False news detection method and device based on user emotion and comment semantic mining

Abstract

The invention relates to the technical field of false news detection, in particular to a false news detection method, equipment and device based on user emotion and comment semantic mining and a computer storage medium. According to the false news detection method, firstly, initial feature representation is carried out on news texts, comments and user emotions, secondly, news representation associated with context is learned according to sentence-level news initial embedding, and news fine-grained semantics are captured for reinforcement learning of subsequent comments at the same time. Thirdly, integrating consideration of different attention of news text sentences to obtain comment semantic embedding of news enhancement. Fourthly, by taking the comments as units, designing a gating mechanism to perform feature fusion on comment semantics and user emotions, and mining associated features of the comments and the user emotions; and 5, capturing change modes of comment semantics and user emotions in the comment area through a long-short-term memory network. And finally, integrating features extracted by each module, and carrying out news category judgment. The accuracy of false news detection is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fake news detection, and particularly to a fake news detection method, device, apparatus and computer storage medium based on user emotion and comment semantic mining. Background Art

[0002] The booming development of social media has changed the way information is disseminated. Through social media, information dissemination is no longer restricted by time and space, and users can easily post content and interact with others. However, it is precisely this convenience and immediacy that, while improving the efficiency of information dissemination, also provides a channel for the rapid spread of fake news.

[0003] The widespread dissemination of fake news has had various negative impacts on society. First of all, it seriously endangers public safety and social order. Secondly, the proliferation of fake news has exacerbated the public's trust crisis in the news. According to a Reuters survey, only 42% of users trust the news they see online, and this lack of trust is partly due to the widespread dissemination of fake news in multiple fields such as politics, economy, and health. Thirdly, the widespread dissemination of fake news has weakened the credibility of the government. Previous studies have also found that fake news can effectively confuse voters' views and affect election results. Therefore, curbing the spread of fake news and reducing its harm to society has become an urgent task to be solved.

[0004] Currently, fake news detection faces many challenges. First of all, fake news is often deliberately edited and published to deceive the public, but with the development of large language models and generative artificial intelligence, the boundary between fake news and real news has become increasingly blurred. Therefore, it is very difficult to accurately judge the authenticity of news only relying on the news content. Secondly, the amount of information on social media is huge, the content forms are rich and diverse, and most of them are user-generated, with a large amount of data noise. Therefore, relying on manual detection of fake news is not only very time-consuming but also inefficient.

[0005] Many researchers have conducted extensive explorations in the field of fake news detection. The existing fake news detection methods can be roughly divided into content-based methods and social context-based methods.

[0006] False news detection based on news content mainly focuses on the information within the news itself. Representative methods can be divided into three categories: 1) Methods based on linguistic features, which identify fake news by capturing the writing style differences and semantic differences between true and false news. However, since fake news is often deliberately edited before publication, with the development of large language models (LLMs) and artificial intelligence-generated content (AIGC), the boundary between fake news and real news has become increasingly blurred. To address this challenge, Wu et al. used LLMs to generate news reconstructions with diverse styles and combined content clues, effectively enhancing the robustness and accuracy of fake news detectors when dealing with style attacks driven by LLMs. 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) Methods based on multi-modal feature fusion, which combine data such as text, images, and videos to improve detection performance. Compared with single-modal methods, multi-modal methods can better identify inconsistencies by integrating information from different sources. However, multi-modal methods also face some challenges: collecting and annotating data usually requires significant costs and time; during the inference process, if data from some modalities is missing or incomplete, it may affect the performance of the model.

[0007] Early research on false news detection mainly focused on manually extracting features from the prior knowledge of false news and training machine learning classifiers based on these features. However, this method is labor-intensive and costly, and it is difficult to capture the deep features of fake news. In addition, it is often difficult to accurately judge the authenticity of news relying solely on news content. To address the above problems, many scholars have introduced social context information (such as user behavior, dissemination structure, etc.) and combined deep learning techniques to infer the authenticity of news, achieving remarkable results. However, methods based on social context highly rely on external information, and the collection and processing of this information are complex and require significant computing resources. In contrast, content-based methods only rely on the news text itself and do not require additional information support. Since news text is usually easily accessible, this method is more convenient and efficient in practice. As the most direct feedback from users to news, comment data can provide important clues for false news detection. Some studies have attempted to use comment data to assist in false news detection. For example, dEFEND improves detection performance by introducing a sentence-comment co-attention mechanism to analyze the relevance between news content and user comments and selecting comments that can explain fake news. Hamed et al. provide auxiliary support for false news detection by extracting the emotional features of news titles and combining emoji information in user comments.

