Method and device for detecting fake news based on user sentiment and review semantic mining

Abstract

The present application relates to the technical field of fake news detection, and particularly to a fake news detection method, device and apparatus based on user emotion and comment semantic mining, and a computer storage medium. The fake news detection method comprises the following steps: firstly, initial feature representation is performed on news text, comments and user emotion; secondly, according to news initial embedding at the sentence level, news representation associated with context is learned, and news fine-grained semantics are captured for subsequent comment enhancement learning; thirdly, different attention of news text sentences is considered to obtain enhanced comment semantic embedding of news; fourthly, in the unit of comments, a gating mechanism is designed to perform feature fusion on comment semantics and user emotion, and to mine associated features of the two; fifthly, a long short-term memory network is used to capture the change mode of comment semantics and user emotion in the comment area; and finally, the features extracted by the modules are comprehensively considered to determine the news category. The present application improves the accuracy of fake news detection.

Description

Technical Field

[0001] This invention relates to the field of fake news detection technology, and in particular to a method, device, apparatus, and computer storage medium for detecting fake news based on user sentiment and comment semantic mining. Background Technology

[0002] The rapid development of social media has transformed the way information is disseminated. Through social media, information is no longer limited by time and space; users can easily publish content and interact with others. However, this very convenience and immediacy, while increasing the efficiency of information dissemination, has also provided channels for the rapid spread of fake news.

[0003] The widespread dissemination of fake news has had numerous negative impacts on society. First, it seriously endangers public safety and social order. Second, the proliferation of fake news exacerbates the public's crisis of trust in the news. According to a Reuters survey, only 42% of users trust the news they see online, a lack of trust stemming in part from the widespread dissemination of fake news in multiple fields, including politics, economics, and health. Third, the widespread dissemination of fake news weakens the credibility of governments. Previous research has also found that fake news can effectively confuse voters and influence election results. Therefore, curbing the spread of fake news and reducing its harm to society has become an urgent task.

[0004] Currently, fake news detection faces numerous challenges. First, fake news is often deliberately edited and published to mislead the public, but with the development of large language models and generative artificial intelligence, the boundary between fake and real news is becoming increasingly blurred. Therefore, it is difficult to accurately determine the authenticity of news based solely on its content. Second, social media is filled with a vast amount of information, diverse content formats, and is mostly user-generated, resulting in significant data noise. Therefore, relying on manual fake news detection is not only extremely time-consuming but also inefficient.

[0005] Many researchers have conducted extensive research in the field of fake news detection. Existing fake news detection methods can be broadly categorized into content-based methods and social context-based methods.

[0006] Fake news detection based on news content primarily focuses on the information within the news itself. Representative methods can be categorized into three types: 1) Language feature-based methods, which identify fake news by capturing differences in writing style and semantics between real and fake news. However, since fake news is often deliberately edited before publication, the boundary between fake and real news is becoming increasingly blurred with the development of LLM and AIGC. To address this challenge, Wu et al. utilized LLM to generate diverse news reconstructions and combined them with content clues, effectively improving the robustness and accuracy of fake news detectors against LLM-driven style attacks. 2) Knowledge-based detection methods, which enrich text representation and improve detection accuracy by introducing external knowledge, such as topic descriptions, source credibility, and knowledge graphs. 3) Multimodal feature fusion methods, which combine text, image, and video data to improve detection performance. Compared to unimodal methods, multimodal methods, by integrating information from different sources, can better identify inconsistencies. However, multimodal approaches also face some challenges: collecting and labeling data often requires significant costs and time; during inference, missing or incomplete data for certain modalities may affect the model's performance.

[0007] Early research on fake news detection primarily focused on manually extracting features from prior knowledge of fake news and training machine learning classifiers based on these features. However, this approach is labor-intensive, costly, and struggles to capture the deep features of fake news. Furthermore, relying solely on news content often fails to accurately determine its authenticity. To address these issues, many researchers have introduced social contextual information (such as user behavior and dissemination structures) and combined it with deep learning techniques to infer the authenticity of news, achieving significant results. However, social context-based methods heavily rely on external information, the collection and processing of which are complex and computationally expensive. In contrast, content-based methods rely solely on the news text itself, requiring no additional information. Since news text is typically readily available, this approach is more convenient and efficient in practice. As the most direct feedback from users to news, comment data can provide crucial clues for fake news detection. Some studies have attempted to utilize comment data to assist in fake news detection. For example, dEFEND introduces a sentence-comment co-attention mechanism to analyze the correlation between news content and user comments, selecting comments that can explain fake news to improve detection performance. Hamed et al. provided auxiliary support for fake news detection by extracting the emotional features of news headlines and combining them with emoticon information from user comments.

[0008] While existing methods for detecting fake news using user comments have achieved some success, most focus on feature mining of individual comments, with few studies exploring sentiment and semantic changes from the perspective of the comment section. Exploring sentiment changes in the comment section is crucial for fake news detection. First, these changes directly reflect shifts in public attitudes towards news events. In user-participated discussions, emotional tendencies directly reflect users' positive (happy, optimistic), negative (bored, pessimistic), or neutral (no obvious emotional tendency) attitudes, and this emotional information can help infer the veracity of the news. For example, research shows that fake news is more likely to evoke negative public emotions compared to real news. Second, the evolution trend of sentiment in the comment section can also reveal the veracity of the news. For example, consider a piece of information from the Weibo-comp dataset... Figure 1 Over time, the emotions in the comments section gradually shifted from initial surprise to anger, and eventually to neutrality. This emotional shift helps reveal the veracity of the news content. When the comments section displays increasingly negative emotions, it may indicate that the news has triggered widespread distrust or aversion, thus providing clues for detecting fake news. Similarly, the evolution of semantics in the comments section is also an important basis for revealing fake news. Figure 1 Over time, the content of discussions in the comments section evolved from initial "uncertainty and concern" to "questioning and criticism," and finally to "clarifying fake news." This semantic evolution reflects the deepening understanding of news events by the public. Initially, the public may have been confused about the veracity of the event, but as the discussion progressed, clearer judgments and clarifications gradually formed. This semantic change also provides a basis for identifying the falsity of news. Existing research, such as the Post2DR model proposed by Yang et al., while using CNNs to model local changes in comment semantics, lacks the ability to capture global changes in comment semantics. Figure 1 Furthermore, it can be observed that news texts typically contain a wealth of information, and the discussions they spark are often diverse. For example, some users are concerned about food safety issues, while others are concerned about consumer rights. News articles provide detailed background and context for events, which helps in accurately interpreting the semantics and sentiment of comments. However, current research on comment representation learning often treats comments as independent entities, focusing solely on the comment text itself. While this approach is simple, it fails to reveal the intrinsic connection between comments and news content.

