Unsupervised unreal information detection method based on group wisdom and content cross inspiration

Abstract

The invention belongs to the technical field of unreal information detection, and particularly relates to an unsupervised unreal information detection method based on group intelligence and content cross inspiration. Comprising the steps that blog articles and comment data are acquired from a social platform, and a blog article-comment interaction network is constructed; constructing prompt words according to the blog article-comment interaction network; according to the cue word, processing the blog article and the comment data by adopting a large language model to obtain an attitude division result of the comment; according to an attitude division result, quantifying an attitude vector of the blog article; semantic embedding is extracted from the blog article; extracting a semantic center vector from the comments corresponding to the blog article according to the attitude division result of the comments; fusing the attitude vector of the blog article, the semantic embedding and the semantic center vector of the comment corresponding to the blog article to obtain comprehensive feature representation of the blog article; performing clustering processing on all the blog articles according to the comprehensive feature representation of each blog article to obtain an unreal information detection result; the method can effectively overcome the limitation of a supervised learning method on unknown unreal information identification, and improves the accuracy of an unreal information detection result.

Description

Technical Field

[0001] This invention belongs to the field of misinformation detection technology, specifically involving an unsupervised misinformation detection method based on collective intelligence and content cross-inspiration. Background Technology

[0002] With the widespread use of social media platforms in daily life, the efficiency of information dissemination has significantly improved, but it has also provided fertile ground for the breeding and spread of misinformation. Compared to traditional media, while social media provides users with convenient access to news and diverse information, the lack of effective fact-checking mechanisms has made it a breeding ground for unverified and false information. Compared to general misinformation, online misinformation is characterized by faster spread, wider impact, and deeper social harm. Misinformation typically refers to unverified, false, or misleading information spread through social media channels such as Weibo. For example, the false information that "the new medical insurance policy in 2024 stipulates that employee medical insurance contributions will no longer be credited to personal accounts" appeared in early 2024. While triggering large-scale online discussions, some false details caused public anxiety, which was only quelled after the relevant authorities clarified the situation. Such cases fully demonstrate that misinformation not only creates social panic and weakens institutional credibility, but may also exacerbate social divisions. Therefore, misinformation detection technology on social networks is receiving increasing attention.

[0003] In the current context of information overload and the proliferation of misinformation, how to help the public effectively distinguish between true and false information, reach consensus, and make rational decisions has become a core issue that urgently needs to be addressed. This prompts us to focus on the key concept of collective intelligence. In human society, collective intelligence is typically manifested through democratic decision-making, crowdsourcing collaboration, expert networks, and wiki-like platforms. Take Wikipedia as an example: this knowledge base, collaboratively built by tens of thousands of volunteers worldwide, has gained widespread recognition for its content quality and authority through continuous content editing, mutual review, and knowledge correction, even surpassing the level of traditional encyclopedias in some areas. In the field of organizational management, cross-functional teams, by integrating the professional knowledge and experience perspectives of members from different backgrounds such as technology, marketing, and management, can form more comprehensive and in-depth problem solutions, significantly improving the quality and execution of strategic decisions.

[0004] Large language models have seen rapid development in the field of Natural Language Processing (NLP). Since the application of neural networks and deep learning in NLP, language models have made significant progress. Their core principle is to learn the statistical regularities and semantic knowledge of human language through pre-training on massive amounts of text data, forming implicit representations that include grammatical structure, semantic relationships, and world knowledge. Large language models employ a self-supervised learning paradigm, capturing distributed representations of vocabulary, syntax, and semantics in large-scale corpora through pre-training tasks such as masked language modeling or autoregressive generation. Some researchers have begun exploring the use of large language models to improve the accuracy of misinformation detection. Directly analyzing content using NLP techniques and examining lexical, syntactic, sentiment, and semantic features can identify misinformation more quickly before it spreads widely, which is an advantage of using large language models for misinformation detection. However, research shows that large language models primarily detect misinformation through common-sense reasoning rather than utilizing prior knowledge of current news facts. Since misinformation often lacks sufficient prior knowledge in its early stages, relying solely on large language models to judge its authenticity may lead to a high risk of misjudgment. Therefore, how to effectively and scientifically use large language models for misinformation analysis still requires further research.