[0008] Although existing methods for detecting fake news using user comments have achieved some success, most of them focus on feature mining of individual comments, and few studies explore the emotional and semantic changes from the perspective of the comment section. Exploring the emotional changes in the comment section is of great significance for fake news detection. First of all, the emotional changes in the comment section can directly reflect the public's attitude towards news events. In the topic discussions participated by users, the emotional tendency can directly reflect the positive (happy, optimistic), negative (bored, pessimistic), or neutral (no obvious emotional tendency) attitudes of users, and this emotional information can help infer the authenticity of the news. For example, research shows that compared with real news, fake news is more likely to trigger negative public emotions. Secondly, the evolution trend of the emotion in the comment section can also reveal the authenticity of the news. Taking an item of information in the Weibo-comp dataset as an example, such as Figure 1 , over time, the emotion in the comment section gradually changed from the initial "surprised" to "angry" and finally to "neutral". This emotional change helps to reveal the authenticity of the news content. When the emotion in the comment section shows more and more negative emotions, it may indicate that the news has triggered widespread distrust or disgust, thus providing clues for the detection of fake news. Similarly, the evolution of the semantics in the comment section is also an important basis for revealing fake news. For example, Figure 1 , over time, the content discussed in the comment section has changed from the initial "uncertainty and worry" to "questioning and criticism" and finally to "clarifying the fake news". This semantic evolution reflects the process of the public's deepening understanding of the news event. The public may be confused about the authenticity of the event in the early stage, but as the discussion deepens, they gradually form a clearer judgment and clarification. This semantic change also provides a basis for identifying the falsity of the news. Existing research such as the Post2DR model proposed by Yang et al. although uses CNN to model the local changes of comment semantics, it lacks the capture of the global perspective of comment semantic changes. From Figure 1 , it can also be observed that news texts usually contain rich information, and the discussions triggered by news are often diverse. For example, some users are concerned about food safety issues, while others are concerned about consumer rights. The news article provides the detailed background and context of the event, which helps to accurately interpret the semantics and emotions of the comments. However, the current research on comment representation learning usually treats comments as independent entities and only focuses on the comment text itself. Although this method is simple, it fails to reveal the internal connection between comments and news content.

[0009] In summary, the accuracy of existing methods for detecting fake news using user comments is relatively low, which is a problem to be solved at present. Summary of the Invention

[0010] To this end, the technical problem to be solved by the present invention is to overcome the problem of low accuracy of the method for detecting fake news by using user comments in the prior art.

[0011] To solve the above technical problem, the present invention provides a method for detecting fake news, including:

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

[0013] Perform context-related encoding representation on the news text embedding to obtain a news encoding representation;

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

[0015] Fuse the personalized representation of the comment set text and the comment set sentiment embedding to obtain a comment semantic sentiment fusion vector representation;

[0016] According to the personalized representation of the comment set text and the comment set sentiment embedding, 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.

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

[0018] Obtain the text embedding of each sentence in the target news through a first pre-trained model to obtain the news text embedding;

[0019] Obtain the text embedding of each comment in the target news comment set through a second pre-trained model to obtain the comment set text embedding;

[0020] Perform sentiment classification on each comment in the target news comment set to obtain the comment set sentiment embedding.

[0021] Preferably, the performing context-related encoding representation on the news text embedding to obtain a news encoding representation includes:

[0022] Use a bidirectional long short-term memory network to encode the news text embedding to obtain pre-order state sequence information and post-order state sequence information;

[0023] Concatenate the hidden layer state of the last time step in the pre-order state sequence information and the hidden layer state of the last time step in the post-order state sequence information to obtain a news encoding representation.

[0024] Preferably, the personalized representation of the comment set text obtained by performing news reinforcement learning on the comment set text embedding according to the news text embedding 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 the personalized representation of the comment set text.