[0009] In conclusion, the existing methods for detecting fake news using user comments have low accuracy, which is a problem that needs to be solved. Summary of the Invention

[0010] Therefore, the technical problem to be solved by the present invention is to overcome the low accuracy of existing methods for detecting fake news using user comments.

[0011] To address the aforementioned technical problems, this invention provides a method for detecting fake news, comprising:

[0012] Obtain the target news and its comment set, and extract the news text embedding, comment set text embedding, and comment set sentiment embedding;

[0013] The news text embedding is subjected to context-associative encoding to obtain the news encoded representation;

[0014] Based on the news text embedding, news reinforcement learning is performed on the comment set text embedding to obtain a personalized representation of the comment set text;

[0015] The personalized representation of the comment set text and the sentiment embedding of the comment set are fused to obtain the semantic-sentiment fusion vector representation of the comments;

[0016] Based on the personalized representation of the comment set text and the sentiment embedding of the comment set, semantic evolution features and user sentiment evolution features are obtained, and fake news classification prediction is performed by combining the news text embedding and the semantic-sentiment fusion vector representation of the comments.

[0017] Preferably, the extraction of news text embeddings, comment set text embeddings, and comment set sentiment embeddings includes:

[0018] The text embedding of each sentence in the target news is obtained by using the first pre-trained model;

[0019] The text embedding of each comment in the target news comment set is obtained by using a second pre-trained model;

[0020] Each comment in the target news comment set is classified by sentiment to obtain the sentiment embedding of the comment set.

[0021] Preferably, the step of performing context-related encoding on the embedded news text to obtain the news encoded representation includes:

[0022] The news text embedding is encoded using a bidirectional long short-term memory network to obtain the preceding state sequence information and the following state sequence information;

[0023] The hidden state of the last time step in the preceding state sequence information and the hidden state of the last time step in the following state sequence information are concatenated to obtain the news encoding representation.

[0024] Preferably, the step of performing news reinforcement learning on the comment set text embedding based on the news text embedding to obtain a personalized representation of the comment set text includes:

[0025] Based on the attention mechanism, news reinforcement learning is performed on each comment text embedding according to the news text embedding to obtain a personalized representation of the comment set text.

[0026] Preferably, the step of fusing the personalized representation of the comment set text and the sentiment embedding of the comment set to obtain the comment semantic-sentiment fusion vector representation includes:

[0027] The personalized representation and sentiment embedding of each comment text are concatenated to generate a gated vector;

[0028] The personalized representation and sentiment embedding of each comment text are linearly transformed and mapped to the same space, generating comment semantic vectors and sentiment vectors with the same dimensions;

[0029] Generate fused comment features based on the gating vector, the comment semantic vector, and the sentiment vector;

[0030] The fused comment feature sequence of the comment set is input into a long short-term memory network to generate state sequence information, and the hidden layer state of the last time step in the state sequence information is selected as the comment semantic sentiment fusion vector.

[0031] Preferably, obtaining semantic evolution features and user sentiment evolution features based on the personalized representation of the comment set text and the sentiment embedding of the comment set includes:

[0032] The personalized representation of the comment set text and the sentiment embedding of the comment set are respectively input into a long short-term memory network to generate corresponding state sequence information, and the hidden state layer of the last time step is selected to generate semantic evolution features and user sentiment evolution features.

[0033] Preferably, the process of classifying and predicting fake news includes:

[0034] The semantic evolution features, user sentiment evolution features, news text embeddings, and comment semantic sentiment fusion vectors are concatenated, and the correlation between features is calculated through a fully connected network and softmax function transformation, and finally the prediction vector is output.

[0035] The present invention also provides a fake news detection device, comprising:

[0036] The initial feature representation module is used to obtain the target news and its comment set, and extract the news text embedding, comment set text embedding, and comment set sentiment embedding;

[0037] The news encoding module is used to perform context-related encoding representation on the embedded news text to obtain a news encoded representation;

[0038] The comment representation learning module is used to perform news reinforcement learning on the comment set text embedding based on the news text embedding to obtain a personalized representation of the comment set text;

[0039] The semantic sentiment fusion module is used to fuse the personalized representation of the comment set text with the sentiment embedding of the comment set to obtain the semantic sentiment fusion vector representation of the comments;

[0040] The news classification module is used to obtain semantic evolution features and user sentiment evolution features based on the personalized representation of the comment set text and the sentiment embedding of the comment set, and to perform fake news classification prediction by combining the news text embedding and the semantic-sentiment fusion vector representation of the comments.

[0041] The present invention also provides a fake news detection device, comprising:

[0042] Memory, used to store computer programs;

[0043] A processor is used to implement the steps of the above-described method for detecting fake news when executing the computer program.

[0044] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for detecting fake news.