[0005] In recent years, significant progress has been made in the research of misinformation detection on social networks. Existing research mainly focuses on the following directions: some studies analyze the structural characteristics of misinformation dissemination and use graph structure methods to integrate contextual and structural information; others integrate text and structural features for representation learning to address data sparsity issues; and still others conduct fine-grained analysis from dimensions such as user characteristics, content features, lexical features, and post sequences. Our observations show that user comments often reflect their attitude towards the authenticity of posts, especially when posts contain misinformation, where comments often contain questioning or opposing voices. Furthermore, misinformation posts typically exhibit significant linguistic features such as exaggerated expressions, emotional language, ambiguous referencing, and logical contradictions, indicating that analyzing the correlation between post text and comments can improve the effectiveness of misinformation identification.

[0006] In summary, current misinformation detection models still have the following problems:

[0007] There is a lack of collective cognitive judgment mechanisms. Users from different backgrounds have different understandings of the same information, and how to effectively integrate the diverse perspectives provided by collective wisdom remains a challenge.

[0008] There is a lack of direct analysis of the text itself. Existing technologies mostly focus on the analysis of social network structure and propagation processes, but as the methods of spreading misinformation become increasingly diversified, these methods have shown limitations.

[0009] Research on supervised detection methods is weak. Compared to supervised learning methods that require a large amount of labeled data, unsupervised methods can still identify false information patterns even with limited training data, and have better adaptability and generalization ability to unknown false information.

[0010] In conclusion, there is an urgent need for a new unsupervised method for detecting misinformation in order to improve the effectiveness of misinformation identification. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this invention proposes an unsupervised method for detecting false information based on collective intelligence and content cross-heuristics. This method includes:

[0012] S1: Obtain blog post and comment data from social media platforms, and construct a blog post-comment interaction network based on the blog post and comment data;

[0013] S2: Construct cue words based on the blog post-comment interaction network; based on the cue words, use a large language model to process the blog post and comment data to obtain the attitude classification results of the comments;

[0014] S3: Quantify the attitude vector of blog posts based on the attitude segmentation results of comment content;

[0015] S4: Extract semantic embeddings from blog posts; extract semantic center vectors from the comments corresponding to the blog posts based on the attitude segmentation results of the comments;

[0016] S5: By integrating the attitude vector, semantic embedding, and semantic center vector of the blog post's corresponding comments, a comprehensive feature representation of the blog post is obtained;

[0017] S6: Cluster all blog posts based on the comprehensive feature representation of each post to obtain the false information detection results.

[0018] Preferably, the process of building a blog post-comment interaction network includes:

[0019] Treating both blog posts and comments as nodes, if there is a propagation relationship between blog post nodes, then there is an edge between the two nodes; if a comment belongs to a blog post, then there is an edge between the corresponding comment node and the blog post node.

[0020] Preferably, step S3 specifically includes:

[0021] The proportion of each attitude comment was obtained based on the attitude classification results;

[0022] Set the weight for each attitude comment;

[0023] By combining the weight of each attitude comment with its corresponding proportion, an attitude vector for the blog post is obtained; among them, the attitude of the comments is divided into support, neutrality, and skepticism.

[0024] Furthermore, the weight of each attitude and its corresponding proportion are expressed as follows:

[0025]

[0026] in, This represents the attitude vector of blog post i. , and These represent the weights of supportive, neutral, and critical comments, respectively. , and The percentages of comments that support, are neutral, and are critical are represented respectively.

[0027] Preferably, the process of extracting semantic center vectors from the comments corresponding to blog posts based on the attitude segmentation results includes:

[0028] The comments corresponding to the blog post are embedded to obtain the comment embedding vector;

[0029] Based on the attitude segmentation results of the comments, the comment embedding vectors of the same attitude are aggregated to obtain the comment embedding vector sum of each attitude comment;

[0030] The comment embedding vector of the attitude comment and the semantic center vector of the corresponding attitude category are used.