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

[0027] Concatenate the personalized representation and the sentiment embedding corresponding to each comment text, and generate a gating vector;

[0028] Perform a linear transformation on the personalized representation and the sentiment embedding corresponding to each comment text to map them to the same space, and generate a comment semantic vector and a sentiment vector with the same dimension;

[0029] Generate a fused comment feature according to the gating vector, the comment semantic vector, and the sentiment vector;

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

[0031] Preferably, the obtaining of the semantic evolution feature and the user sentiment evolution feature according to the personalized representation of the comment set text and the comment set sentiment embedding includes:

[0032] Input the personalized representation of the comment set text and the comment set sentiment embedding into a long short-term memory network respectively to generate corresponding state sequence information, and select the hidden state layer at the last time step to generate the semantic evolution feature and the user sentiment evolution feature.

[0033] Preferably, the false news classification prediction includes:

[0034] Concatenate the semantic evolution feature, the user sentiment evolution feature, the news text embedding, and the comment semantic sentiment fusion vector, and calculate the correlation relationship between the features through a fully connected network and a softmax function transformation, and finally output a prediction vector.

[0035] The present invention also provides a false news detection device, including:

[0036] An initial feature representation module, configured to obtain a target news and its comment set, and extract a news text embedding, a comment set text embedding, and a comment set sentiment embedding;

[0037] A news encoding module, which is used to perform context - associated encoding representation on the news text embedding to obtain a news encoding representation;

[0038] A comment representation learning module, which is used to perform news - enhanced learning on the comment set text embedding according to the news text embedding to obtain a personalized representation of the comment set text;

[0039] A semantic - emotion fusion module, which is used to fuse the personalized representation of the comment set text and the comment set emotion embedding to obtain a comment semantic - emotion fusion vector representation;

[0040] A news classification module, which is used to obtain semantic evolution features and user emotion evolution features according to the personalized representation of the comment set text and the comment set emotion embedding, and combine the news text embedding and the comment semantic - emotion fusion vector representation to perform false news classification prediction.

[0041] The present invention also provides a false news detection device, including:

[0042] A memory, which is used to store a computer program;

[0043] A processor, which is used to implement the steps of the above - mentioned false news detection method when executing the computer program.

[0044] The present invention also provides a computer - readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the above - mentioned false news detection method are implemented.

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

[0046] For the false news detection method of the present invention, firstly, initial feature representations of news texts, comments and user emotions are performed. Secondly, according to the news initial embedding at the sentence level, context - associated news representations are learned to capture the fine - grained semantics of news for subsequent enhanced learning of comments. Thirdly, the news - enhanced comment semantic embedding is obtained by considering different attentions of news text sentences. Fourthly, a gating mechanism is designed for each comment to fuse the comment semantics and user emotions and mine the associated features of the two. Fifthly, the change patterns of comment semantics and user emotions in the comment area are captured through a long - short - term memory network. Finally, the features extracted by each module are comprehensively used to determine the news category. The present invention improves the accuracy of false news detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to make the content of the present invention easier to be clearly understood, the following further details the present invention according to the specific embodiments of the present invention in combination with the drawings, where:

[0048] Figure 1 It is a schematic diagram of an example from the Weibo-comp dataset;

[0049] Figure 2 It is a flowchart for implementing a false news detection method provided by the present invention;

[0050] Figure 3 It is a flowchart for implementing a false news detection method provided by an embodiment of the present invention;

[0051] Figure 4 It is a schematic diagram of the performance of different settings of the maximum number of comments on two datasets;

[0052] Figure 5 It is a schematic diagram of the performance of different settings of the maximum number of news sentences on two datasets;

[0053] Figure 6 It is about d ce Schematic diagram of the performance of different settings on two datasets;

[0054] Figure 7 It is a distribution diagram of the emotional evolution characteristics of the comment area of the emotional consistency case in the Weibo-comp dataset;

[0055] Figure 8 It is a graph of the emotional change in the comment area of the emotional consistency case in the Weibo-comp dataset;

[0056] Figure 9 It is a distribution diagram of the emotional evolution characteristics of the comment area of the emotional reversal case in the Weibo-comp dataset;

[0057] Figure 10 It is a graph of the emotional change in the comment area of the emotional reversal case in the Weibo-comp dataset;

[0058] Figure 11 It is a distribution diagram of the emotional evolution characteristics of the comment area of the emotional reversal case in the RumourEval-19 dataset;

[0059] Figure 12 It is a graph of the emotional change in the comment area of the emotional reversal case in the RumourEval-19 dataset;

[0060] Figure 13 It is a graph of the cumulative emotional score in the comment area of the emotional reversal case in the RumourEval-19 dataset. Specific implementation manners

[0061] The core of the present invention is to provide a false news detection method, device, equipment and computer storage medium, which effectively improves the accuracy of false news detection.