[0045] The technical solution of the present invention has the following advantages compared with the prior art:

[0046] The fake news detection method described in this invention first performs initial feature representation on the news text, comments, and user sentiment. Second, based on the initial news embedding at the sentence level, it learns context-related news representations, capturing fine-grained semantics of the news while simultaneously providing reinforcement learning for subsequent comments. Third, it comprehensively considers different attention levels within the news text sentences to obtain news-enhanced comment semantic embeddings. Fourth, it designs a gating mechanism to fuse comment semantics and user sentiment features on a comment-by-comment basis, mining the correlation between the two. Fifth, it captures the changing patterns of comment semantics and user sentiment within the comment section using a Long Short-Term Memory (LSTM) network. Finally, it integrates the features extracted from each module to determine the news category. This invention improves the accuracy of fake news detection. Attached Figure Description

[0047] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings, wherein:

[0048] Figure 1 This is an example illustration from the Weibo-comp dataset;

[0049] Figure 2 A flowchart illustrating the implementation of a fake news detection method provided by this invention;

[0050] Figure 3 A flowchart illustrating the implementation of a fake news detection method provided in an embodiment of the present invention;

[0051] Figure 4 This is a performance diagram illustrating the different settings for the maximum number of comments on two datasets.

[0052] Figure 5 This is a performance diagram illustrating the different settings for the maximum number of news sentences on two datasets.

[0053] Figure 6 It is d ce The performance diagrams for different settings on the two datasets are shown.

[0054] Figure 7 This is a distribution map of sentiment evolution features in the comment section of sentiment consistency cases from the Weibo-comp dataset;

[0055] Figure 8 This is a sentiment change graph in the comment section of a sentiment consistency case study from the Weibo-comp dataset;

[0056] Figure 9 This is a distribution map of sentiment evolution features in the comment section of sentiment reversal cases from the Weibo-comp dataset;

[0057] Figure 10 This is a sentiment change graph in the comment section of a sentiment reversal case from the Weibo-comp dataset;

[0058] Figure 11 This is a distribution map of sentiment evolution features in the comment section of sentiment reversal cases from the RumourEval-19 dataset;

[0059] Figure 12 This is a sentiment change graph in the comment section of a sentiment reversal case from the RumourEval-19 dataset;

[0060] Figure 13 This is a graph showing the cumulative sentiment scores in the comment section of sentiment reversal cases from the RumourEval-19 dataset. Detailed Implementation

[0061] The core of this invention is to provide a method, apparatus, device, and computer storage medium for detecting fake news, which effectively improves the accuracy of fake news detection.

[0062] To enable those skilled in the art to better understand the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely 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.

[0063] Please refer to Figure 2 , Figure 2 The flowchart illustrates the implementation of a fake news detection method provided by this invention; the specific operation steps are as follows:

[0064] S101: Obtain the target news and its comment set, and extract the news text embedding, comment set text embedding, and comment set sentiment embedding;

[0065] S102: Perform context-related encoding on the embedded news text to obtain a news encoding representation;

[0066] S103: Based on the news text embedding, perform news reinforcement learning on the comment set text embedding to obtain a personalized representation of the comment set text;

[0067] S104: The personalized representation of the comment set text and the sentiment embedding of the comment set are fused to obtain the semantic sentiment fusion vector representation of the comments;

[0068] S105: Based on the personalized representation of the comment set text and the sentiment embedding of the comment set, obtain semantic evolution features and user sentiment evolution features, and combine the news text embedding and the comment semantic sentiment fusion vector representation to perform fake news classification prediction.

[0069] Please refer to Figure 3 , Figure 3 This is a flowchart illustrating the implementation of a fake news detection method provided in an embodiment of the present invention.

[0070] Based on the above embodiments, this embodiment will provide a detailed description of step S101:

[0071] In some embodiments, extracting news text embeddings, comment set text embeddings, and comment set sentiment embeddings includes:

[0072] The text embedding of each sentence in the target news is obtained by using the first pre-trained model;

[0073] The text embedding of each comment in the target news comment set is obtained by using a second pre-trained model;

[0074] Each comment in the target news comment set is classified by sentiment to obtain the sentiment embedding of the comment set.

[0075] Fake news content is typically related to current events or trending topics and is deliberately crafted before publication to attract attention and generate widespread discussion. Because the text usually revolves around a specific event or topic, the context is often highly relevant. Therefore, capturing the contextual information of news content to improve the quality of text embedding is crucial for fake news detection.

[0076] In some embodiments, the first pre-trained model may be, for example, a model that can capture contextual information of news content to improve the quality of text embedding; this embodiment does not limit the specific model.

[0077] The BERT model has achieved excellent performance in multiple fields such as natural language processing and image classification. In text representation learning, compared to previous models, BERT, based on a bidirectional transformer, fully utilizes textual context information, greatly alleviating the problem of ambiguous word embeddings. Simultaneously, it uses a self-attention mechanism to weight words to obtain the vectorized representation of sentences, effectively addressing long-distance contextual dependencies.

[0078] In one embodiment, the first pre-trained model uses a pre-trained BERT model for news embedding. For a news article consisting of M sentences, P = {s1, s2, s3, s4, s5...s} M},in, Sentence-level embeddings are obtained using the pre-trained model Bert. T P ∈R M×d .in Let d be the embedding of the i-th sentence, and d be the dimension of the sentence embedding. The calculation method is as follows:

[0079]

[0080] Here, PreModel represents the pre-trained model BERT. For English text, the bert-base-uncased model trained on an unlabeled English corpus is selected; for Chinese text, the Chinese-Roberta-wwmext model is selected to embed the news text.

[0081] In order to map comments from text to computer-processable vectors containing semantic features, in some embodiments, a second pre-trained model is sampled to encode the comments; this embodiment does not limit the pre-trained model.

[0082] In one embodiment, similar to news embedding, the second pre-trained model uses a BERT pre-trained model to encode the comments. For the comment set C = {c1, c2, c3, ..., c...} corresponding to the news article... |C|}, where c i This represents the i-th comment, with the comment text embedded as follows: Among them, T C ∈R |C|×d d represents the text embedding dimension. The text embedding for the i-th comment. The specific calculation method is as follows:

[0083]

[0084] PreModel represents the pre-trained model. Similar to news embedding, the model selected varies depending on the text language type.

[0085] Comments are a primary means for people to participate in topic discussions and typically contain rich emotional signals. To extract emotional information from user comments and convert it into numerical vectors, this embodiment does not limit the specific emotional classification method.

