Microblog rumor detection method and system based on static and dynamic knowledge enhancement

Abstract

The invention discloses a microblog rumor detection method and system based on static and dynamic knowledge enhancement. The method comprises the following steps: firstly, preprocessing a microblog text, including text cleaning, word segmentation, part-of-speech tagging and named entity recognition, so as to obtain a structured text and entity information; and then knowledge linking and semantic enhancement are performed on the text entities by utilizing a pre-constructed field static knowledge graph, entity attributes, relationships and related facts are extracted, and deep combination of texts and background knowledge is realized. For core declarations that the knowledge graph is not covered or has disputes, a dynamic search enhancement mechanism is further introduced, authoritative external evidences are obtained through a search engine, and dynamic knowledge representation is constructed. And then respectively evaluating semantic consistency and logic consistency between the microblog content and the static knowledge and between the microblog content and the dynamic evidence by adopting a knowledge consistency verification module, and quantifying the credibility of the microblog content. Finally, the text semantic features, the knowledge enhancement features, the user behavior features, the propagation features and the knowledge credibility scores are fused, a multi-dimensional comprehensive feature vector is constructed, and rumor recognition is completed through a deep neural network classifier. Compared with a traditional detection technology depending on content or a propagation structure, the method has the advantages that structured knowledge and real-time external evidences can be utilized at the same time, complex, novel and high-timeliness rumors can be accurately recognized, and more reliable and efficient technical support is provided for social media content auditing, public opinion monitoring and public safety management.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of social network information credibility analysis and false information identification, and in particular to a microblog rumor detection method and system based on static and dynamic knowledge enhancement. BACKGROUND

[0002] With the rapid popularization of mobile Internet and social media, social network platforms represented by microblogs have become an important channel for the public to obtain news information and exchange views. However, social media has the characteristics of fast transmission, low access threshold, and lack of strict audit mechanism, which leads to false information and rumors that can spread rapidly in a short time, seriously interfering with public awareness, and even causing social panic and group incidents. Therefore, how to effectively identify and curb the spread of rumors in social networks has become an important problem to be solved in the fields of natural language processing, social computing, and public safety management.

[0003] Existing rumor detection research can be mainly divided into two categories: content feature-based methods and propagation structure-based methods. Content feature-based methods usually extract semantic features, sentiment features, and other information from microblog text for classification and judgment, but this type of method relies on existing labeled corpus for supervised learning, and the degree of utilization of external knowledge is limited. When facing complex rumors that need to be judged in combination with factual background or domain knowledge, the recognition effect is still insufficient. Propagation structure-based methods focus on analyzing the diffusion path, user interaction relationship, and propagation mode of rumors in social networks, which can describe the structural characteristics of rumor propagation, but in actual application, the utilization of text semantic information is often insufficient, and the collaborative modeling ability between content information and propagation structure is limited. In addition, most existing methods still mainly rely on static features or offline modeling, which is difficult to effectively adapt to the actual scenarios of rumor content evolution over time and the rapid emergence of new rumors.

[0004] In recent years, some research has attempted to integrate external knowledge or multi-modal information to improve the robustness of rumor detection, but there are still obvious deficiencies. Publication No. CN120216780A (Rumor detection method based on information propagation structure) discloses a method that constructs a tweet propagation tree, a retweet subgraph, and a global user forwarding network, combines a deep walk generation tree algorithm and a WL algorithm to extract local and global structural features, uses a pre-trained large language model to extract content features, and inputs structural encoding, content encoding, and position encoding into a Graph Transformer for fusion and classification. This method fully utilizes the complexity of the propagation structure, but its effectiveness is highly dependent on the completeness of user interaction data, and there are obvious limitations for early rumor identification or data scenarios lacking propagation depth.

[0005] Publication No. CN120336605A (Multi-modal rumor detection system based on multi-domain perception) discloses a system that obtains aligned features of text and images through a multi-view feature extraction module, learns cross-domain shared knowledge and domain-specific knowledge in a multi-domain perception module, integrates knowledge using a multi-level knowledge fusion module, and finally completes discrimination through a classifier. This method has certain advantages in multi-modal and multi-domain knowledge fusion, but still mainly relies on the representation learned by the model internally, and does not introduce a real-time external knowledge supplement mechanism, so there are deficiencies in handling cross-event migration or emerging rumor topics.

[0006] In summary, although the prior art has made certain progress in propagation structure modeling, multi-modal feature fusion, and deep representation learning, there are still the following deficiencies in actual rumor detection applications: (1) the coverage of static knowledge graph is limited and cannot respond to newly emerging entities and events in time; (2) there is a lack of real-time knowledge verification mechanism for rumor content, making it difficult to dynamically obtain the latest evidence; (3) external resources such as search engines are not fully utilized for dynamic knowledge enhancement, which restricts the recognition ability of new rumors and rumors related to sudden events. SUMMARY

[0007] In view of the above defects of the prior art, the present application provides a microblog rumor detection method and system based on static and dynamic knowledge enhancement, aiming to solve the technical problems of insufficient knowledge coverage, lack of real-time verification ability and difficulty in dealing with new rumors of traditional rumor detection methods.

[0008] The present application proposes a comprehensive detection method that combines static knowledge graph and dynamic search knowledge enhancement. This method provides a structured knowledge base through a pre-constructed domain knowledge graph, and combines a real-time search mechanism to obtain the latest external evidence, achieving accurate identification of microblog rumors. This method not only utilizes structured knowledge for semantic reasoning, but also adapts to rapidly changing social events through dynamic search, significantly improving the accuracy and timeliness of rumor detection.

[0009] Specifically, the application firstly pre-processes the micro-blog text, including word segmentation, entity recognition and key information extraction; then uses the pre-constructed static knowledge graph to perform knowledge matching and semantic enhancement on the entities in the text; for entities not covered by the knowledge graph or controversial statements, dynamic knowledge retrieval is performed through a search engine to obtain external evidence such as related web pages and news reports; then a knowledge consistency verification module is designed to calculate the credibility score by comparing the consistency of the micro-blog content with the knowledge graph and the search results; finally, the text semantic features, knowledge enhanced features, user behavior features and propagation mode features are fused to construct a multi-level feature representation, and a deep learning classifier is used for rumor judgment. The application can not only effectively identify various rumors, but also provide strong technical support for social media content review and public opinion supervision.