[0062] To enable those skilled in the art to better understand the solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention.

[0063] Please refer to Figure 2 , Figure 2 which is a flowchart of the implementation of a false news detection method provided by the present invention; the specific operation steps are as follows:

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

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

[0066] S103: Perform news reinforcement learning on the comment set text embedding according to the news text embedding to obtain the personalized representation of the comment set text;

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

[0068] S105: Obtain the semantic evolution feature and the user sentiment evolution feature according to the personalized representation of the comment set text and the comment set sentiment embedding, and combine the news text embedding and the comment semantic sentiment fusion vector representation to perform false news classification prediction.

[0069] Please refer to Figure 3 , Figure 3 which is a flowchart of the implementation of a false news detection method provided by an embodiment of the present invention.

[0070] Based on the above embodiments, this embodiment will elaborate on step S101:

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

[0072] Obtain the text embedding of each sentence in the target news through a first pre-trained model to obtain the news text embedding;

[0073] Obtain the text embedding of each comment in the target news comment set through a second pre-trained model to obtain the comment set text embedding;

[0074] Perform sentiment classification on each comment in the target news comment set to obtain the sentiment embedding of the comment set.

[0075] The content of fake news is usually related to current events or hot topics. In order to attract people's attention and extensive discussion, it will be deliberately modified before being released. Since the text content usually revolves around a certain event or topic, the context is often highly relevant. Therefore, capturing the context 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 can be, for example, a model that can capture the context information of news content to improve the quality of text embedding. This embodiment does not limit the specific model.

[0077] The proposed Bert model has achieved excellent performance in many fields such as natural language processing and image classification. In terms of text representation learning, compared with previous models, Bert is based on a bidirectional transformer, fully utilizes text context information, and greatly alleviates the problem of ambiguous word embeddings. At the same time, it uses a self-attention mechanism to weight words to obtain a vector representation of the sentence, and better considers the long-distance context dependence problem.

[0078] In one embodiment, the first pre-trained model selects a pre-trained Bert model for news embedding. For news P = {s1, s2, s3, s4, s5... s M} composed of M sentences, where Obtain sentence-level embeddings through the pre-trained model Bert T P ∈R M×d 。Among them is the text embedding of the i-th sentence, and d is the dimension of the sentence embedding. The calculation method of is as follows:

[0079]

[0080] Among them, PreModel represents the pre-trained model Bert. For English text, select the model bert-base-uncased trained on an unlabeled English corpus; for Chinese text, select the Chinese-Roberta-wwmext model to embed news text.

[0081] In order to map comments from text form to a computer-processable vector form 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 selects the Bert pre-trained model to encode the comments. For the comment set C = {c1, c2, c3,..., c |C|} corresponding to the news, where c i represents the i-th comment, and the comment text embedding is where, T C ∈ R |C|×d , d is the text embedding dimension, is the text embedding of the i-th comment. The specific calculation method of

[0083]

[0084] is as follows: where PreModel represents the pre-trained model. Similar to news embedding, the selected model varies according to the text language type.

[0085] Comments are the main way for people to participate in topic discussions and usually contain rich emotional signals. In order to extract emotional information from users' comments and convert it into digital vectors, this embodiment does not limit the specific emotion classification method.

[0086] In one embodiment, for English text, the widely used Ekman-based emotion classification in the field is selected, and emotions are divided into six categories: {Anger, Disgust, Fear, Joy, Sadness, Surprise}; for Chinese data, through the emotion analysis service provided by Baidu Smart Cloud, the text emotions are divided into nine categories: {Boredom, Sadness, Like, Neutral, Optimistic, Fear, Happy, Angry, Pessimistic}. The user emotion feature set is represented as E c = {e1, e2, e3,... e |C|}, for the comment c i , its emotion vector is e i , and the calculation method of e i is as follows:

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

[0088] where, e i ∈ R 1×k , and the values in the vector represent the probability that the text belongs to the corresponding emotion category, and k is the number of emotion categories.