[0086] In one embodiment, for English text, the widely used Ekman basic sentiment classification is selected, dividing sentiment into six categories: {Anger, Disgust, Fear, Joy, Sadness, Surprise}. For Chinese data, the sentiment analysis service provided by Baidu's intelligent platform is used to divide text sentiment into nine categories: {Annoyance, Sadness, Like, Neutral, Optimism, Fear, Happiness, Anger, Pessimism}. The user sentiment feature set is represented as E c ={e1,e2,e3,...e |C|}, regarding comment c i Its emotional vector is e i e i The calculation method is as follows:

[0087] e i =emotion_model(c i )#(3)

[0088] Among them, e i ∈R 1×k The numerical values ​​in the vector represent the probability that the text belongs to the corresponding sentiment category, and k is the number of sentiment categories.

[0089] Based on the above embodiments, this embodiment will provide a detailed description of step S102:

[0090] The initial representation of news text can capture the semantics within sentences relatively well. In order to better learn the contextual relationships at the sentence level of news, the news text embedding is represented by contextual relationship encoding. This embodiment does not limit the specific encoding method.

[0091] In one embodiment, performing context-related encoding on the news text embedding to obtain a news encoding representation includes:

[0092] The news text embedding is encoded using a bidirectional long short-term memory network to obtain the preceding state sequence information and the following state sequence information:

[0093] The news is further encoded using a bidirectional LSTM. The news sentence embedding sequence is used. As input to a bidirectional LSTM, it is encoded to obtain preceding and following information. The forward state vector at time step t is... The backward state vector is The formal expression is as follows:

[0094]

[0095] Forward network computing as an example, The calculation process is obtained through the forget gate, input gate, memory update unit, and output gate structure of the feedforward long short-term memory network, as shown below:

[0096]

[0097] Among them, s t This is the input for the t-th time step. These correspond to the input gate, forget gate, memory unit, and output gate structures of the forward LSTM, respectively. The backward network computation process is similar to the forward process.

[0098] The hidden layer state of the last time step in the preceding state sequence information. and the hidden layer state of the last time step in the subsequent state sequence information. To splice,

[0099] The news representation T obtained after associating the context sentences s .

[0100]

[0101] in, d s This refers to the text embedding dimension during the news encoding stage.

[0102] Based on the above embodiments, this embodiment will provide a detailed description of step S103:

[0103] In some embodiments, news content is used as contextual information for comments, and news reinforcement learning is performed on the text embedding of the comment set; the specific method is not limited.

[0104] Inspired by the work of Wu et al., this study considers readers' reading habits when fusing features from image and text semantics. From the perspective of public reading habits, when reading news text, users typically focus their attention on certain sentences and paragraphs, and their comments are often only related to specific parts of the news. Different users also focus on different information. Therefore, this invention uses news content as contextual information for comments and applies an attention mechanism to learn news-enhanced personalized comment representations.

[0105] In one embodiment, performing news reinforcement learning on the comment set text embedding based on the news text embedding to obtain a personalized representation of the comment set text includes:

[0106] Based on the attention mechanism, news reinforcement learning is performed on each comment text embedding according to the news text embedding to obtain a personalized representation of the comment set text:

[0107] With the initial embedding of news and initial embedded comments As input, an attention mechanism is used to obtain the sequence of news-enhanced comment representations. in This represents the i-th comment after news enhancement. Specifically, it is derived from the initial embedding T of the input. P and T C A linear transformation yields the query matrix Q, the key matrix K, and the value matrix V. The calculation process for matrices Q, K, and V is as follows:

[0108]

[0109] Among them, W q W k W v The parameters are for a linear transformation. The attention mechanism measures the attention of each sentence s in a news article, given a news segment. j Comment c i The effect of attention score is calculated using the following formula.

[0110]

[0111] After the attention score is calculated, the softmax function is applied to normalize the attention score, converting the score into a probability distribution to obtain the weighting coefficient α. i According to the weighting coefficient α iThe weighted calculation of V yields the enhanced comment representation of the i-th news item. α i and The calculation formula is as follows:

[0112]

[0113] Where d c for The embedding dimension.

[0114] Based on the above embodiments, this embodiment will provide a detailed description of step S104:

[0115] In some embodiments, the personalized representation of the comment set text and the emotional embedding of the comment set are fused together. This embodiment does not limit the specific fusion method.

[0116] In one embodiment, inspired by the GRU update gate control mechanism, this invention innovatively introduces a gating mechanism to fuse the semantics and sentiment of comments, thereby gaining a more comprehensive understanding of comment data:

[0117] The personalized representation of the comment set text and the sentiment embedding of the comment set are fused to obtain the comment semantic sentiment fusion vector representation, which includes:

[0118] The personalized representation and sentiment embedding for each comment text are concatenated to generate a gated vector:

[0119] In one specific embodiment, comments are embedded. With emotion vector e t After concatenation, the data is processed through a fully connected layer, and a gated vector u is generated using the sigmoid function. t u t Used to control the degree to which semantic and emotional information is preserved.

[0120]

[0121] The personalized representation and sentiment embedding of each comment text are linearly transformed and mapped to the same space, generating comment semantic vectors and sentiment vectors with the same dimensions:

[0122] In one specific embodiment, the semantic vector and emotional vector e t A linear transformation is performed to map the vectors to the same space, and then a hyperbolic tangent function transformation is used to obtain emotion vectors with the same dimension. and comment semantic vector h t :

[0123]

[0124]

[0125] Generate fused comment features based on the gating vector, the comment semantic vector, and the sentiment vector:

[0126] In one specific embodiment, the gate vector u is combined with t By comprehensively considering the semantic vector and sentiment vector of the comment, a fused feature containing both semantic and sentiment information of the comment is finally obtained.

[0127]

[0128] The fused comment feature sequence of the comment set is input into a long short-term memory network to generate state sequence information, and the hidden layer state of the last time step in the state sequence information is selected as the comment semantic sentiment fusion vector.