[0031] Preferably, the method of integrating the attitude vector of the blog post, the semantic embedding, and the semantic center vector of the corresponding comment is concatenation.

[0032] Preferably, the process of clustering all blog posts based on the comprehensive feature representation of each blog post includes:

[0033] Blog posts with typical characteristics of misinformation were selected from the blog post collection as beacon posts;

[0034] Calculate the similarity between the composite feature representation of all blog posts and the composite feature representation of the selected beacon blog posts;

[0035] Based on similarity, an unsupervised clustering algorithm is used to process all blog posts to obtain clustering results; the cluster containing the beacon blog posts in the clustering results is taken as the false information cluster, i.e., the false information detection result.

[0036] The beneficial effects of this invention are as follows:

[0037] This invention innovatively integrates the analytical capabilities of collective intelligence and large language models, proposing an unsupervised method for detecting misinformation that requires no manual annotation and possesses strong generalization ability. By constructing a multi-dimensional post-comment interaction network, it deeply mines the collective cognitive features contained in user comments, achieving a comprehensive evaluation of information authenticity from semantic, sentiment, and structural dimensions. Compared with traditional detection methods that rely on labeled data or single text features, this invention effectively overcomes the limitations of supervised learning methods in identifying unknown misinformation, significantly improving the adaptability and scalability of the detection model.

[0038] This method utilizes a large language model to perform fine-grained sentiment analysis on comment content, accurately identifying attitudes such as support, skepticism, and neutrality. By fusing quantified attitude vectors and semantic embedding vectors, a comprehensive representation reflecting the characteristics of information dissemination is constructed. Employing an unsupervised clustering algorithm based on semantic similarity, using typical posts containing false information as beacons, it achieves automated false information detection on large-scale social data, avoiding the dependence on labeled data in traditional methods.

[0039] The application of this invention can significantly improve the accuracy and efficiency of misinformation detection, outperforming many existing unsupervised methods on public datasets, and even surpassing some supervised learning methods in specific scenarios. It exhibits particularly outstanding detection performance in the Chinese language context, providing effective technical support for content governance on social media platforms and a cleaner cyberspace. This method has strong engineering application value, adaptable to different language environments and diverse social media platform scenarios, and provides an innovative solution for building a trustworthy network environment. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the unsupervised false information detection method based on collective intelligence and content cross-inspiration in this invention.

[0041] Figure 2 This figure shows the performance comparison results of the present invention and the comparison model on the CHECKED and PHEME datasets. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] This invention proposes an unsupervised method for detecting misinformation based on collective intelligence and content cross-heuristics. The method includes the following:

[0044] Observations reveal that tweets (posts) from different users on the same topic share significant commonalities in language style, emotional tone, and keyword frequency. For example, in the Chinese virus misinformation dataset used in this experiment, user tweets on topics related to education and lifestyle adjustments during a specific period often expressed concern or anticipation; while on topics related to the international pandemic and medical resources, they tended to express worry and anxiety. These commonalities reflect users' social behavior patterns on specific topics, and the formation of these patterns may be influenced by the nature of the topic and user interactions.

[0045] Meanwhile, tweets and their comments are closely linked, with tweet content often directly influencing the tone, emotion, and diversity of viewpoints in the comments. For example, neutral tweets may spark discussions of differing opinions, while tweets with strong emotional biases tend to elicit emotional responses. Furthermore, this invention observes that the nature of a tweet's content largely determines the attitudes and emotional tone of the comments section. For instance, when a tweet evokes sympathy or describes a tragic event, even if it is proven false, the comments tend to be supportive and positive. This phenomenon may be because people tend to respond with sympathy and concern when faced with tragic situations, ignoring the veracity of the facts. Conversely, when a tweet involves real events that actually exist and negatively impact people's lives and work, the comments section often displays strong opposition and skepticism. This emotional response may stem from public concern about social issues and dissatisfaction with the underlying causes of policies or events.