[0089] Based on the above embodiments, this embodiment details step S102:

[0090] The initial representation of news text can better capture the semantics within a sentence. To better learn the context correlation at the news sentence level, context correlation encoding representation is performed on the news text embedding, and the specific encoding method is not limited in this embodiment.

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

[0092] Encoding the news text embedding using a bidirectional long short-term memory network to obtain the previous state sequence information and the subsequent state sequence information:

[0093] Using bidirectional LSTM to further encode the news. Taking the news sentence embedding sequence as the input of the bidirectional LSTM, and obtaining the previous information and the subsequent information through encoding. Among them, the forward state vector at the t-th time step is The backward state vector is The formal expression is as follows:

[0094]

[0095] Taking the forward network calculation as an example, The calculation process of is obtained through the forget gate, input gate, memory update unit, and output gate structures of the forward long short-term memory network, and the calculation process is as follows:

[0096]

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

[0098] Concatenating the hidden layer state at the last time step in the previous state sequence information and the hidden layer state at the last time step in the subsequent state sequence information to obtain the news representation T after context sentence correlation

[0099] s .

[0100]

[0101] Among them, d s is the text embedding dimension in the news encoding stage.

[0102] Based on the above embodiments, this embodiment details step S103: ​

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

[0104] Inspired by the work of Wu et al., when fusing the features of pictures and text semantics, this study takes into account the reading habits of readers. From the perspective of the public's reading habits, when users read news texts, their attention usually focuses on certain paragraphs, and the published comment content is often only related to some paragraphs of the news, and different users focus on different information. Therefore, the present invention uses the news content as the context information of the comments and applies the attention mechanism to learn the personalized comment representation enhanced by the news.

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

[0106] Based on the attention mechanism, performing news enhanced learning on each comment text embedding according to the news text embedding to obtain the personalized representation of the comment set text:

[0107] Taking the initial embedding of the news and the initial embedding of the comment as inputs, through the attention mechanism, obtaining the comment representation sequence enhanced by the news where is the i-th comment representation enhanced by the news. Specifically, through the input initial embeddings T P and T C linear transformations are performed to obtain the query matrix Q, the key matrix K, and the value matrix V. The calculation processes of the matrices Q, K, and V are as follows:

[0108]

[0109] where, W q , W k , W v are the parameters of the linear transformation. The attention mechanism measures the influence of each sentence s j in the news on the comment c i under the given news paragraph, and calculates the attention score using the following formula

[0110]

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

[0112]

[0113] where d c is the embedding dimension of.

[0114] Based on the above embodiments, this embodiment details step S104:

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

[0116] In one embodiment, inspired by the GRU update gate control mechanism, the present invention innovatively introduces a gating mechanism to fuse the semantics and sentiment of comments, so as to more comprehensively understand the comment data:

[0117] Fusing the personalized representation of the comment set text and the comment set sentiment embedding, the obtained comment semantic sentiment fusion vector representation includes:

[0118] Concatenate the personalized representation and sentiment embedding corresponding to each comment text, and generate a gating vector:

[0119] In one specific embodiment, the comment embedding is concatenated with the sentiment vector e t , and after passing through a fully connected layer and generating a gating vector u through the sigmoid function t , u t is used to control the retention degree of semantic and sentiment information.

[0120]

[0121] Map the personalized representation and sentiment embedding corresponding to each comment text to the same space through linear transformation, and generate comment semantic vectors and sentiment vectors with the same dimension:

[0122] In one specific embodiment, the semantic vector and the sentiment vector e t are linearly transformed and mapped to the same space, and then the sentiment vector with the same dimension is obtained through the hyperbolic tangent function transformation and the comment semantic vector h t :

[0123]

[0124]

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

[0126] In a specific embodiment, combine the gating vector u t Comprehensively consider the comment semantic vector and the sentiment vector, and finally obtain a fused feature that includes the semantic information and sentiment information of the comment

[0127]

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

[0129] In a specific embodiment, learn the internal correlation of the fused feature sequence through a long short-term memory network (LSTM) of:

[0130]

[0131] Take the representation at the last time step as the semantic and sentiment fusion vector of the overall comment, denoted by . Among them, d ce is the dimension of.

[0132] Based on the above embodiments, this embodiment details step S105:

[0133] For all comments on a piece of news, if the emotional and semantic features of these comments are analyzed separately, some static and isolated results can be obtained, and it is impossible to understand the overall semantic and emotional change trends in the comment area, nor can the deep emotional motivation and group resonance effect behind the comments be analyzed.