[0010] The first aspect of the application provides a micro-blog rumor detection method based on static and dynamic knowledge enhancement, comprising the following steps: S1: data loading and preprocessing: loading a micro-blog data set, the data set containing micro-blog text and its related attributes, performing noise cleaning, word segmentation, named entity recognition and semantic coding on the text to obtain text semantic features and entity information; S2: static knowledge enhancement: based on the entities obtained in step 1, using a pre-constructed knowledge graph to perform entity linking, and extracting the attributes, relationships and related facts of the entities, generating static knowledge enhanced features through knowledge graph embedding and attention aggregation mechanism; S3: dynamic search enhancement: for entities that cannot be matched in the knowledge graph or key statements in the text, generating corresponding search queries and obtaining related web page content through a search engine interface, screening effective evidence based on source credibility and semantic consistency, and constructing dynamic search enhanced features; S4: knowledge consistency verification: integrating static knowledge facts and dynamic search evidence to evaluate the consistency of the text content, calculating the knowledge credibility score based on semantic matching, relationship verification and logical reasoning, which is used to reflect the degree of conformity between the text and the multi-source knowledge; S5: multi-dimensional feature fusion: fusing the text semantic features, static knowledge enhanced features, dynamic search enhanced features and knowledge credibility score to construct a multi-dimensional fusion representation; S6: construction, training and rumor detection of deep neural network classifier: building a deep neural network classifier, using the multi-dimensional fusion representation, and using the deep neural network classifier to judge the micro-blog content for rumors, outputting the rumor probability and the corresponding confidence.

[0011] The specific steps of the data preprocessing of step S1 are: S1.1: loading a micro-blog data set, the data set being a public micro-blog rumor data set or a data set collected and constructed according to keywords / events, containing Each record contains fields such as Weibo ID, text content, posting time, user information, number of reposts, number of comments, and tags. S1.2: Clean the text content fields to remove noise information such as URL links, special characters, and emoticons, while retaining the core text content; S1.3: One of the word segmentation and part-of-speech tagging tools, Jieba / THULAC / LTP / HanLP, is used to segment the cleaned text and perform part-of-speech tagging. S1.4: Use BiLSTM-CRF or BERT-CRF models to identify entities such as personal names, place names, organization names, time, and numbers in the text, and record the entity text, category, and location for subsequent knowledge alignment; S1.5: Using a pre-trained language BERT model to process the original text Encoding is performed to obtain a semantic representation vector of the original text. This vector captures the overall semantic information of the text and serves as the basis for subsequent feature fusion.

[0012] The specific steps of static knowledge enhancement described in step S2 are as follows: S2.1: Load the pre-built domain knowledge graph ,in It is a collection of entities, and is not required to cover all entities appearing in the text. For a set of relations, For the set of triples , indicating the head entity Through relationships With tail entity For text entities that cannot find a corresponding node in the knowledge graph, they will not participate in subsequent knowledge enhancement steps. S2.2: For each entity identified in step S1.4 In the knowledge graph, entity linking is performed by mapping text entities to standard entity nodes in the knowledge graph through methods such as string matching and semantic similarity calculation, defining entity link scores, and selecting the candidate entity with the highest score as the link result. S2.3: For successfully linked entities, extract their multi-hop neighbor information in the knowledge graph, and obtain the entity's [unclear - possibly related to a random walk strategy]. Jump to the neighbor; S2.4: Loaded domain knowledge graph Each entity in and relationships Vectorization is performed, and their respective definitions are defined. 3D embedding vectors yield entity vectors With relation vector This vector supports a hierarchical knowledge graph embedding method, which combines the translation properties of TransE and the rotation properties of RotatE to construct a hybrid embedding loss function. S2.5: Targeting text containing... Entities that are successfully linked Utilizing its embedding vector With text semantic vectors By performing knowledge attention aggregation, a static knowledge-enhanced representation of the text is obtained. Through the attention aggregation mechanism, the model can assign different weights to different entities based on the semantics of the text, giving higher contributions to entities more relevant to rumor detection. As a static knowledge augmentation vector, it will be used in subsequent steps along with text semantic representation. To integrate.

[0013] The specific steps of the dynamic search enhancement described in step S3 are as follows: S3.1: After extracting claims from the Weibo text, to avoid unnecessary external searches for unverifiable content, the verifiability of each claim is determined. A dependency parsing-based claim extraction method is used to extract claims from the Weibo text. Identify structured declarations in the text; S3.2: Based on the verifiable statement obtained in step S3.1, examine the entities involved in the statement. Knowledge coverage is assessed. If an entity's score in step S2.2 reaches a preset threshold, it is considered successfully linked to the knowledge graph, and its coverage is recorded as 1. If the score does not reach the threshold, the knowledge graph is considered to have failed to cover the entity, and its coverage is recorded as 0. For entities with a coverage of 0, a diverse set of entity-oriented queries is constructed. The query set includes the following types of query expressions: entity name query, which uses the entity name as the basic search term to obtain the most direct entity-related information; explanatory query, which is used to obtain definitional and encyclopedic content; type extension query, which combines the entity with its type information; and context extension query, which combines the entity name with its context keywords. S3.3: The verification statement obtained in step S3.1 and the query set generated in step S3.2 for entities with insufficient knowledge coverage. Based on this, a unified candidate query set is constructed. The set includes two types of query items: one is a declaration-level query constructed from validation statements, and the other is a diversified query constructed from uncovered entities; S3.4: After obtaining the Top-K external search results, for each webpage result... A multi-dimensional credibility assessment was conducted, and the webpage's authority score was determined by... It is composed of a weighted combination of source credibility features; S3.5: After completing the webpage credibility assessment, each search evidence extract its main text content, and use the pre-training language BERT model for semantic coding to get evidence representation vector , using the credibility of different source evidence and content consistency, introducing uncertainty perception evidence aggregation mechanism, calculating the final dynamic knowledge enhancement vector.