[0046] Furthermore, there are potential connections between comments under the same tweet, often manifested in language content, emotional tone, and consistency of viewpoint. Commonalities in comments often manifest as similar stances or emotional reactions, especially under tweets related to social events, where commenters may collectively express anger or support. Consistency in language and emotion among comments also reveals potential relationships; for example, similar or repetitive comments may originate from collaborative generation by those spreading misinformation, potentially related to the spread of false information. Significant emotional divergence, on the other hand, may indicate the controversial or divisive nature of the information dissemination.

[0047] Therefore, exploring the potential relationships among various users in social networks is of great significance for analyzing the nature of tweets in this invention. In social networks, each user promotes discussion by sharing insights, constructing an information exchange network, gathering different experiences and perspectives, and demonstrating the power of collective wisdom. We will analyze the interaction patterns between tweets, between tweets and comments, and between comments to extract a more comprehensive public opinion landscape and provide a more scientific and rational basis for the analysis of complex events.

[0048] Text data from tweets and comments on social networks are input into the model. This forms the foundational data layer for the entire system, containing the basic units of information dissemination and collective feedback from user interactions.

[0049] Subsequently, Large Language Models (LLMs) were used to extract deep semantic features and sentiment distributions from these texts. This process not only understands the surface meaning of the text but also captures the underlying attitudes, transforming unstructured text into computationally achievable feature representations.

[0050] Next, leveraging the concept of collective intelligence, the model integrates this diverse semantic information from different users. This step is crucial; it aggregates individual users' perspectives to form a collective understanding and evaluation of the post's content, thereby overcoming the limitations of a single perspective.

[0051] Ultimately, the system employs an unsupervised learning method, using the comprehensive features generated in the preceding steps to evaluate the authenticity of posts. This method does not rely on pre-labeled data and can automatically identify information with characteristics of false information, thus completing the detection task. Figure 1 As shown, the specific processing procedure of the present invention is as follows:

[0052] S1: Obtain blog post and comment data from social media platforms, and construct a blog post-comment interaction network based on the blog post and comment data.

[0053] By acquiring tweet and user comment data from social media, and defining tweet nodes, comment nodes, and directed connections between them, a multi-level interaction graph structure is constructed to reflect the relationships between tweets and the posting relationships between tweets and comments; specifically:

[0054] Assuming the tweet is This is a message posted by a user. We define it as a node, represented as:

[0055]

[0056] in Indicates the first One tweet.

[0057] Each tweet There can be multiple comments below, and these comments form a set, denoted as . :

[0058]

[0059] in Indicates the first 1 comment.

[0060] The edges of a tweet-comment interaction network can be represented as:

[0061] The relationship between tweets:

[0062] Assuming that there may be some kind of dependency or propagation relationship between tweets, such as retweets, citations, or topic relevance, this invention names this relationship Relationships Between Posts (RBP). The set of tweets 𝑇 is represented as:

[0063]

[0064] Define the relationship between tweets as directed edges. , indicating tweet With tweets There is some kind of relationship between them (e.g.) Forwarded or and (The topics are related). The formal definition is as follows:

[0065]

[0066] in Tweet With tweets The propagation or semantic association between them.

[0067] The relationship between tweets and comments:

[0068] Each tweet (T) typically sparks discussion among users. This is to define the tweet (T) and its comments. The relationship between Posts and Comments is named Relationships Between Posts and Comments (RBPC), and represented by a directed graph where each tweet node points to a comment node. This forms the following relationship:

[0069]

[0070] in Tweet Rather than comment The relationship between them.

[0071] S2: Construct cue words based on the blog post-comment interaction network; based on the cue words, use a large language model to process the blog post and comment data to obtain the attitude classification results of the comments.

[0072] To obtain a more comprehensive picture of public opinion on social networks, the blog post-comment interaction network needs to be optimized. Specifically, in addition to the two types of edge relationships mentioned above, the blog post-comment interaction network should also include relationships between comments:

[0073] Each comment All of these can express a certain attitude, such as support, opposition, neutrality, or skepticism. Let a function A(cj) represent the attitude type of comment cj, which can take the following values:

[0074]

[0075] Based on the similarity of attitude types, Relationships Between Comments (RBC) is used to represent the relationships between comments. for:

[0076]

[0077] That is, when two comments ci and cj have the same attitude type, a commonality relationship is defined between them.