[0134] This embodiment starts from the perspective of the comment area, aiming to capture the patterns of semantic and emotional changes among users participating in the news discussion.

[0135] In an embodiment, obtain the semantic evolution feature and the user emotional evolution feature according to the personalized representation of the comment set text and the comment set emotion embedding, including:

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

[0137] Specifically:

[0138] The comment representation enhanced by news and the sentiment representation E c ={e1, e2, e3,..., e |C|} are used as module inputs, and the correlation features of the input sequence are learned through two LSTM networks respectively. The hidden state at the last time step contains the information of each time step of the input sequence from front to back. Therefore, the hidden state at this time step is taken as the semantic change feature h c . The calculation process of the sentiment change feature is similar to that of the semantic change feature. The formal expressions of the comment semantic evolution feature h c and the user sentiment evolution feature h e are as follows:

[0139] h c = LSTM(T pec )#(22)

[0140] h e = LSTM(E c )#(23)

[0141] Based on the above embodiments, we obtain the news embedding T s , the comment semantic evolution feature h c , the user sentiment evolution feature h e and the semantic sentiment fusion feature

[0142] In one embodiment, the false news classification prediction includes:

[0143] Concatenate the semantic evolution feature, the user sentiment evolution feature, the news text embedding and the comment semantic sentiment fusion vector, and calculate the correlation relationship between the features through the fully connected network and the softmax function transformation, and finally output the prediction vector.

[0144] Specifically:

[0145] Concatenate these four feature vectors, and then calculate the correlation relationship between the features through a layer of fully connected network and the softmax function transformation, and finally output the prediction vector The value of the vector represents the probability that the news belongs to different categories. The calculation process is as follows:

[0146]

[0147] Among them, concate is the splicing operation, and W p is the parameter matrix of the linear transformation, and b p is the bias term of this layer.

[0148] The cross-entropy loss function is used to train the model as follows:

[0149]

[0150] Among them, 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 in the present invention to verify the effectiveness of the model includes an English dataset and a Chinese dataset.

[0152] The English dataset is RumourEval-19, which is derived from the rumor detection task of the SemEval-2019 competition. The news in the dataset is from Twitter and Reddit platforms and is released by Gorrell et al. The news in this dataset includes three categories: real, fake, and unverified. In the experiments of the present invention, the dataset division method is the same as that of the official competition questions, and the ratio of the training set, test set, and validation set is 7:2:1.

[0153] The Chinese dataset is Weibo-Comp, which is provided by the "Network Rumor Detection during the Epidemic" competition organized by the CCF Big Data Working Group and covers news content in 8 fields on the Weibo platform, including technology, politics, business and finance, social life, culture and entertainment, sports and examinations, and medicine and health, as well as some content related to the epidemic.

[0154] The news and comment-related information of the RumourEval-19 and Weibo-comp datasets are shown in Table 1:

[0155] Table 1 Basic situation of the RumourEval-19 and Weibo-comp datasets

[0156] Dataset RumourEval-19 Weibo-Comp Total number of news 445 6444 Fake news 185 3581 True news 138 2863 Unverified news 122 - Number of comments 7990 102372 Average number of comments 18 16

[0157] The present invention selects multiple representative false news detection methods for comparison. The comparison models are as follows:

[0158] LSTM: LSTM is a deep learning model, and the input is sequence data. In the false news detection task, usually the word sequence and semantic embedding are used as the input, and the average pooling operation is performed on the results calculated by the network to obtain the final text representation.

[0159] Text_CNN: Text_CNN is a model that first applies a convolutional network to the text classification task and learns news semantic representations at different granularities by setting the size of the convolutional kernel.

[0160] BERT: Based on the pre-trained model, this method fine-tunes the fully-connected layer network to adapt to the fake news detection task.

[0161] dEFEND: dEFEND is an interpretable fake news detection framework that adopts a sentence-comment co-attention mechanism to improve the performance of fake news detection and makes full use of user comments to explain 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 through a hierarchical attention network. This method has been widely compared in the misinformation detection task.

[0164] Dual_Emo: This model conducts fake news detection from the perspective of emotion, mines the relationship between the emotions of the publisher and users based on the text content of news and comments to improve the accuracy of fake news detection. This method takes HSA-BLSTM as the baseline model and introduces bidirectional emotion features. After adding bidirectional emotion features, the model performance is further improved compared with HSA-BLSTM.

