False news verification method and system based on heterogeneous evidence collaborative fusion
Through the collaborative fusion of heterogeneous evidence and the multi-head attention mechanism, the problems of insufficient evidence coverage and shallow relationship mining in existing fake news detection are solved, and efficient and accurate detection of fake news is achieved.
Patent Information
- Application Number
- CN202510951719.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-10
- Publication Date
- 2025-10-28
AI Technical Summary
Existing fake news detection methods have limitations in evidence utilization and knowledge fusion, especially in their inaccurate judgment of emerging information and their neglect of key semantic information in complex relationship networks between entities, which limits the model's ability to accurately assess the credibility of news.
By introducing a collaborative fusion mechanism for heterogeneous evidence, using search engines to obtain unstructured documents as evidence sources, and combining the multi-head attention mechanism and TransE model, we deeply mine the association network between entities, generate deep semantic vector representations, and improve evidence quality and relationship understanding.
It significantly expands the breadth and quality of evidence sources, improves the accuracy and robustness of fake news detection, and solves the problems of insufficient evidence coverage, shallow relationship mining, and noise interference in traditional methods.
Smart Images
Figure CN120849686A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the fields of knowledge graphs, retrieval enhancement generation, and natural language processing, specifically to a method and system for verifying fake news based on the collaborative fusion of heterogeneous evidence. Background Technology
[0002] In the field of fake news detection, the quality and relevance of evidence, as well as the effectiveness of the prediction mechanism, are key factors in detection efficiency. Traditional methods typically extract information from commonly used corpora such as Wikipedia, which is limited by outdated or incomplete data. This leads to inaccurate judgments on emerging information and breaking events. Furthermore, one of the key challenges in fake news detection in the knowledge graph domain lies in understanding the diversity and relevance of entity representations. Accurately identifying and associating heterogeneous entity representations is fundamental to in-depth analysis of information consistency and credibility. However, current knowledge graph-based fake news detection methods, due to their overemphasis on entity nodes, severely neglect the key semantic information carried by the complex relationship networks between entities. This results in insufficient comprehensiveness, depth, and accuracy in the introduction of external knowledge, ultimately limiting the model's ability to accurately assess the credibility of news. This invention requires a technical solution that can extract relevant evidence more deeply and utilize knowledge graph relationship information in a more structured way.
[0003] In view of this, those skilled in the art are working to develop a new type of fake news detection system and method to overcome the aforementioned shortcomings in the existing technology. Summary of the Invention
[0005] To address the limitations of existing fake news detection methods in evidence utilization and knowledge fusion, this invention proposes a fake news verification method based on the collaborative fusion of heterogeneous evidence. Unlike methods that rely on commonly used corpora to retrieve relevant information, this invention uses external unstructured documents retrieved through search engines as sources of unstructured evidence, resulting in greater timeliness. It innovatively introduces a similarity-based ranking and filtering mechanism to calculate the semantic similarity between searched document paragraphs and the news to be detected, thereby ranking and selecting the most abundant unstructured evidence. This significantly expands the breadth of evidence sources and improves the quality of the evidence.
[0006] Firstly, this disclosure proposes a method for verifying fake news based on the collaborative fusion of heterogeneous evidence. The method includes: acquiring heterogeneous evidence based on the news text to be detected, wherein the heterogeneous evidence includes textual evidence and entity evidence. The textual evidence is text acquired based on the news text to be detected; entities directly acquired from the news text to be detected are central entities; one-hop neighbor entities obtained from the wikidata general knowledge graph based on the central entity are neighbor entities; and entity triples obtained from the central entity and neighbor entities are entity evidence. Based on the textual evidence and the news text to be detected, a textual evidence set P and a news text vector c' to be detected are obtained through two screening processes. The entity evidence is embedded using the TransE model to obtain the central entity embedding E' and the neighbor entity embedding EC″. The news text vector c' to be detected is used as a query, and its corresponding textual evidence set P is used as the key and value. A multi-head attention mechanism is used to calculate the textual evidence representation p. f The text vector c' to be detected is used as the query, and its corresponding central entity is embedded in E' as the key and value. The central entity representation e is calculated using a multi-head attention mechanism. f The text vector c' to be detected is used as the query, its corresponding central entity embedding E' is used as the key, and the neighbor entities embedding EC″ are used as the values. The neighbor entity representation ec is calculated using a multi-head attention mechanism. f The central entity is represented as e. f Neighbor entity representation (ec) f Textual evidence representation p f After splicing, the data is input into a trained multilayer perceptron to obtain the probability that the news to be detected is true.