[0014] The specific steps of knowledge consistency verification in step S4 are: S4.1: The key statements in the microblog text are respectively analyzed for semantic consistency with static knowledge enhancement information and dynamic search enhancement information, and the matching degree between the text content and multi-source knowledge is calculated; S4.2: According to the consistency analysis result, the knowledge credibility score reflecting the authenticity of microblog content is obtained, when there is obvious conflict between the text content and multi-source knowledge, the risk weight of being judged as rumor is increased; when part of the content lacks effective verification, its influence in the final discrimination is reduced; 6. The microblog rumor detection method based on static and dynamic knowledge enhancement according to claim 1, wherein the specific steps of multi-dimensional feature fusion in step S5 are: S5.1: Obtain multi-dimensional features from the foregoing steps, including text semantic features obtained in step S1 , static knowledge enhancement features obtained in step S2 , dynamic search enhancement features obtained in step S3 , and knowledge credibility score obtained in step S4 ; S5.2: Adopt content-knowledge fusion architecture to jointly represent learning of text semantic features and two types of knowledge enhancement features, first perform nonlinear transformation on feature vector to obtain content enhancement features.

[0015] The specific steps of constructing, training and rumor detection of the deep neural network classifier in step S6 are: S6.1: Build a deep neural network classifier, which adopts a multi-layer feature fusion architecture, combines BERT pre-training language model, knowledge graph embedding module, attention mechanism and multi-layer perception to build an end-to-end rumor detection network, and the network structure includes: text coding layer, knowledge enhancement layer, feature fusion layer and classification output layer, wherein the classification output layer adopts a multi-layer neural network as a classifier, and introduces an attention mechanism for feature importance modeling, and the classifier includes two layers of nonlinear conversion to obtain intermediate feature representation; S6.2: Use a loss function combining supervised learning and contrast learning, and add a knowledge constraint term to calculate the total loss; S6.3: During the training phase, the model parameters are updated using the joint loss function described in step S6.2, so that the model can simultaneously optimize classification performance, feature discriminativeness, and knowledge consistency constraints. S6.4: After training, the model enters the inference phase. For the input Weibo text, the model uses fused features. Perform forward computation and obtain the class probability vector through the Softmax output described in step S6.1. The component corresponding to the rumor category is used as the rumor probability. Furthermore, a confidence index is calculated based on the probability distribution to make a final judgment on the credibility of the Weibo content; S6.5: Input the test set into the final deep neural network to complete the rumor identification of Weibo text in the test set; compare the identification results with the real labels of each sample in the test set, and calculate the model's accuracy, precision, recall and F1 score.

[0016] A second aspect of the present invention provides a method and system for detecting Weibo rumors based on static and dynamic knowledge enhancement, comprising: The module includes a data preprocessing module, a static knowledge enhancement module, a dynamic search module, a knowledge consistency verification module, and a feature fusion and classification module. The data preprocessing module is used to load the Weibo dataset, clean the Weibo text, segment it, perform named entity recognition and semantic encoding, and generate text semantic features and entity information. The module includes a data preprocessing module, a static knowledge enhancement module, a dynamic search module, a knowledge consistency verification module, a multi-dimensional feature fusion module, and a deep neural network classifier construction and training module. The data preprocessing module is used to load the Weibo dataset, clean the Weibo text, segment it, perform named entity recognition and semantic encoding, and generate text semantic features and entity information. The static knowledge enhancement module is used to perform entity linking based on a pre-built knowledge graph and generate static knowledge enhancement features through knowledge graph embedding and attention aggregation mechanisms. The dynamic search module is used to generate queries for entities or statements not covered by the knowledge graph, perform real-time external searches, and generate dynamic search enhancement features based on source credibility and evidence consistency assessment. The knowledge consistency verification module is used to compare the consistency between the microblog text and static knowledge and dynamic evidence, calculate the knowledge credibility score, and analyze potential contradictions and uncertainties. The multi-dimensional feature fusion module is used to fuse text semantic features, static knowledge enhancement features, dynamic search enhancement features, and knowledge credibility scores to construct a multi-dimensional fusion representation.

[0017] The deep neural network classifier construction, training, and rumor detection module is used to build a deep neural network classifier, which uses multi-dimensional feature fusion representation to identify rumors in Weibo content and outputs the probability of rumors and the corresponding confidence level.

[0018] The working principle of this invention is as follows: First, the microblog text is preprocessed and entity recognition is performed. Then, knowledge enhancement is performed on known entities using a static knowledge graph. For content not covered by the knowledge graph, dynamic knowledge retrieval is performed through a search engine. Then, the credibility of the text content is evaluated through a knowledge consistency verification module. Finally, text semantics, knowledge enhancement, user behavior and dissemination features are integrated, and rumors are classified using a deep neural network.

[0019] The innovation of this invention lies in addressing the problem that traditional rumor detection methods lack external knowledge support and struggle to handle novel rumors. It proposes an innovative solution that integrates static knowledge graphs and dynamic search. By constructing a domain knowledge graph to provide a structured knowledge foundation and combining it with a real-time search mechanism to obtain the latest external evidence, multi-level knowledge verification of rumor content is achieved. The main innovations include: (1) A two-layer knowledge enhancement mechanism is proposed. Addressing the problem that existing rumor detection methods mainly rely on text content and dissemination patterns and lack a knowledge verification mechanism for the authenticity of rumor content, this invention innovatively proposes a two-layer knowledge enhancement method that integrates static knowledge graphs and dynamic search. This method first uses a pre-constructed domain knowledge graph to perform knowledge matching on text entities, extracting structured entity attributes and relational information; then, for content not covered by the knowledge graph, it performs dynamic knowledge retrieval through search engine APIs to obtain external evidence such as authoritative websites and news reports. This knowledge enhancement mechanism, combining static and dynamic approaches, ensures rapid matching of common entities and facts while adapting to newly emerging events and entities, significantly improving the knowledge coverage and timeliness of rumor detection.