[0078] To understand the relationships between comments, cue words are constructed based on the blog post-comment interaction network. Prompts are used to guide the LLM (Large Language Model) in analyzing tweet content, and these prompts are embedded in both the tweet text and comments to optimize the input format. For example, using the Llama large language model for sentiment analysis, the input is provided as follows: "This is a message: xxxx…; Below this message is a comment: Oh my god…", with explicit prompts telling the model its task, such as "Please analyze…". Experimental results show that the LLM can deeply understand and analyze the sentiment and context of comments, thereby generating reasonable responses. Simultaneously, the LLM significantly reduces human bias in the sentiment analysis process, providing more consistent and objective sentiment and attitude classifications, resulting in more objective results. This method, through fine-tuning the large model, can help it further adapt to downstream tasks in experiments.

[0079] Based on the prompt words, after classifying the opinions of the comments under the tweet using a Large Language Model (LLM), three main opinion types were extracted: supportive, neutral, and critical.

[0080] Assume that tweets and comments on the same topic constitute a dataset. ,in This represents the i-th tweet. Tweet Below After obtaining the categorization results for the comments, the comment data was divided into three categories:

[0081]

[0082] In this way, the present invention obtains structured comment data and obtains the results of comment attitude classification.

[0083] S3: Quantify the attitude vector of blog posts based on the attitude segmentation results of the comments.

[0084] Based on the attitude classification results, the proportion of each attitude was obtained: Let the proportions of the three attitudes be respectively... , , .

[0085] To maximize the distinguishability between true and false information in the feature space, this invention sets the weight of each attitude comment as a support weight based on the semantic representation differences of different attitudes. Questioning weight Neutral weight This asymmetric weighting strategy assigns a higher penalty coefficient to "skepticism," not only achieving adversarial separation of positive and negative polarities in the vector direction but also highlighting the high fidelity of skepticism in identifying potential misinformation, thereby effectively preventing false information features from being overwhelmed by a large number of blindly supportive comments.

[0086] By combining the weight of each attitude comment with its corresponding proportion, an attitude vector for blog posts and tweets is obtained. Attitude vector Defined as:

[0087]

[0088] in, This represents the attitude vector of blog post i. , and These represent the weights of supportive, neutral, and critical comments, respectively. , and The percentages of comments that support, are neutral, and are critical are represented respectively.

[0089] so, It's a tweet The three-dimensional attitude vector reflects the attitude distribution of comments under the tweet.

[0090] By defining the user comment network as described above, we can obtain the correspondence between all comments under each tweet and the original tweet, thereby generating a social network opinion landscape.

[0091] S4: Extract semantic embeddings from blog posts; extract semantic center vectors from the comments corresponding to the blog posts based on the attitude segmentation results of the comments.

[0092] In order to capture tweets Its comment collection The semantic relationships between them are explained using the Transformer model in this invention. Convert the tweet text and the corresponding comments into high-dimensional embedding vectors.

[0093] Let the embedding vector of the tweet be... The set of embedding vectors for comments is The calculation formula is as follows:

[0094]

[0095] in This indicates that the Sentence Transformer model is used to transform text into a fixed-length vector representation.

[0096] To further aggregate the embedded representations of the three types of opinion comments, this invention, based on the social network public opinion landscape, aggregates the comment embedding vectors of the same attitude according to the attitude classification results, obtaining the comment embedding vector sum for each attitude comment:

[0097]

[0098] The comment embedding vector of attitude comments and the semantic center vector of the corresponding attitude category are used; these vectors The tweet indicated The semantic distribution center of the three types of comments.

[0099] S5: By integrating the attitude vector, semantic embedding, and semantic center vector of the blog post's corresponding comments, a comprehensive feature representation of the blog post is obtained.