[0007] In one embodiment of the above technical solution, the embedding based on entity evidence using the TransE model includes: predicting the links of entities by using the idea that the sum of the head entity and the relation in the vector space is approximately equal to the tail entity, to obtain the embedding of the central entity and the embedding of the neighbor entities corresponding to each entity; and calculating the average value based on the neighbor entity embeddings to obtain the entity neighbor embedding corresponding to each news item.
[0008] In one embodiment of the above technical solution, the two screenings include: filtering the initial set of evidence links obtained based on the news text to be detected based on a predefined list of fake news websites to achieve initial screening; encoding the initially screened text evidence and the news text to be detected using the Stella model to generate evidence text vectors and news text vectors to be detected c', wherein the evidence text vectors are composed of evidence paragraphs d. i p k ', where i is the document identifier and k is the paragraph identifier; calculate d for each evidence paragraph. i p kThe dot product similarity between the vector c' and the news text vector c' to be detected is calculated, and the K paragraphs with the highest similarity are retained to achieve secondary filtering, where K is a preset value.
[0009] In one embodiment of the above technical solution, the loss function of the multilayer perceptron during training is the binary cross-entropy loss.
[0010] In one embodiment of the above technical solution, each evidence segment d is calculated. i p i The dot product similarity between vector c' and the news text vector c' to be detected is as follows: In the formula: This is an embedding function used to map paragraphs and sentences to dense vectors. The specific expression for this function is: It represents the embedding of the i-th token in paragraph p, where |p| is the number of tokens in paragraph p.
[0011] In one embodiment of the above technical solution, the filtering includes deduplication and limiting the maximum number of crawls, and removing duplicate domain name links.
[0012] Secondly, this disclosure proposes a computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods described above.
[0013] Thirdly, this disclosure proposes a fake news verification system based on heterogeneous evidence collaborative fusion. The system includes a heterogeneous evidence acquisition module, a preprocessing module, a multi-head attention mechanism module, and a perception module. Specifically: the heterogeneous evidence acquisition module is configured to acquire heterogeneous evidence based on the news text to be detected. This heterogeneous evidence includes textual evidence and entity evidence. The textual evidence is text acquired from the news text to be detected. Entities directly obtained from the text are central entities. One-hop neighbor entities obtained from the wikidata general knowledge graph based on the central entity are neighbor entities. Entity triples obtained from the central entity and neighbor entities are used as entity evidence. The preprocessing module is configured to perform two filtering operations based on the textual evidence and the news text to be detected to obtain a textual evidence set P and a news text vector c' to be detected. Based on the entity evidence, the TransE model is used for embedding to obtain the central entity embedding E' and the neighbor entity embedding EC″. The multi-head attention mechanism module is configured to use the news text vector c' to be detected as a query, and its corresponding textual evidence set P as the key and value, and use the multi-head attention mechanism to calculate the textual evidence representation p. f The text vector c' to be detected is used as the query, and its corresponding central entity is embedded in E' as the key and value. The central entity representation e is calculated using a multi-head attention mechanism. fThe text vector c' to be detected is used as the query, its corresponding central entity embedding E' is used as the key, and the neighbor entities embedding EC″ are used as the values. The neighbor entity representation ec is calculated using a multi-head attention mechanism. f The sensing module is configured to represent the central entity e. f Neighbor entity representation (ec) f Textual evidence representation p f After splicing, the data is input into a trained multilayer perceptron to obtain the probability that the news to be detected is true.
[0014] Beneficial technical effects: This approach utilizes a real-time search engine interface to acquire dynamic webpage evidence, overcoming the time-sensitive limitations of static corpora; it proposes an entity expansion mechanism to extract entity context, i.e., entity neighbors, thus expanding the coverage of structured evidence. Furthermore, when embedding text vectors, the Stella model, a sentence vector embedding model, is used to effectively address the problem of insufficient relational dependencies caused by excessively long sentences. A multi-head attention mechanism is employed to mine the relevance between evidence and news. Therefore, based on three core modules—dynamic evidence acquisition, neighbor expansion knowledge modeling, and hierarchical attention refinement—this solution systematically solves the bottlenecks in traditional fake news detection, such as insufficient evidence coverage, shallow relational mining, severe noise interference, and weak long text modeling, achieving significant breakthroughs in accuracy, robustness, and efficiency. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 This is a flowchart illustrating the specific implementation of the method proposed in this invention.