[0020] (2) A credibility assessment method based on knowledge consistency verification is proposed. By calculating the semantic consistency between Weibo content, knowledge graph information, and search results, the credibility of the text content is quantitatively assessed. This method not only considers the matching degree between text and knowledge but also introduces logical reasoning and contradiction detection mechanisms, enabling it to identify common rumor characteristics such as temporal logic errors, numerical exaggerations, and factual distortions. Compared to traditional methods that rely solely on textual features, this invention can deeply verify rumors at the knowledge level, improving the interpretability and accuracy of the detection.

[0021] (3) A multi-dimensional feature fusion-based rumor detection framework was constructed. A hierarchical feature representation system was built by comprehensively utilizing textual semantic features, static knowledge enhancement features, and dynamic search features. Through the effective fusion of multi-source heterogeneous features using deep neural networks, a comprehensive modeling of rumors was achieved. Compared to single-dimensional detection methods, this invention can comprehensively judge from multiple perspectives such as content authenticity, knowledge consistency, user credibility, and anomalous propagation, significantly improving the robustness and generalization ability of rumor detection.

[0022] The beneficial effects of this invention are: it can effectively identify various rumors in social networks, especially complex rumors that require external knowledge to judge, providing important technical support for social media content review and public opinion supervision. Attached Figure Description

[0023] The accompanying drawings are provided to give a visual representation of the invention and, together with the embodiments of the invention, are used to explain the principles of the invention. They do not constitute a limitation on the scope of protection of the invention.

[0024] Figure 1 This is the overall framework diagram of the present invention; Figure 2 This is an example diagram of the static knowledge graph of the present invention; Figure 3 This is a flowchart of the dynamic search knowledge enhancement process of the present invention; Figure 4 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of protection of the present invention.

[0026] Example 1

[0027] See Figure 1 and Figure 4 This embodiment provides a Weibo rumor detection method based on static and dynamic knowledge enhancement. This method can be applied to the automatic detection and identification of rumors related to the COVID-19 pandemic, and is used to determine the authenticity of Weibo information involving content such as pandemic transmission, prevention and control measures, and treatment methods. The specific implementation steps of this method are as follows: S1: Data loading and preprocessing: Load the Weibo dataset, which contains Weibo text and its related attributes, and perform noise cleaning, word segmentation, named entity recognition and semantic encoding on the text to obtain text semantic features and entity information; S2: Static knowledge enhancement: Based on the entities obtained in step 1, entities are linked using a pre-built knowledge graph, and the attributes, relationships, and related facts of the entities are extracted. Static knowledge enhancement features are generated through knowledge graph embedding and attention aggregation mechanisms. S3: Dynamic Search Enhancement: For entities or key claims in text that fail to match in the knowledge graph, generate corresponding search queries and obtain relevant web page content through the search engine interface. Valid evidence is filtered based on source credibility and semantic consistency to construct dynamic search enhancement features. S4: Knowledge Consistency Verification: Combining static knowledge facts with dynamic search evidence, the consistency of text content is evaluated. Based on semantic matching, relation verification and logical reasoning, a knowledge credibility score is calculated to reflect the degree of conformity between the text and multi-source knowledge. S5: Multi-dimensional feature fusion: This method integrates text semantic features, static knowledge enhancement features, dynamic search enhancement features, and knowledge credibility scores to construct a multi-dimensional fusion representation. S6: Construction, training and rumor detection of deep neural network classifiers: Build a deep neural network classifier, use multi-dimensional fusion representation, use the deep neural network classifier to identify rumors in Weibo content, and output the probability of rumors and the corresponding confidence level.

[0028] Furthermore, the specific steps of data preprocessing in step S1 are as follows: S1.1: Load the Weibo dataset. The dataset can be a publicly available Weibo rumor dataset or a dataset built based on keywords / events, containing... Each record contains fields such as Weibo ID, text content, posting time, user information (number of followers, verification status, etc.), number of reposts, number of comments, and tags (true / false, tags can be obtained based on manual annotation or aligned with authoritative debunking sources); S1.2: Clean the text content fields to remove noise information such as URL links, special characters, and emoticons, while retaining the core text content; S1.3: Use one of the word segmentation and part-of-speech tagging tools such as Jieba / THULAC / LTP / HanLP to perform word segmentation and part-of-speech tagging on the cleaned text; S1.4: Use BiLSTM-CRF or BERT-CRF models to identify entities such as people's names, place names, organization names, time, and numbers in the text, and record the entity text, category, and location for subsequent knowledge alignment; S1.5: Use a pre-trained language BERT model (Chinese BERT-base or a similar pre-trained model) to process the original text. Encode the text to obtain the semantic representation vector of the original text. :

[0029] in This is the hidden state vector at position [CLS] output by the BERT model. The vector dimension (usually 768 or 1024) captures the overall semantic information of the text, serving as the basis for subsequent feature fusion.

[0030] Furthermore, the specific steps of static knowledge enhancement described in step S2 are as follows: S2.1: Load the pre-built domain knowledge graph ,in It is a collection of entities, and is not required to cover all entities appearing in the text. For a set of relations, For the set of triples , indicating the head entity Through relationships With tail entity For text entities that cannot find a corresponding node in the knowledge graph, they may not participate in subsequent knowledge enhancement steps. S2.2: For each entity identified in step S1.4 In the knowledge graph, entity linking is performed by mapping text entities to standard entity nodes in the knowledge graph through methods such as string matching and semantic similarity calculation, and defining entity link scores:

[0031] in For text entities, Candidate entities for knowledge graph The calculation formula is , For string differences based on Levenshtein edit distance; The calculation formula is , For the contextual semantic representation of entities in text, Aliases for entities in a knowledge graph or semantic representations of descriptive text can both be obtained by encoding pre-trained language BERT models; , The entity type obtained during the named entity recognition process in step S1.4 is the text entity. For knowledge graph entities in construction The type label defined at that time. are weight parameters and Select the candidate entity with the highest score as the link result:

[0032] S2.3: For successfully linked entities, extract their multi-hop neighbor information in the knowledge graph, and obtain the entity's [unclear - possibly related to a random walk strategy]. Skip Neighborhood: in To maximize the number of hops, knowledge path sampling employs a biased random walk, prioritizing relationships with high importance.

