is deemed insufficient, a backtracking mechanism is executed, returning to the previous agent to re-search for evidence; when there are insufficient points of contradiction, a backtracking mechanism is executed, returning to the previous agent to re-detect contradictions between key factual elements and evidence. Specifically: A retrieval insufficiency backtracking mechanism is set up between the retrieval agent and the contradiction analysis agent. When the evidence coverage is lower than the preset threshold, the retrieval agent is triggered to rewrite the retrieval query and perform a new retrieval. A backtracking mechanism for insufficient contradictions is set up between the contradiction analysis agent and the conclusion verification agent. When the number of contradiction sets is lower than the threshold or the confidence level of evidence consistency is in the discrimination boundary range, the contradiction analysis agent is triggered to re-execute the contradiction detection.

7. A fake news detection system based on collaborative reasoning and adaptive backtracking, characterized in that, include: The event set construction module is configured to perform deep semantic parsing, alignment and fusion, and structured processing on the acquired multimodal news input data to obtain the event set. The multi-agent collaboration module is configured to build a collaborative reasoning model that includes multiple agents. Through the collaborative work of multiple agents, key fact elements are extracted, evidence is retrieved, contradictions between key fact elements and evidence are detected, and fake news is identified and verified. The backtracking module is configured to execute a backtracking mechanism when the evidence coverage is insufficient, returning to the previous agent to re-retrieve the evidence; and to execute a backtracking mechanism when there are insufficient contradictions, returning to the previous agent to re-detect contradictions between key factual elements and evidence. The output module is configured to integrate the event set, evidence retrieval results, and contradiction set obtained from contradiction detection, calculate the evidence consistency confidence level, output the fake news discrimination result, and generate a structured and interpretable evidence chain as the final result of fake news detection.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps of the fake news detection method based on collaborative reasoning and adaptive backtracking as described in any one of claims 1-6.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps in the fake news detection method based on collaborative reasoning and adaptive backtracking as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the steps in the fake news detection method based on collaborative reasoning and adaptive backtracking as described in any one of claims 1-6.

Images