[0129] In one specific embodiment, a fusion feature sequence is learned through a Long Short-Term Memory (LSTM) network. Internal correlation:

[0130]

[0131] Take the representation of the last time step As a semantic and emotional fusion vector of the overall commentary, using It indicates. Among them, d ce yes Dimensions.

[0132] Based on the above embodiments, this embodiment will provide a detailed description of step S105:

[0133] If we analyze the emotional and semantic features of each comment on a news story in isolation, we can only obtain some static and isolated results. We cannot understand the overall semantic and emotional trends in the comment section, nor can we analyze the deep emotional motivations and group resonance effects behind the comments.

[0134] This example takes a perspective from the comments section, aiming to capture patterns of semantic and emotional changes among users participating in news discussions.

[0135] In one embodiment, obtaining semantic evolution features and user sentiment evolution features based on the personalized representation of the comment set text and the sentiment embedding of the comment set includes:

[0136] The personalized representation of the comment set text and the sentiment embedding of the comment set are respectively input into a long short-term memory network to generate corresponding state sequence information, and the hidden state layer of the last time step is selected to generate semantic evolution features and user sentiment evolution features.

[0137] Specifically:

[0138] Commentary enhanced by news Emotional expression E c ={e1,e2,e3,...,e |C| As the module input, the association features of the input sequence are learned through two LSTM networks, and the hidden state at the last time step is used. The hidden state at each time step of the input sequence is contained, so the hidden state at that time step is taken as the semantic change feature h. c Characteristics of emotional changes The calculation process is similar to that of semantic change features. (Comment on semantic evolution features h) c and user emotional evolution characteristics h e The formal expression is shown below:

[0139] h c =LSTM(T) pec )#(twenty two)

[0140] h e =LSTM(E c )#(twenty three)

[0141] Based on the above embodiments, we obtained the news embedding T s Comment semantic evolution features h c User emotional evolution characteristics h e and semantic sentiment fusion features

[0142] In one embodiment, performing fake news classification prediction includes:

[0143] The semantic evolution features, user sentiment evolution features, news text embeddings, and comment semantic sentiment fusion vectors are concatenated, and the correlation between features is calculated through a fully connected network and softmax function transformation, and finally the prediction vector is output.

[0144] Specifically:

[0145] These four feature vectors are concatenated, and then the correlation between the features is calculated by passing a fully connected network and a softmax function transformation, finally outputting the prediction vector. The values ​​of the vector represent the probability that a news item belongs to a different category. The calculation process is as follows:

[0146]

[0147] Where concate is the concatenation operation, W p Let b be the parameter matrix of the linear transformation. p This is the bias term for this layer.

[0148] The model is trained using the cross-entropy loss function, as shown below:

[0149]

[0150] Where Θ represents the parameters that the model needs to learn, and y represents the true label vector.

[0151] Based on the above embodiments, the dataset used by the present invention to verify the effectiveness of the model includes an English dataset and a Chinese dataset.

[0152] The English dataset is RumorEval-19, which originates from the rumor detection task of the SemEval-2019 competition. The news articles in the dataset are sourced from Twitter and Reddit platforms and were published by Gorrell et al. The news articles in this dataset are categorized into three types: real, fake, and unverified. In the experiments of this invention, the dataset partitioning method is consistent with the official competition problem, with the ratio of training set, test set, and validation set being 7:2:1.

[0153] The Chinese dataset is Weibo-Comp, provided by the "Detection of Online Rumors During Epidemics" competition organized by the CCF Big Data Working Group. It covers news content in eight fields on the Weibo platform, including technology, politics, business and finance, social life, culture and entertainment, sports and exams, and medicine and health, as well as some content related to the epidemic.

[0154] Table 1 shows the news and comment information related to the RumourEval-19 and Weibo-comp datasets:

[0155] Table 1. Basic information about the RumorEval-19 and Weibo-comp datasets.

[0156] Total number of news items 445 6444 Fake News 185 3581 Real News 138 2863 Unverified news 122 - Number of comments 7990 102372 Average number of comments 18 16

[0157] This invention selects several representative fake news detection methods for comparison. The comparison model is as follows:

[0158] LSTM: LSTM is a deep learning model that takes sequence data as input. In fake news detection tasks, word sequences and semantic embeddings are usually taken as input, and the results calculated by the network are subjected to average pooling to obtain the final text representation.

[0159] Text_CNN: Text_CNN is the first model to apply convolutional networks to text classification tasks, learning news semantic representations of different granularities by setting the size of the convolutional kernels.

[0160] BERT: This method fine-tunes the fully connected layer network based on the pre-trained model, thereby adapting it to the fake news detection task.

[0161] dEFEND: dEFEND is an interpretable fake news detection framework that uses a sentence-comment co-focus mechanism to improve fake news detection performance and fully utilizes user comments to interpret fake news.

[0162] HAN: HAN uses a hierarchical attention network to model news content at both the word and sentence levels.

[0163] HSA-BLSTM: HSA-BLSTM models news and comments at the word level, post level, and sub-event level using a hierarchical attention network. This method has been widely compared in the task of fake news detection.

[0164] Dual_Emo: This model detects fake news from an emotional perspective. It mines the relationship between publisher sentiment and user sentiment based on the text content of news and comments to improve the accuracy of fake news detection. This method uses HSA-BLSTM as the baseline model and introduces bidirectional sentiment features. After adding bidirectional sentiment features, the model's performance is further improved compared to HSA-BLSTM.

[0165] MTEFN: This model uses sentiment classification as an auxiliary task, combining sentiment classification and news classification to build a fake news detection framework. The aim is to improve the model's ability to extract key features from the data, thereby improving the accuracy of fake news detection.

[0166] AGWu: This model is a fake news detection model that simultaneously considers sentiment features and textual features. AGWu proposes a dual sentiment feature extraction mechanism based on deep normalized attention and designs a random forest classifier with adaptive genetic weight updates.

[0167] To ensure a fair comparison, the original parameters of the above models will not be changed.