[0033] in For entities The set of outgoing edge relations, For relationship In the map Importance within; S2.4: Loaded domain knowledge graph Each entity in and relationships Vectorization is performed, and their respective definitions are defined. 3D embedding vectors yield entity vectors With relation vector This vector can support hierarchical knowledge graph embedding methods, combined with the translation properties of TransE (relation vectors). Head entity vector Translate to tail entity vector The neighborhood of RotatE and the rotation properties (relation vectors) For entity vectors Apply phase rotation to make it close to Construct a hybrid embedding loss function: in:

[0034] in For the negative sample set, For Hadamard product, To embed entities into a projection onto their type subspace The operation, To balance the hyperparameters; S2.5: Targeting text containing... Entities that are successfully linked Utilizing its embedding vector With text semantic vectors Knowledge attention aggregation is performed to obtain a static knowledge-enhanced representation of the text:

[0035] Attention weight Calculated via gating mechanism:

[0036] in Meta-features of entities (such as entity type, frequency of occurrence, node degree, etc.). For element-wise multiplication, For the Sigmoid function, As learnable parameters, the model uses an attention aggregation mechanism to assign different weights to different entities based on text semantics, giving higher contributions to entities more relevant to rumor detection. The final result is... As a static knowledge augmentation vector, it will be used in subsequent steps along with text semantic representation. To integrate.

[0037] Furthermore, the specific steps of the dynamic search enhancement described in step S3 are as follows: S3.1: After extracting claims from the Weibo text, to avoid unnecessary external searches for unverifiable content, the verifiability of each claim is determined. A dependency parsing-based claim extraction method is used to extract claims from the Weibo text. Structured declarations are identified in Chinese:

[0038] in As the main body, As a predicate, As an object, For a set of modifying components, the verifiability of a declaration is defined as follows: in As an indicator function, the fact-judgment predicate can be determined based on a predefined set of predicates or a classification model. S3.2: Based on the verifiable statement obtained in step S3.1, examine the entities involved in the statement. Knowledge coverage is assessed. If an entity's score in step S2.2 reaches a preset threshold, it is considered successfully linked to the knowledge graph, and its coverage is recorded as 1. If the score does not reach the threshold, the knowledge graph is considered to have failed to cover the entity, and its coverage is recorded as 0. For entities with a coverage of 0, a diverse set of entity-oriented queries is constructed. The query set includes the following types of query expressions: entity name query, which uses the entity name as the basic search term to obtain the most direct entity-related information; explanatory query, such as " "What is it?" is used to retrieve descriptive and encyclopedic content; type-extended queries combine entities with their type information, such as " ",in Category attributes for entities can be obtained through named entity recognition or entity linking processes; context-expanded queries combine entity names with their contextual keywords, such as " ",in Keywords extracted from the context of the sentence containing the entity can be obtained through fixed window extraction, dependency parsing, or keyword extraction methods based on TF-IDF / TextRank. S3.3: The verifiable claims obtained in step S3.1 and the query set generated in step S3.2 for entities with insufficient knowledge coverage. Based on this, a unified candidate query set is constructed. The set includes two types of query terms: one is declaration-level queries constructed from verifiable declarations, and the other is diversified queries constructed from entities not covered, namely:

[0039] in This represents a single query expression obtained from any of the above sources, which is the basic unit for priority sorting in this step. All query items in Its importance score is calculated as follows:

[0040] in, The centrality of the query in the original text can be determined based on its syntactic position, word frequency weight, or semantic relevance. The uncertainty of the model's prediction of query-related content can be calculated using the entropy value or confidence index of the output probability distribution. The potential controversy reflected in the query can be estimated by the degree of contradiction between historical information sources, sentiment polarization characteristics, or inconsistencies in related reports. All query items are then ranked according to the aforementioned importance scores, and [the following is selected]. The top-K queries with the highest scores are used as external retrieval targets; S3.4: After obtaining the Top-K external search results, for each webpage result... A multi-dimensional credibility assessment was conducted, and the webpage's authority score was determined by... It is composed of a weighted combination of source credibility features, defined as follows: in, These represent characteristics such as the domain name of a government agency or university, the source of mainstream media, the overall ranking of the webpage, the implementation of security protocols, and the domain registration age. The corresponding weighting coefficients can be obtained through prior settings or model learning, based on the probability of website fraud. The original authority score is adjusted to obtain the calibrated credibility score: in It can be obtained through website reputation databases, URL security detection models, or other risk identification algorithms, and is used to identify potentially false or misleading sources; S3.5: After completing the webpage credibility assessment, analyze each piece of search evidence. The main text content is extracted and semantically encoded using a pre-trained BERT language model to obtain the evidence representation vector. By leveraging the credibility and consistency of evidence from different sources, an uncertainty-aware evidence aggregation mechanism is introduced to calculate the final dynamic knowledge enhancement vector: Among them, the adjusted weights Taking into account the credibility of the evidence and its consistency with other highly credible evidence, the definition is as follows:

[0041] Each piece of evidence is assigned an initial weight. (or This can be set based on the webpage credibility score in step S3.4, ensuring consistency. This indicates the degree of semantic consistency between a piece of evidence and a set of highly credible evidence. It is calculated by comparing the evidence with the set of highly credible evidence. The average similarity of the evidence in the two cases, where: Among them, threshold Used to screen for evidence with higher credibility, if If empty, the entire evidence set can be used as a substitute to ensure the stability of the aggregation process.