[0100] The aforementioned process obtained each tweet Three-dimensional attitude vectors of the three types of comments This can scientifically and intuitively reflect the distribution of supportive, skeptical, and neutral attitudes under a tweet. Through this multi-layered data processing and analysis, we can not only capture the semantic interaction between tweets and comments but also uncover deeper patterns in opinion distribution. By fusing this vector with the blog post's attitude vector and semantic embedding, a comprehensive feature representation with fused semantics can be generated for each tweet. Specifically: the semantic embedding vector of the tweet... Three types of comment semantic center vectors , , And the attitude vector of the tweet The pieces are then combined to form the post. The comprehensive characteristics are represented.

[0101] S6: Cluster all blog posts based on the comprehensive feature representation of each post to obtain the false information detection results.

[0102] In existing technologies, misinformation detection methods mainly rely on supervised learning, which requires a large amount of labeled data, or analyze only the textual features of the post itself, often neglecting the collective cognitive value contained in user comments. Traditional methods, such as detection models based on graph neural networks, consider the propagation structure but are insufficient in mining the semantics and sentiment of comments; while methods based on pre-trained language models can understand textual semantics, they struggle to effectively integrate the characteristics of group attitude distribution. In contrast, this invention achieves an innovative breakthrough by constructing a multi-dimensional feature representation, integrating post semantics, comment semantic centers, and attitude vectors, and employing an unsupervised clustering method to accurately identify misinformation without labeled data.

[0103] The process of clustering all blog posts based on their comprehensive feature representations includes:

[0104] Step 1: Select blog posts with typical characteristics of misinformation from the blog post collection as beacon posts. .

[0105] Empirical observations show that, compared to factual information, misinformation tweets often employ inflammatory, exaggerated, and emotional language to attract attention, evoke emotional responses, and incite online controversy. They differ significantly from ordinary information in both semantic and emotional characteristics. These tweets, aimed at gaining attention, are defined as "tweets from online misinformation publishers," and they play a crucial role in distorting public opinion and misleading public perception.

[0106] The beacon tweet screening process aims to identify the most representative "online misinformation publisher tweets" from candidate tweets. This screening is achieved by evaluating the salient features of tweets within a specific topic context, primarily including: the use of inflammatory or emotionally manipulative language, a clear tendency towards emotional polarization, and frequently repeated templated expressions. By extracting tweets with these typical characteristics, this invention can accurately capture the core traits of misinformation publishers' behavior. These selected tweets will serve as strong reference points (i.e., beacons), providing guidance for downstream clustering and classification tasks, thereby enhancing the robustness and interpretability of the detection mechanism.

[0107] Step 2: Calculate the similarity between the composite feature representation of all blog posts and the composite feature representation of the selected beacon blog posts.

[0108] In some preferred embodiments of the present invention, cosine distance is used as a metric to calculate the similarity between the comprehensive feature representation of all posts in the dataset and the comprehensive feature representation of the selected beacon post. This distance simultaneously considers the semantic content of the post, the distribution of group attitudes in the comments, and sentiment tendencies.

[0109] Step 3: Based on similarity, use an unsupervised clustering algorithm to process all blog posts and obtain clustering results; the cluster containing the beacon blog posts in the clustering results is taken as the false information cluster, i.e., the false information detection result.

[0110] Based on the clustering results, the cluster containing the beacon blog post is taken as the false information cluster, i.e., the false information detection result. The blog posts in the false information cluster are closest to the beacon blog post in the feature space, i.e., the semantic and sentiment features are most similar.

[0111] Simulation verification of the present invention:

[0112] Data on real and false information from Weibo and Twitter was obtained through existing platforms (GitHub), specifically the CHECKED and PHEME datasets. Simulation tests were conducted on the present invention and existing models, resulting in a performance comparison table, as follows: Figure 2 As shown, the Unsup metric in the model indicates that it follows an unsupervised learning method, while the Sup metric indicates that it follows a supervised learning method. Furthermore, bolded results represent the best performance in supervised and unsupervised scenarios, while underlined results represent the best performance in unsupervised scenarios. Experimental results show that this invention not only significantly outperforms traditional unsupervised methods (such as PCA), but also rivals or surpasses some advanced unsupervised models (such as Graph-GAN, NIT) and even some supervised models (such as AGAT) in key metrics such as accuracy, recall, and F1 score, demonstrating its effectiveness in capturing deep features when label data is lacking. Secondly, the model performs particularly well on the Chinese dataset (CHECKED), even outperforming some supervised algorithms (such as SVM-TS, DTC), highlighting its strong generalization ability and robustness in Chinese misinformation detection tasks by adapting to the characteristics of the Chinese language and social networks.