[0168] For the RumorEval-19 dataset, the dataset partitioning method used in this invention is consistent with the competition organizers, with a training, test, and validation set ratio of 7:2:1. For the Weibo-comp dataset, the training, validation, and test sets are partitioned in a 6:2:2 ratio, meaning 60% of the data is used for training, 20% for validation, and 20% for testing. Other model parameter settings are shown in Table 2.

[0169] Table 2 Hyperparameter Settings

[0170]

[0171]

[0172] For the RumorEval-19 dataset, the evaluation metrics used in this invention are consistent with the official competition criteria, namely Macro-F1 and RMSE. For the Weibo-comp dataset, four common classification metrics are selected: accuracy, precision, recall, and F1 score. The experimental results are shown in Table 3, with the best-performing results highlighted in bold.

[0173] Table 3 shows the model's performance on the RumorEval-19 and Weibo-comp datasets.

[0174]

[0175] As shown in Table 3, among the baseline methods, firstly, among LSTM, TextCNN, BERT, HAN, and MTEFN methods that extract features only from news content, MTEFN performs best on the RumorEval-19 dataset, while BERT achieves the best performance on the Weibo-comp dataset. This may be because the RumorEval-19 dataset has a more pronounced sentiment bias compared to the Weibo-comp dataset. MTEFN, when jointly trained with news classification and sentiment classification, can more accurately detect fake news. BERT, with its powerful semantic understanding and feature extraction capabilities, performs exceptionally well on the Weibo-comp dataset.

[0176] Secondly, among methods that extract features from both news and commentary levels simultaneously, dEFEND significantly improves detection performance by employing a sentence-comment co-attention mechanism, outperforming methods that only consider news content. On the RumorEval-19 dataset, dEFEND's Mac-F1 score is 0.6% higher than MTEFN; on the Weibo-comp dataset, dEFEND outperforms MTEFN by 6.6%, 3.4%, 9.3%, and 6.5% in accuracy, precision, recall, and F1 score, respectively. Furthermore, HSA-BLSTM also outperforms most models that only consider news text. These results demonstrate that combining news and commentary features is more effective in fake news detection than relying solely on news content.

[0177] Among methods that consider comment information, HSA-BLSTM and dEFEND focus on mining semantic features from comments, while Dual-Emo and AGWu focus on extracting sentiment features. Dual-Emo, by mining bidirectional sentiment signals from news articles and comments, improves accuracy, precision, recall, and F1 score by 6.8%, 4.1%, 8.7%, and 6.5% respectively compared to MTEFN on the Weibo-comp dataset. On the RumorEval-19 dataset, Dual-Emo improves Mac-F1 by 1.3% and reduces RMSE by 5.5%. While AGWu utilizes deep learning techniques to extract dual sentiment signals and combines an adaptive genetic weight update mechanism with random forests for feature filtering and classification, the lack of direct training in the feature extraction part may explain its poor performance on both datasets.

[0178] In summary, the proposed method outperforms the comparative model on both datasets. For example, compared to the best-performing comparative model (Dual_Emo), SECS improves the F1 score by 2.9% on the Weibo-comp dataset and achieves a 2.4% improvement in the Mac-F1 score on the RumorEval-19 dataset. Experimental results on both datasets demonstrate that the proposed method performs well in fake news detection tasks, further supporting the necessity of mining semantic and sentiment features from the comment section perspective for fake news detection.

[0179] To test the effectiveness of each component of the method of the present invention, five ablation experiments were set up.

[0180] The first group verifies the effectiveness of the news encoding module. The news encoding module is removed, and the news-level text representation obtained directly from the pre-trained module is used as the news embedding for the final news classification module, while other modules remain unchanged. This sub-model is named SECS-ne. The experimental results are shown in Table 4. The classification metrics for both datasets show a decrease, indicating that the sentence-level news semantic learning module helps the model understand the relationships between sentences within a news article.

[0181] The second sub-model removes the comment representation learning based on reading habits and directly uses the initial comment representation as input for the comment evolution learning part and the comment semantics and user sentiment fusion module. This sub-model is named SECS-rhcl. The results are shown in Table 4. The model performance decreases on both datasets, with a more significant decrease on the RumorEval-19 dataset, where Mac-F1 and RMSE decrease by 9.5% and 19.9%, respectively. Therefore, considering multiple segments of the news text and selectively performing comment enhancement based on their relevance to the comments helps improve the performance of fake news detection.

[0182] The third sub-model, which hides the semantic evolution learning part of comments, concatenates the remaining feature vectors from the news classification module. This sub-model is named SECS-sem. The results are shown in Table 4. On the RumorEval-19 dataset, SECS-sem achieves an 11.7% reduction in Mac-F1 score compared to SECS; on the Weibo-comp dataset, it achieves a 3.5% reduction in F1 score. This indicates that capturing user semantic changes from the perspective of the comment section helps improve the accuracy of fake news detection.

[0183] In the fourth group, the user sentiment evolution learning part was removed, and the news classification module concatenated the remaining feature vectors. This sub-model was named SECS-eem. The results are shown in Table 4. Compared with SECA, the various indicators of SECS-eem also decreased to varying degrees, which means that constructing the correlation of sentiment changes among different users participating in news is beneficial to performance improvement.

[0184] The fifth group evaluated whether designing a gating mechanism to fuse user semantic and sentiment information helped improve model performance. Therefore, the comment semantic and sentiment fusion module was removed, and the remaining features were concatenated during the news classification stage. This sub-model was named SECS-fce. The results are shown in Table 4. Compared to the RumorEval-19 dataset, SECS-fce showed a significant performance drop on the Weibo-comp dataset. This is related to the fact that the Weibo-comp dataset has a larger number of comments, capturing more semantic information and sentiment signals than RumorEval-19.

[0185] Table 4 Comparison of detection performance of the SECS model and its sub-models

[0186]

[0187] To evaluate the impact of different parameter settings on the performance of this invention, we conducted a parameter sensitivity analysis, mainly focusing on the maximum number of comments, the maximum number of news sentences, and the hidden layer dimension d of the LSTM in the comment semantics and user sentiment fusion module. ce Three parameters.