[0042] Furthermore, the specific steps of knowledge consistency verification described in step S4 are as follows: S4.1: Based on the entity links and knowledge graph facts obtained in step S2, evaluate the consistency between the current Weibo text and the knowledge graph. Let the set of entities identified in the current text be... The set of entity relations extracted from the text is The verifiable set is Consistency scores are calculated at three levels: entity, relation, and logic, and a comprehensive knowledge graph consistency score is obtained through a weighted average. Among them: entity-level consistency Used to characterize the degree of matching between the textual description of an entity and the relevant facts about that entity in the knowledge graph; relation-level consistency. Used to characterize whether the relationships between entity pairs in a text correspond to a set of triples in a knowledge graph. Consistency; Logical consistency The natural language reasoning model is used to determine the supporting, neutral, or contradictory relationships between textual statements and facts that can be inferred from the knowledge graph, and these relationships are mapped to numerical scores. The three consistency scores are obtained by weighted summation. in These are the weighting coefficients; S4.2: Based on the results obtained in step S3 Search evidence and the corresponding evidence vector representation And credibility weight Calculate the consistency score between the textual statement and the search evidence, and let... This is the vector representation of the current declaration to be verified. For vector-based semantic similarity (e.g., cosine similarity), the dynamic search consistency score is defined as:

[0043] in The initial credibility weight of the evidence is set based on the webpage credibility score in step S3.4; S4.3: Introduce an adaptive fusion strategy to balance the importance of static knowledge graphs and dynamic search evidence. The number of entities successfully linked to the knowledge graph (from S2.2). If the total number of entities in the current text is given, then the knowledge graph coverage rate is defined as follows: The consistency score of static knowledge graphs Consistency score with dynamic search The results are then combined to obtain the final knowledge credibility score:

[0044] The score It serves as a knowledge credibility feature in subsequent feature fusion and classification stages, and is used to construct knowledge constraint terms in the loss function.

[0045] Furthermore, the specific steps of multi-dimensional feature fusion described in step S5 are as follows: S5.1: Obtain multi-dimensional features from the preceding steps, including the text semantic features obtained in step S1. The static knowledge enhancement features obtained in step S2 The dynamic search enhancement features obtained in step S3 And the knowledge credibility score obtained in step S4 ; S5.2: Employing a content-knowledge fusion architecture, this method jointly learns representations of textual semantic features and two types of knowledge-enhanced features. First, it processes the feature vectors... Perform a nonlinear transformation to obtain content enhancement features:

[0046] in This represents a multilayer perceptron, with the symbols [·;·] indicating vector concatenation. The knowledge credibility score is then concatenated with the content enhancement features to obtain the final fused features.

[0047] Furthermore, the specific steps for constructing, training, and detecting rumors using the deep neural network classifier described in step S6 are as follows: S6.1: Construct a deep neural network classifier. This classifier employs a multi-layer feature fusion architecture, combining a BERT pre-trained language model, a knowledge graph embedding module, an attention mechanism, and a multi-layer perceptron to build an end-to-end rumor detection network. The network structure includes: a text encoding layer, a knowledge enhancement layer, a feature fusion layer, and a classification output layer. The classification output layer uses a multi-layer neural network as the classifier and introduces an attention mechanism for feature importance modeling. The classifier includes two non-linear transformation layers to obtain intermediate feature representations. Apply attention weights to intermediate features:

[0048] Features are weighted according to attention weights:

[0049] Final output: Probability of being a rumor

[0050] Among the symbols This represents element-wise multiplication and the weights. It can determine the contribution of different features to the final judgment; Table 1 BERT Text Encoder Network Parameters

[0051] Table 2 Parameters of Knowledge Graph Embedding Module

[0052] Table 3. Parameters of Multi-Dimensional Feature Fusion and Deep Neural Network Classification

[0053] S6.2: A loss function combining supervised learning and contrastive learning is adopted, with a knowledge constraint term added. The total loss is defined as: in, For cross-entropy loss, The comparative learning loss is used to bring the feature representations of similar samples closer together and to distinguish between dissimilar samples. The knowledge consistency constraint loss is used to encourage the model output to align with the knowledge credibility score. Consistent; S6.3: During the training phase, the model parameters are updated using the joint loss function described in step S6.2, enabling the model to simultaneously optimize classification performance, feature discriminativeness, and knowledge consistency constraints. AdamW is selected as the optimizer, with an initial learning rate set to 2e-5 and a weight decay coefficient of 0.01. A learning rate scheduling strategy combining linear warm-up and linear decay is adopted, with the learning rate warm-up ratio set to 0.1 (i.e., the first 10% of training steps are used for learning rate warm-up). During training, an early stopping strategy is used to prevent overfitting. The patience parameter M is set to 10, meaning that if the F1 score on the validation set does not improve for 10 consecutive rounds, the network is considered to have converged, training is stopped, and the model parameters with the highest F1 score on the validation set are saved as the final model; otherwise, the network parameters are updated and training returns to step 3.1 to continue until the maximum number of training rounds (10 rounds) is reached or the early stopping condition is met.

[0054] S6.4: After training, the model enters the inference phase. For the input Weibo text, the model uses fused features. Perform forward computation and obtain the class probability vector through the Softmax output described in step S6.1. The component corresponding to the rumor category is used as the rumor probability. Furthermore, a confidence index can be calculated based on the probability distribution to make a final judgment on the credibility of Weibo content.

[0055] S6.5: Input the test set into the final deep neural network to complete the rumor identification of Weibo text in the test set. Compare the identification results with the true labels of each sample in the test set, and calculate the model's accuracy, precision, recall, and F1 score. On the publicly available Weibo rumor dataset, the network achieved an average identification accuracy of 92.3% and an F1 score of 91.8%, effectively improving the accuracy of Weibo rumor detection.

[0056] In the public health and safety scenario of "COVID-19 rumor detection," this embodiment can integrate static knowledge, dynamic search, knowledge consistency verification, and multi-feature information to effectively identify various types of medical and health rumors, and has significant application value and promotion potential.

[0057] Example 2

[0058] This embodiment provides a system for implementing the Weibo rumor detection method based on static and dynamic knowledge enhancement described in Embodiment 1, comprising: The system includes a data preprocessing module, a static knowledge enhancement module, a dynamic search module, a knowledge consistency verification module, and a feature fusion and classification module.