[0113] In summary, this invention proposes an unsupervised method for detecting misinformation specifically designed for social networks. This method comprehensively utilizes the semantic relationships between tweets and related comments, user interaction patterns, and attitude distribution features revealed by sentiment analysis. Through multi-dimensional feature fusion, the model can effectively identify tweets with typical characteristics of misinformation dissemination, improving the accuracy of misinformation detection results.

[0114] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An unsupervised method for detecting false information based on collective intelligence and content cross-heuristics, characterized in that, Includes the following steps: S1: Obtain blog post and comment data from social media platforms, and construct a blog post-comment interaction network based on the blog post and comment data; S2: Construct cue words based on the blog post-comment interaction network; based on the cue words, use a large language model to process the blog post and comment data to obtain the attitude classification results of the comments; S3: Quantify the attitude vector of blog posts based on the attitude segmentation results of comment content; S4: Extract semantic embeddings from blog posts; extract semantic center vectors from the comments corresponding to the blog posts based on the attitude segmentation results of the comments; S5: By integrating the attitude vector, semantic embedding, and semantic center vector of the blog post's corresponding comments, a comprehensive feature representation of the blog post is obtained; S6: Cluster all blog posts based on the comprehensive feature representation of each post to obtain the false information detection results.

2. The unsupervised method for detecting false information based on collective intelligence and content cross-heuristics as described in claim 1, characterized in that, The process of building a blog post-comment interaction network includes: Treating both blog posts and comments as nodes, if there is a propagation relationship between blog post nodes, then there is an edge between the two nodes; if a comment belongs to a blog post, then there is an edge between the corresponding comment node and the blog post node.

3. The unsupervised method for detecting false information based on collective intelligence and content cross-heuristics as described in claim 1, characterized in that, Step S3 specifically includes: The proportion of each attitude comment was obtained based on the attitude classification results; Set the weight for each attitude comment; By combining the weight of each attitude comment with its corresponding proportion, an attitude vector for the blog post is obtained; among them, the attitude of the comments is divided into support, neutrality, and skepticism.

4. The unsupervised method for detecting false information based on collective intelligence and content cross-heuristics as described in claim 3, characterized in that, The weight of each attitude and its corresponding proportion are expressed as follows: ； in, This represents the attitude vector of blog post i. , and These represent the weights of supportive, neutral, and critical comments, respectively. , and The percentages of comments that support, are neutral, and are critical are represented respectively.

5. The unsupervised method for detecting false information based on collective intelligence and content cross-heuristics as described in claim 1, characterized in that, The process of extracting semantic center vectors from the comments corresponding to blog posts based on the attitude segmentation results includes: The comments corresponding to the blog post are embedded to obtain the comment embedding vector; Based on the attitude segmentation results of the comments, the comment embedding vectors of the same attitude are aggregated to obtain the comment embedding vector sum of each attitude comment; The comment embedding vector of the attitude comment and the semantic center vector of the corresponding attitude category are used.

6. The unsupervised method for detecting false information based on collective intelligence and content cross-heuristics as described in claim 1, characterized in that, The method of integrating the attitude vector, semantic embedding, and semantic center vector of the blog post and its corresponding comments is concatenation.

7. The unsupervised method for detecting false information based on collective intelligence and content cross-heuristics as described in claim 1, characterized in that, The process of clustering all blog posts based on their comprehensive feature representations includes: Blog posts with typical characteristics of misinformation were selected from the blog post collection as beacon posts; Calculate the similarity between the composite feature representation of all blog posts and the composite feature representation of the selected beacon blog posts; Based on similarity, an unsupervised clustering algorithm is used to process all blog posts to obtain clustering results; the cluster containing the beacon blog posts in the clustering results is taken as the false information cluster, i.e., the false information detection result.

Images