[0017] Figure 2 This is a flowchart of the webpage retrieval and filtering process in this invention.
[0018] Figure 3 This is a flowchart of the entity and entity neighbor retrieval process in this invention. Detailed Implementation
[0019] Methods for detecting fake news generally fall into three categories: 1) Content-based fake news detection. This method uses pre-trained models to extract features and then combines them with other deep learning components to attempt to teach the model the general patterns of real and fake news in the training set. During the testing phase, it classifies fake news by determining which pattern it most closely matches. This method often has limitations because fake news is usually too similar to real news in terms of textual features. 2) Social context-based fake news detection. This method extracts features by constructing specific networks, such as user interaction networks, user social structures, and news dissemination networks. However, methods based on social context information require a large amount of social context information and cannot effectively curb the spread of fake news. 3) Evidence-based fake news detection. Recent research increasingly uses evidence verification techniques to detect fake news. Evidence-based fake news detection methods enhance the interpretability of detection by providing evidence of authenticity. This invention focuses on evidence-based methods to improve the interpretability of claim verification.
[0020] This invention proposes a fake news verification method based on heterogeneous evidence collaborative fusion, which integrates evidence knowledge graphs and large language models for enhanced verification. Existing methods for fake news detection using knowledge graphs often only focus on the entity itself or its direct attributes. This invention particularly emphasizes utilizing the "contextual" information of entities in the knowledge graph. Through entity linking tools, not only are key entities in the news identified, but further, the "neighboring nodes" information of each entity in the knowledge graph (i.e., directly connected entities and their relationships) is extracted. These entity contexts (neighboring information) are integrated as auxiliary knowledge into the model learning process, deeply mining the complex relationship network between entities, providing richer structured evidence, and helping to understand the potential relationships between news entities. Traditional text representation methods (such as Word2Vec, GloVe) have shortcomings in capturing long-distance semantic dependencies and handling noise. This invention introduces a large language model to generate deep semantic vector representations for the news text to be detected and unstructured evidence paragraphs, effectively capturing long-distance semantic dependencies within and across texts. When fusing information from different evidence sources, a multi-head attention mechanism is crucially applied to dynamically evaluate the importance of different evidence for fake news detection. This makes the process of incorporating external knowledge more reasonable and efficient, and significantly suppresses the interference of irrelevant or noisy information on the final detection and judgment.
[0021] The following description, in conjunction with the accompanying drawings, clearly and completely describes how the technical solution of this case is implemented. Obviously, the described embodiments are only a part of the embodiments of this case, and not all of them. Based on the embodiments in this case, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.
[0022] S1, Heterogeneous Evidence Set
[0023] When dealing with news being detected, we denote it as 'c'. We then use a search engine to crawl the text as textual evidence, treating this textual evidence as unstructured evidence. An entity linking tool extracts entities and entity context from this text as structured evidence.
[0024] See Figure 2 As shown, a specific implementation method for extracting textual evidence as unstructured evidence through a search engine is as follows: First, obtain an initial evidence link set D = {d1, d2, ..., d...} through the Web Search API (Google Custom Search JSON API). n}, d i ={d i p1,d i p2,…,d i p k The documents retrieved from the webpage are initially organized based on the search engine's relevance algorithm, with the most relevant documents at the top of the list. Therefore, when the API returns, URLs sorted by search engine relevance are returned. Next, the returned documents undergo initial filtering based on a predefined list of fake news websites, i.e., credible source filtering. The designed list is sourced from authoritative fake information databases (e.g., snopes, politifact websites). Deduplication and limiting the maximum number of crawls are performed, removing duplicate domain links and retaining the top 10 links. Then, the webpage content D = {d1, d2, ..., d...} from the links is downloaded concurrently. 10 The text content of these web pages constitutes the preliminary filtered textual evidence.