[0188] Since the number of sentences and comments varies across different news items, setting different maximum numbers of comments and news sentences can affect model performance. In configuring these two parameters, we considered both truncation and padding strategies. Based on the number of sentences and comments for each news item in the dataset, four parameter settings were implemented: maximum, minimum, average, and the average of the maximum and minimum values. Specifically, in the RumorEval-19 dataset, experiments were conducted with the maximum number of comments set to 193, 1, 18, and 97, and the maximum number of news sentences set to 9, 1, 3, and 5. In the Weibo-comp dataset, experiments were conducted with the maximum number of comments set to 45, 5, 16, and 25, and the maximum number of news sentences set to 63, 1, 3, and 32.

[0189] For d ce For the parameters, we selected some common LSTM hidden layer dimensions for parameter sensitivity analysis. We then sequentially analyzed the parameters of d... ce Experiments were conducted with values ​​of 32, 64, 128, and 256.

[0190] Figure 4 and Figure 5 The results show that for both datasets, the model achieves optimal performance when the maximum number of comments and the maximum number of news sentences are set to their respective maximum values ​​in the datasets; furthermore, the model performance improves as the values ​​of these two parameters increase. This indicates that although padding introduces some data noise, compared to truncation, padding effectively avoids the omission of key information and contributes to improved model performance.

[0191] also, Figure 6 This shows that when the comment and user sentiment fusion module is used, the LSTM hidden layer dimension d ce The model performs optimally when the hidden layer dimension is set to 64. When the hidden layer dimension is too small, LSTM struggles to fully capture the sequential relationships between comment semantics and sentiment; conversely, when the dimension is too large, redundant features are introduced, increasing noise and interfering with the capture of key information. ce Setting it to 64 provides LSTM with sufficient expressive power, enabling it to effectively learn the internal correlation between sentiment and semantics in fused features, thereby improving the overall performance of the model.

[0192] These results demonstrate that maximizing data structure information and balancing feature representation capabilities are crucial for improving model performance in sentiment classification tasks.

[0193] To evaluate the effectiveness of the user sentiment evolution learning module, this invention uses data from the RumorEval-19 and Weibo-comp datasets to deeply analyze the sentiment evolution features extracted by the module. Specifically, this invention reduces the dimensionality of the sentiment evolution features to a three-dimensional space using t-SNE to observe their scatter distribution in the three-dimensional space, and compares and verifies the feature distribution with the actual sentiment changes in the comment section. The sentiment changes in the comment section are visualized by plotting the dominant sentiment of each comment, thus intuitively presenting the sentiment evolution pattern of the comment section. This invention selects two cases for analysis: one with strong sentiment consistency in comments and the other with significant sentiment reversal in comments.

[0194] In the case of the Weibo-comp dataset, such as Figure 7 The sentiment in the comments section is highly neutral, with the sentiment distribution mainly concentrated in one area. This indicates strong consistency in sentiment and no significant emotional fluctuations. This characteristic is consistent with the corresponding... Figure 8 The sentiment change graph echoes this, clearly demonstrating the stability of sentiment in the comment section. This result indicates that the user sentiment evolution learning module can accurately capture the stable patterns of sentiment in the comment section and effectively learn sentiment consistency characteristics.

[0195] In contrast, the Weibo-comp dataset shows a significant shift in sentiment in the comment section in the case of sentiment reversal. Figure 9 The distribution of emotional evolution characteristics reveals that emotions gradually shift from an initial "neutral" state to "pessimistic" negative emotions, and then gradually return to "neutral." This alternating evolutionary feature between "neutral" and "pessimistic" states is consistent with... Figure 10 The sentiment fluctuations in the comment sentiment change graph are consistent. Furthermore, in another case using the RumourEval-19 dataset, such as... Figure 11-13 As shown, the sentiment in the comment section gradually transitions from negative to positive, and the corresponding sentiment change graph also shows a pattern of sentiment gradually developing in a positive direction over time. To more intuitively observe the sentiment changes in the comment section, this invention, based on the sentiment reversal case from the RumorEval-19 dataset, plotted a sentiment win-loss graph on top of the sentiment change graph. See Figure 13 .

[0196] according to Figure 13The results show that before the 97th comment, although the cumulative sentiment score fluctuated, it remained negative, indicating that most comments before this point carried negative sentiment. However, with the appearance of the 97th comment, the curve crossed zero, and the cumulative score turned from negative to positive, marking a reversal in sentiment. After the sentiment reversal, the cumulative sentiment score rose rapidly and significantly, indicating a strong reversal. The continued accumulation of positive sentiment scores also reflects the gradual recovery of public sentiment from negative to positive.

[0197] These cases clearly demonstrate that the user emotion evolution learning module can accurately capture patterns of emotional changes in the comment section. Whether in scenarios of emotional consistency or emotional reversal, it effectively reflects the flow and evolution of emotions within the comment section. Through the application of this module, we can gain a deeper understanding of the evolution of user emotions and their actual manifestation in social interactions.

[0198] In summary, this invention comprehensively utilizes comment data and studies the problem of fake news detection from the perspective of the comment section. It innovatively designs a fake news detection method to explore the evolution patterns of comment semantics and user sentiment. Experimental results show that the model performs well on both Chinese and English datasets, effectively improving the efficiency of fake news detection.

[0199] This invention also provides a fake news detection device; the specific device may include:

[0200] The initial feature representation module is used to obtain the target news and its comment set, and extract the news text embedding, comment set text embedding, and comment set sentiment embedding;

[0201] The news encoding module is used to perform context-related encoding representation on the embedded news text to obtain a news encoded representation;

[0202] The comment representation learning module is used to perform news reinforcement learning on the comment set text embedding based on the news text embedding to obtain a personalized representation of the comment set text;

[0203] The semantic sentiment fusion module is used to fuse the personalized representation of the comment set text with the sentiment embedding of the comment set to obtain the semantic sentiment fusion vector representation of the comments;

[0204] The news classification module is used to obtain semantic evolution features and user sentiment evolution features based on the personalized representation of the comment set text and the sentiment embedding of the comment set, and to perform fake news classification prediction by combining the news text embedding and the semantic-sentiment fusion vector representation of the comments.