[0059] Data preprocessing module: Used to load Weibo datasets, perform text cleaning, word segmentation, part-of-speech tagging and named entity recognition, and output structured text and entity lists.

[0060] Static knowledge enhancement module: Used to load domain knowledge graph, perform entity linking and knowledge matching on identified entities, extract attributes, relationships and facts, and generate static knowledge enhancement vectors.

[0061] Dynamic Search Module: Used to process entities and key claims not covered by the knowledge graph, obtain external evidence through a search engine, and generate dynamic search enhancement vectors.

[0062] Knowledge Consistency Verification Module: Used to evaluate the semantic consistency between Weibo content and static knowledge and dynamic evidence, and obtains a knowledge credibility score through similarity calculation, fact verification and logical reasoning.

[0063] Multi-dimensional feature fusion module: used to fuse text semantic features, static knowledge enhancement features, dynamic search enhancement features and knowledge credibility scores to construct a multi-dimensional fusion representation.

[0064] The module for building, training, and detecting rumors in deep neural network classifiers is used to build deep neural network classifiers. It utilizes multi-dimensional feature fusion representations to identify rumors in Weibo content and outputs the probability of a rumor and its corresponding confidence level.

[0065] The above embodiments are merely specific implementations of the present invention, and the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or improvements made by those skilled in the art within the scope disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A Weibo rumor detection method based on static and dynamic knowledge enhancement, characterized in that, Includes the following steps: S1: Data loading and preprocessing: Load the Weibo dataset, which contains Weibo text and its related attributes, and perform noise cleaning, word segmentation, named entity recognition and semantic encoding on the text to obtain text semantic features and entity information; S2: Static knowledge enhancement: Based on the entities obtained in step 1, entities are linked using a pre-built knowledge graph, and the attributes, relationships, and related facts of the entities are extracted. Static knowledge enhancement features are generated through knowledge graph embedding and attention aggregation mechanisms. S3: Dynamic Search Enhancement: For entities or key claims in text that fail to match in the knowledge graph, generate corresponding search queries and obtain relevant web page content through the search engine interface. Valid evidence is filtered based on source credibility and semantic consistency to construct dynamic search enhancement features. S4: Knowledge Consistency Verification: Combining static knowledge facts with dynamic search evidence, the consistency of text content is evaluated. Based on semantic matching, relation verification and logical reasoning, a knowledge credibility score is calculated to reflect the degree of conformity between the text and multi-source knowledge. S5: Multi-dimensional feature fusion: This method integrates text semantic features, static knowledge enhancement features, dynamic search enhancement features, and knowledge credibility scores to construct a multi-dimensional fusion representation. S6: Construction, training and rumor detection of deep neural network classifiers: Build a deep neural network classifier, use multi-dimensional fusion representation, use the deep neural network classifier to identify rumors in Weibo content, and output the probability of rumors and the corresponding confidence level.

2. The Weibo rumor detection method based on static and dynamic knowledge enhancement according to claim 1, characterized in that, The specific steps of data preprocessing described in step S1 are as follows: S1.1: Load the Weibo dataset. The dataset can be a publicly available Weibo rumor dataset or a dataset constructed by collecting data based on keywords / events, containing... Each record contains fields such as Weibo ID, text content, posting time, user information, number of reposts, number of comments, and tags. S1.2: Clean the text content fields to remove noise information such as URL links, special characters, and emoticons, while retaining the core text content; S1.3: One of the word segmentation and part-of-speech tagging tools, Jieba / THULAC / LTP / HanLP, is used to segment the cleaned text and perform part-of-speech tagging. S1.4: Use BiLSTM-CRF or BERT-CRF models to identify entities such as personal names, place names, organization names, time, and numbers in the text, and record the entity text, category, and location for subsequent knowledge alignment; S1.5: Using a pre-trained language BERT model to process the original text Encoding is performed to obtain a semantic representation vector of the original text. This vector captures the overall semantic information of the text and serves as the basis for subsequent feature fusion.

3. The Weibo rumor detection method based on static and dynamic knowledge enhancement according to claim 1, characterized in that, The specific steps of static knowledge enhancement described in step S2 are as follows: S2.1: Load the pre-built domain knowledge graph ,in It is a collection of entities, and is not required to cover all entities appearing in the text. For a set of relations, For the set of triples , indicating the head entity Through relationships With tail entity For text entities that cannot find a corresponding node in the knowledge graph, they will not participate in subsequent knowledge enhancement steps. S2.2: For each entity identified in step S1.4 In the knowledge graph, entity linking is performed by mapping text entities to standard entity nodes in the knowledge graph through methods such as string matching and semantic similarity calculation, defining entity link scores, and selecting the candidate entity with the highest score as the link result. S2.3: For successfully linked entities, extract their multi-hop neighbor information in the knowledge graph, and obtain the entity's [unclear - possibly related to a random walk strategy]. Jump to the neighbor; S2.4: Loaded domain knowledge graph Each entity in and relationships Vectorization is performed, and their respective definitions are defined. 3D embedding vectors yield entity vectors With relation vector This vector supports a hierarchical knowledge graph embedding method, which combines the translation properties of TransE and the rotation properties of RotatE to construct a hybrid embedding loss function. S2.5: Targeting text containing... Entities that are successfully linked Utilizing its embedding vector With text semantic vectors By performing knowledge attention aggregation, a static knowledge-enhanced representation of the text is obtained. Through the attention aggregation mechanism, the model can assign different weights to different entities based on the semantics of the text, giving higher contributions to entities more relevant to rumor detection. As a static knowledge augmentation vector, it will be used in subsequent steps along with text semantic representation. To integrate.