[0025] See Figure 3 As shown, a specific implementation method for using entity linking tools to extract entities and entity contexts from the news text to be detected as structured evidence is as follows: Using an entity linking tool, such as TagMe, the entities in the news text to be detected are extracted and aligned to the wikidata general knowledge graph G. For example, "monarch butterfly" is associated with "Monarch butterfly", and "milkweed" is associated with "Milkweed", thus obtaining the entity sequence in the news text to be detected. According to the linked central entity Retrieve the skip neighbor entities from the wikidata general knowledge graph G. , This refers to the number of neighboring entities of the central entity. The context subgraph of the entity is stored using RDF triples, and the resulting data is saved in the form of (entity, relation, entity). If... The context entity set is , express and The relationship between two entities, such as Ginkgo - belonging to - Ginkgo phylum. Then the central entity sequence. The corresponding context entity set is , where n is the number of entities in the set.
[0026] The aforementioned text and entity triples constitute a heterogeneous set of evidence.
[0027] S2, Heterogeneous Representation Learning
[0028] Using LLM to encode text semantics, TransE embeds structured knowledge triple semantics.
[0029] For the initially filtered textual evidence D = {d1, d2, ..., d...} 10},d i ={d i p1,d i p2,…,d i p k The text to be detected, c, is encoded using the Stella sentence embedding model. Stella is a text encoding model that supports embeddings of up to 8192 dimensions, capable of capturing semantic information from long texts and generating evidence text vectors d′. i ={d i p1′,d i p2′,…,d i p k Let '} be the document identifier and k be the paragraph identifier, and let c' be the text vector of the news to be detected. The evidence text vector d' is also included. i The component is the evidence paragraph, i.e., d. i p1′,d i p2′,…,d i p k ′ represents the evidence paragraphs.
[0030] For structured evidence, the TransE model is used for entity and neighbor embedding. The basic idea of the TransE model is that if a triple (head entity, relation, tail entity) is true, then the sum of the "head entity" and "relation" in the vector space is approximately equal to the tail entity. For example: vec(Rome) + vec(is-capital-of) ≈ vec(Italy).
[0031] This allows for the completion of missing triples (monarch butterfly, instance of, ?), i.e., link prediction. Through this knowledge graph embedding model, the embedding of the central entity can be obtained: h is the vector dimension, and the embeddings of neighboring entities for each entity are also included. .
[0032] Each news item corresponds to a neighbor entity embedded EC″={e1c″,e2c″,…,e n c″}, where e i c″ is obtained by calculating the average value of the neighbor entity embeddings:
[0033]
[0034] S3. Screening Evidence
[0035] See Figure 2 As shown, the evidence paragraph d retrieved is calculated. i p k The dot product similarity between vector c' and the target news text vector c' is calculated, and the K paragraphs with the highest similarity are retained, where K is a preset value. The similarity calculation formula is as follows: This is an embedding function used to map paragraphs and sentences to dense vectors. The specific expression for this function is: Where stella(p,i) is the embedding of the i-th token in paragraph p, and |p| is the number of tokens in paragraph p.
[0036] The above method sorts unstructured text evidence in descending order of similarity and extracts the Top-K paragraphs, achieving a secondary filtering of unstructured text evidence. Specifically, for each webpage (maximum 10), only the K paragraphs most relevant to the news are retained. For example, setting K to 3 yields the filtered text evidence set P = {p1, p2, ..., p...}. n}, n≤30.
[0037] In some implementations, the initial screening and secondary screening are combined into an evidence filter.
[0038] S4, Quantitative Fusion
[0039] A multi-head attention architecture is built to enable interaction between the news text to be detected and various pieces of evidence.
[0040] A multi-head attention mechanism is employed to independently capture information in different subspaces. Its basic operations include: mapping the input vector to query (Q), key (K), and value (V) respectively; calculating attention weights; and weighted summation to obtain the output representation. The core calculation formula is:
[0041]
[0042] Where d k Let be the dimension of the key vector. First, a multi-head attention mechanism is used to calculate the attention coefficient between the news text c to be detected and the text evidence set P. The news text vector c' is Q, and the unstructured evidence vector P is K / V, according to the formula:
[0043]
[0044] The refined textual evidence representation p after weighted aggregation is calculated. f h is the dimension of the key vector.
[0045] For structured knowledge, two attention networks are designed. For the central entity attention, the news vector c is used as the query Q, and the central entity embedding E′ is used as the K / V, according to the formula:
[0046]
[0047] The central entity representation e after weighted aggregation is calculated. f h is the dimension of the key vector.
[0048] Regarding neighbor entity attention, the news vector c is still used as the query Q, the center entity embedding E′ is K, and the neighbor entity embedding EC″ is V, according to the formula:
[0049]
[0050] The weighted aggregation of neighbor entity representations (ec) is calculated. f h is the dimension of the key vector.