[0205] The fake news detection device in this embodiment is used to implement the aforementioned fake news detection method. Therefore, the specific implementation of the fake news detection device can be found in the embodiment section of the fake news detection method above. For example, the initial feature representation module, news encoding module, comment representation learning module, semantic sentiment fusion module, and news classification module are used to implement steps S101, S102, S103, S104, and S105 in the above-mentioned fake news detection method, respectively. Therefore, its specific implementation can be referred to the description of the corresponding embodiments, which will not be repeated here.

[0206] A specific embodiment of the present invention also provides a fake news detection device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the aforementioned fake news detection method.

[0207] A specific embodiment of the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the aforementioned fake news detection method.

[0208] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0209] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0210] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0211] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0212] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for detecting fake news, characterized in that, include: Obtain the target news and its comment set, and extract the news text embedding, comment set text embedding, and comment set sentiment embedding; The news text embedding is subjected to context-associative encoding to obtain the news encoded representation; Based on the news text embedding, news reinforcement learning is performed on the comment set text embedding to obtain a personalized representation of the comment set text; The personalized representation of the comment set text and the vector corresponding to each comment in the comment set sentiment embedding are fused through a gating mechanism to obtain the fusion feature of each comment. The fusion features of all comments are input into the first long short-term memory network in the order of the comments, and the hidden layer state of the last time step is selected as the semantic sentiment fusion vector representation of the comment. The personalized representation of the comment set text is input into the second long short-term memory network in the order of the comments, and the hidden layer state of the last time step is selected as the semantic evolution feature. The sentiment embedding of the comment set is input into a third long short-term memory network in chronological order of the comments, and the hidden layer state of the last time step is selected as the user sentiment evolution feature. The system fuses news encoding representation, semantic evolution features, user sentiment evolution features, and comment semantic sentiment fusion vector representation, and then uses the fused features to predict fake news classification.

2. The method for detecting fake news according to claim 1, characterized in that, The extraction of news text embeddings, comment set text embeddings, and comment set sentiment embeddings includes: The text embedding of each sentence in the target news is obtained by using the first pre-trained model; The text embedding of each comment in the target news comment set is obtained by using a second pre-trained model; Each comment in the target news comment set is classified by sentiment to obtain the sentiment embedding of the comment set.

3. The method for detecting fake news according to claim 1, characterized in that, The process of performing context-related encoding on the embedded news text to obtain the news encoded representation includes: The news text embedding is encoded using a bidirectional long short-term memory network to obtain the preceding state sequence information and the following state sequence information; The hidden state of the last time step in the preceding state sequence information and the hidden state of the last time step in the following state sequence information are concatenated to obtain the news encoding representation.

4. The method for detecting fake news according to claim 1, characterized in that, The step of performing news reinforcement learning on the comment set text embedding based on the news text embedding to obtain a personalized representation of the comment set text includes: Based on the attention mechanism, news reinforcement learning is performed on each comment text embedding according to the news text embedding to obtain a personalized representation of the comment set text.

5. The method for detecting fake news according to claim 1, characterized in that, The step of fusing the personalized representation of the comment set text with the sentiment embedding of the comment set to obtain the comment semantic sentiment fusion vector representation includes: The personalized representation and sentiment embedding of each comment text are concatenated to generate a gated vector; The personalized representation and sentiment embedding of each comment text are linearly transformed and mapped to the same space, generating comment semantic vectors and sentiment vectors with the same dimensions; Generate fused comment features based on the gating vector, the comment semantic vector, and the sentiment vector; The fused comment feature sequence of the comment set is input into a long short-term memory network to generate state sequence information, and the hidden layer state of the last time step in the state sequence information is selected as the comment semantic sentiment fusion vector.

6. The method for detecting fake news according to claim 1, characterized in that, The step of obtaining semantic evolution features and user sentiment evolution features based on the personalized representation of the comment set text and the sentiment embedding of the comment set includes: The personalized representation of the comment set text and the sentiment embedding of the comment set are respectively input into a long short-term memory network to generate corresponding state sequence information, and the hidden state layer of the last time step is selected to generate semantic evolution features and user sentiment evolution features.

7. The method for detecting fake news according to claim 1, characterized in that, The process of classifying and predicting fake news includes: The semantic evolution features, user sentiment evolution features, news text embeddings, and comment semantic sentiment fusion vectors are concatenated, and the correlation between features is calculated through a fully connected network and softmax function transformation, and finally the prediction vector is output.

8. A fake news detection device, characterized in that, include: The initial feature representation module is used to obtain the target news and its comment set, and extract the news text embedding, comment set text embedding, and comment set sentiment embedding; The news encoding module is used to perform context-related encoding representation on the embedded news text to obtain a news encoded representation; The comment representation learning module is used to perform news reinforcement learning on the comment set text embedding based on the news text embedding to obtain a personalized representation of the comment set text; The semantic sentiment fusion module is used to fuse the personalized representation of the comment set text and the vector corresponding to each comment in the sentiment embedding of the comment set through a gating mechanism to obtain the fusion feature of each comment. The fusion features of all comments are input into the first long short-term memory network in the order of the comments, and the hidden layer state of the last time step is selected as the semantic sentiment fusion vector representation of the comment. The temporal evolution module is used to input the personalized representation of the comment set text into the second long short-term memory network according to the time order of the comments, and select the hidden layer state of the last time step as the semantic evolution feature. The sentiment embedding of the comment set is input into a third long short-term memory network in chronological order of the comments, and the hidden layer state of the last time step is selected as the user sentiment evolution feature. The news classification module is used to fuse news encoding representation, semantic evolution features, user sentiment evolution features, and comment semantic sentiment fusion vector representation, and to predict fake news classification based on the fused features.

9. A fake news detection device, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of a fake news detection method as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the fake news detection method as described in any one of claims 1 to 7.

Images