4. The Weibo rumor detection method based on static and dynamic knowledge enhancement according to claim 1, characterized in that, The specific steps of the dynamic search enhancement described in step S3 are as follows: S3.1: After extracting claims from the Weibo text, to avoid unnecessary external searches for unverifiable content, the verifiability of each claim is determined. A dependency parsing-based claim extraction method is used to extract claims from the Weibo text. Identify structured declarations in the text; S3.2: Based on the verifiable statement obtained in step S3.1, examine the entities involved in the statement. Knowledge coverage is assessed. If an entity's score in step S2.2 reaches a preset threshold, it is considered successfully linked to the knowledge graph, and its coverage is recorded as 1. If the score does not reach the threshold, the knowledge graph is considered to have failed to cover the entity, and its coverage is recorded as 0. For entities with a coverage of 0, a diverse set of entity-oriented queries is constructed. The query set includes the following types of query expressions: entity name query, which uses the entity name as the basic search term to obtain the most direct entity-related information; Explanatory queries are used to retrieve defining or encyclopedic content. Type-based extended queries combine an entity with its type information; context-based extended queries combine an entity name with its context keywords. S3.3: The verification statement obtained in step S3.1 and the query set generated in step S3.2 for entities with insufficient knowledge coverage. Based on this, a unified candidate query set is constructed. The set includes two types of query items: one is a declaration-level query constructed from validation statements, and the other is a diversified query constructed from uncovered entities; S3.4: After obtaining the Top-K external search results, for each webpage result... A multi-dimensional credibility assessment was conducted, and the webpage's authority score was determined by... It is composed of a weighted combination of source credibility features; S3.5: After completing the webpage credibility assessment, analyze each piece of search evidence. The main text content is extracted and semantically encoded using a pre-trained BERT language model to obtain the evidence representation vector. By leveraging the credibility and consistency of evidence from different sources, an evidence aggregation mechanism based on uncertainty perception is introduced to calculate the final dynamic knowledge enhancement vector.

5. The Weibo rumor detection method based on static and dynamic knowledge enhancement according to claim 1, characterized in that, The specific steps of knowledge consistency verification described in step S4 are as follows: S4.1: Perform semantic consistency analysis on the key statements in the Weibo text with static knowledge enhancement information and dynamic search enhancement information respectively, and calculate the degree of matching between the text content and multi-source knowledge; S4.2: Based on the consistency analysis results, obtain the knowledge credibility score that reflects the authenticity of the Weibo content. When there is a significant conflict between the text content and multi-source knowledge, increase the risk weight of being judged as a rumor; when some content lacks effective verification, reduce its impact in the final judgment.

6. The Weibo rumor detection method based on static and dynamic knowledge enhancement according to claim 1, characterized in that, The specific steps of multi-dimensional feature fusion described in step S5 are as follows: S5.1: Obtain multi-dimensional features from the preceding steps, including the text semantic features obtained in step S1. The static knowledge enhancement features obtained in step S2 The dynamic search enhancement features obtained in step S3 And the knowledge credibility score obtained in step S4 ; S5.2: Employing a content-knowledge fusion architecture, this method jointly learns representations of textual semantic features and two types of knowledge-enhanced features. First, it processes the feature vectors... A nonlinear transformation is performed to obtain content enhancement features.

7. The Weibo rumor detection method based on static and dynamic knowledge enhancement according to claim 1, characterized in that, The specific steps for constructing, training, and detecting rumors using the deep neural network classifier described in step S6 are as follows: S6.1: Construct a deep neural network classifier. The deep neural network classifier adopts a multi-layer feature fusion architecture, combining a BERT pre-trained language model, a knowledge graph embedding module, an attention mechanism, and a multi-layer perceptron to build an end-to-end rumor detection network. The network structure includes: a text encoding layer, a knowledge enhancement layer, a feature fusion layer, and a classification output layer. The classification output layer uses a multi-layer neural network as the classifier and introduces an attention mechanism for feature importance modeling. The classifier includes two non-linear transformation layers to obtain intermediate feature representations. S6.2: Use a loss function that combines supervised learning and contrastive learning, and add a knowledge constraint term to calculate the total loss. S6.3: During the training phase, the model parameters are updated using the joint loss function described in step S6.2, so that the model can simultaneously optimize classification performance, feature discriminativeness, and knowledge consistency constraints. S6.4: After training, the model enters the inference phase. For the input Weibo text, the model uses fused features. Perform forward computation and obtain the class probability vector through the Softmax output described in step S6.

1. The component corresponding to the rumor category is used as the rumor probability. Furthermore, a confidence index is calculated based on the probability distribution to make a final judgment on the credibility of the Weibo content; S6.5: Input the test set into the final deep neural network to complete the rumor identification of Weibo text in the test set; compare the identification results with the real labels of each sample in the test set, and calculate the model's accuracy, precision, recall and F1 score.

8. A system for implementing the Weibo rumor detection method based on static and dynamic knowledge enhancement as described in any one of claims 1-7, comprising: The module includes a data preprocessing module, a static knowledge enhancement module, a dynamic search module, a knowledge consistency verification module, a multi-dimensional feature fusion module, and a deep neural network classifier construction and training module. The data preprocessing module is used to load the Weibo dataset, clean the Weibo text, segment it, perform named entity recognition and semantic encoding, and generate text semantic features and entity information. The static knowledge enhancement module is used to perform entity linking based on a pre-built knowledge graph and generate static knowledge enhancement features through knowledge graph embedding and attention aggregation mechanisms. The dynamic search module is used to generate queries for entities or statements not covered by the knowledge graph, perform real-time external searches, and generate dynamic search enhancement features based on source credibility and evidence consistency assessment. The knowledge consistency verification module is used to compare the consistency between the microblog text and static knowledge and dynamic evidence, calculate the knowledge credibility score, and analyze potential contradictions and uncertainties. The multi-dimensional feature fusion module is used to fuse text semantic features, static knowledge enhancement features, dynamic search enhancement features, and knowledge credibility scores to construct a multi-dimensional fusion representation. The deep neural network classifier construction, training, and rumor detection module is used to build a deep neural network classifier, which uses multi-dimensional feature fusion representation to identify rumors in Weibo content and outputs the probability of rumors and the corresponding confidence level.

Images