[0051] The central entity is represented as e. f Representing neighboring entities as ec f The structured evidence obtained by splicing together ultimately represents g. f =[e f ;ec f ].
[0052] S5. Test Results
[0053] By classifying multi-source features and integrating news vectors and refined knowledge features, the detection results are output.
[0054] First, the structured knowledge representation g f Compared with textual evidence, p f By concatenating the features, we obtain the fused feature representation: F = [g f ;p f Next, the fused feature F is concatenated with the news text representation c′ to be detected, forming the final input vector S = [c′; g′]. f ;p f ].
[0055] Next, the final input vector S is fed into a multilayer perceptron (MLP) to learn the relationship between input features and output labels through nonlinear transformation. In each layer, the fully connected operation is defined as: h = σ(Wx + b), where x is the input, W is the weight matrix, b is the bias term, and v is the activation function.
[0056] MLP classifier outputs predicted probability The training objective is to minimize the binary cross-entropy loss between the predicted result and the true label y. The cross-entropy loss function is as follows:
[0057]
[0058] Through continuous iteration and optimization, the model's performance in the task of classifying true and false news is improved.
[0059] When applying inference, the input is a news item to be detected, and the output is the probability that it is true news. In some embodiments, the output can also be the probability that it is fake news.
[0060] Based on the above description of the implementation methods, those skilled in the art can correspondingly implement a fake news verification system based on heterogeneous evidence collaborative fusion. For example, the system includes a heterogeneous evidence acquisition module, a preprocessing module, a multi-head attention mechanism module, and a perception module; wherein: the heterogeneous evidence acquisition module is configured to acquire heterogeneous evidence based on the news text to be detected, the heterogeneous evidence including text evidence and entity evidence, the text evidence being text acquired based on the news text to be detected, entities directly acquired from the text being the central entity, and skip neighbor entities acquired from the wikidata general knowledge graph based on the central entity being neighbor entities, and entity triples acquired based on the central entity and neighbor entities being entity evidence; the preprocessing module is configured to perform two filtering operations based on the text evidence and the news text to be detected to obtain a text evidence set P and a news text vector c' to be detected, and to embed the entity evidence using the TransE model to obtain the central entity embedding E' and the neighbor entity embedding EC″; the multi-head attention mechanism module is configured to use the news text vector c' to be detected as a query, and its corresponding text evidence set P as a key and value, and to calculate the text evidence representation p using the multi-head attention mechanism. f The text vector c' to be detected is used as the query, and its corresponding central entity is embedded in E' as the key and value. The central entity representation e is calculated using a multi-head attention mechanism. f The text vector c' to be detected is used as the query, its corresponding central entity embedding E' is used as the key, and the neighbor entities embedding EC″ are used as the values. The neighbor entity representation ec is calculated using a multi-head attention mechanism. f The sensing module is configured to represent the central entity e. f Neighbor entity representation (ec)f Textual evidence representation p f After splicing, the data is input into a trained multilayer perceptron to obtain the probability that the news to be detected is true.
[0061] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods and systems of this disclosure can be implemented using software plus necessary general-purpose hardware, or they can be implemented using dedicated hardware, including dedicated integrated circuits, dedicated CPUs, dedicated memory, dedicated components, etc. Generally, any function performed by a computer program can be easily implemented using corresponding hardware, and the specific hardware structure used to implement the same function can be diverse, such as analog circuits, digital circuits, or dedicated circuits. However, for the purposes of this disclosure, software implementation is often a preferred implementation method.
[0062] Although the embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this disclosure is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this disclosure, and all of these are within the scope of protection of this disclosure.
Claims
1. A method for verifying fake news based on heterogeneous evidence collaborative fusion, characterized in that, The method includes: Heterogeneous evidence is obtained based on the news text to be detected. The heterogeneous evidence includes text evidence and entity evidence. The text evidence is the text obtained based on the news text to be detected. The entity directly obtained based on the news text to be detected is the central entity. The one-hop neighbor entity obtained from the wikidata general knowledge graph based on the central entity is the neighbor entity. The entity triple obtained based on the central entity and the neighbor entity is the entity evidence. Based on textual evidence and the news text to be detected, two rounds of filtering are performed to obtain the textual evidence set P and the news text vector c' to be detected. Based on entity evidence, the TransE model is used for embedding to obtain the central entity embedding E. ′ Neighbor entities are embedded in EC; Using the news text vector c' to be detected as the query and its corresponding text evidence set P as the key and value, the text evidence representation p is calculated using a multi-head attention mechanism. f ; The text vector c' to be detected is used as the query, and its corresponding central entity is embedded in E. ′ The central entity representation e is calculated using a multi-head attention mechanism, with the key and value being the central entity. f ; The text vector c' to be detected is used as the query, and its corresponding central entity is embedded in E. ′ Using EC″ as the key and the neighbor entity embedding as the value, the neighbor entity representation ec is calculated using a multi-head attention mechanism. f ; The central entity is represented as e. f Neighbor entity representation (ec) f Textual evidence representation p f After splicing, the data is input into a trained multilayer perceptron to obtain the probability that the news to be detected is true.
2. The method according to claim 1, characterized in that, The embedding of entity evidence using the TransE model includes: By using the idea that the sum of the head entity and the relation in the vector space is approximately equal to the tail entity, we can predict the links of entities and obtain the embedding of the central entity and the embedding of the neighbor entities corresponding to each entity. The average value of the neighbor entity embeddings is calculated to obtain the entity neighbor embeddings corresponding to each news item.
3. The method according to claim 1, characterized in that, The two screenings include: The initial set of evidence links obtained from the search of the news text to be detected is filtered based on a predefined list of fake news websites to achieve initial screening; The initially filtered textual evidence and the news text to be detected are encoded using the Stella model to generate evidence text vectors and news text vectors to be detected, c'. The evidence text vectors are composed of evidence paragraphs. i p k ′, where i is the document identifier and k is the paragraph identifier; Calculate d for each piece of evidence i p k The dot product similarity between the vector c' and the news text vector c' to be detected is calculated, and the K paragraphs with the highest similarity are retained to achieve secondary filtering, where K is a preset value.
4. The method according to claim 1, characterized in that, The loss function used during training of the multilayer perceptron is the binary cross-entropy loss.
5. The method according to claim 3, characterized in that, Calculate d for each piece of evidence i p k The dot product similarity between vector c' and the news text vector c' to be detected is as follows: In the formula: This is an embedding function used to map paragraphs and sentences to dense vectors. The specific expression for this function is: stella(p,i) is the embedding of the i-th token in paragraph p, and |p| is the number of tokens in paragraph p.
6. The method according to claim 3, characterized in that, The filtering includes deduplication and limiting the maximum number of crawls, and removing duplicate domain name links.
7. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed according to any one of claims 1 to 6.
8. A fake news verification system based on heterogeneous evidence collaborative fusion, characterized in that, The system includes a heterogeneous evidence acquisition module, a preprocessing module, a multi-head attention mechanism module, and a perception module; wherein: The heterogeneous evidence acquisition module is configured to acquire heterogeneous evidence based on the news text to be detected. The heterogeneous evidence includes text evidence and entity evidence. The text evidence is text acquired based on the news text to be detected. The entity directly acquired based on the text is the central entity. The one-hop neighbor entity of the central entity is obtained from the wikidata general knowledge graph as the neighbor entity. The entity triple obtained based on the central entity and the neighbor entity is the entity evidence. The preprocessing module is configured to perform two rounds of filtering based on textual evidence and the news text to be detected to obtain a textual evidence set P and a news text vector c' to be detected. Based on entity evidence, it uses the TransE model for embedding to obtain the central entity embedding E. ′ Neighbor entities are embedded in EC; The multi-head attention mechanism module is configured to use the news text vector c' to be detected as the query, and its corresponding text evidence set P as the key and value, and use the multi-head attention mechanism to calculate the text evidence representation p. f The text vector c' of the news to be detected is used as the query, and its corresponding central entity is embedded in E. ′ The central entity representation e is calculated using a multi-head attention mechanism, with the key and value being the central entity. f The text vector c' of the news to be detected is used as the query, and its corresponding central entity is embedded in E. ′ Using EC″ as the key and the neighbor entity embedding as the value, the neighbor entity representation ec is calculated using a multi-head attention mechanism. f ; The sensing module is configured to represent the central entity e f Neighbor entity representation (ec) f Textual evidence representation p f After splicing, the data is input into a trained multilayer perceptron to obtain the probability that the news to be detected is true.