[0165] MTEFN: This model takes emotion classification as an auxiliary task, jointly constructs a fake news detection framework for emotion classification and news classification tasks, aiming to improve the model's ability to extract key features from data, thereby improving the accuracy of fake news detection.

[0166] AGWu: This model is a fake news detection model that considers both emotion features and text features. AGWu proposes a dual emotion feature extraction mechanism based on deep normalization attention and designs a random forest classifier with adaptive genetic weight update.

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

[0168] For the RumourEval-19 dataset, the dataset partitioning method of the present invention is the same as that of the competition organizers. The ratio of the training set, test set, and validation set is 7:2:1. For the Weibo-comp dataset, the training set, validation set, and test set are partitioned in a ratio of 6:2:2, that is, 60% of the data is used for training, 20% for validation, and 20% for testing. The other parameter settings of the model are shown in Table 2:

[0169] Table 2 Hyperparameter Settings

[0170]

[0171]

[0172] For the RumourEval-19 dataset, the evaluation metrics settings of the present invention are the same as those of the official competition questions, which are Macro-F1 and RMSE. For the Weibo-comp dataset, four common classification metrics, namely accuracy, precision, recall, and F1-score, are selected. The experimental results are shown in Table 3, and the best results are shown in bold:

[0173] Table 3 Performance of the Model on RumourEval-19 and Weibo-comp Datasets

[0174]

[0175] As shown in the experimental results in Table 3, among the baseline methods, first, among the LSTM, TextCNN, BERT, HAN, and MTEFN methods that only extract features from news content, MTEFN performs best on the RumourEval-19 dataset, while BERT achieves the best performance on the Weibo-comp dataset. This may be because, compared with the Weibo-comp dataset, the news sentiment tendency of the RumourEval-19 dataset is more obvious. When jointly training news classification and sentiment classification, MTEFN can detect fake news more accurately. And BERT performs well on the Weibo-comp dataset due to its strong semantic understanding and feature extraction capabilities.

[0176] Secondly, among the methods that extract features from both news and comment aspects simultaneously, dEFEND extracts features through a sentence-comment co-attention mechanism, significantly improving the detection performance, and its performance is better than the methods that only consider news content. On the RumourEval-19 dataset, the Mac-F1 of dEFEND is 0.6% higher than that of MTEFN; on the Weibo-comp dataset, dEFEND leads MTEFN by 6.6%, 3.4%, 9.3% and 6.5% respectively in terms of accuracy, precision, recall and F1 value. In addition, the detection performance of HSA-BLSTM also exceeds most of the models that only consider news text. These results indicate that combining news and comment features is more effective in fake news detection than simply relying on news content.

[0177] Among the methods that consider comment information, HSA-BLSTM and dEFEND focus on mining semantic features from comments, while Dual-Emo and AGWu focus on the extraction of emotional features. By mining the bidirectional emotional signals of news and comments, Dual-Emo improves by 6.8%, 4.1%, 8.7% and 6.5% respectively in terms of accuracy, precision, recall and F1 value compared with MTEFN on the Weibo-comp dataset; on the RumourEval-19 dataset, the Mac-F1 of Dual-Emo increases by 1.3% and the RMSE decreases by 5.5%. Although AGWu uses deep learning technology to extract dual emotional signals and combines random forests through an adaptive genetic weight update mechanism for feature filtering and classification, since the feature extraction part is not directly trained, this may be the reason for the poor performance of this method on the two datasets.

[0178] Generally speaking, the method proposed in the present invention performs better than the comparison models on the two datasets. For example, compared with the best-performing comparison model (Dual_Emo), SECS improves the F1 index by 2.9% on the Weibo-comp dataset and achieves a 2.4% improvement in the Mac-F1 index on the RumourEval-19 dataset. The experimental results on the two datasets show that the method proposed in the present invention has good performance in the fake news detection task, further supporting the necessity of mining comment semantics and emotional features from the perspective of the comment area for fake news detection.

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

[0180] The first group is to verify the effectiveness of the news encoding module. Remove the news encoding module and directly obtain the news-level text representation through the pre-trained module as the news embedding of the final news classification module, while keeping other modules unchanged. Name this sub-model SECS-ne. The experimental results are shown in Table 4. The classification metrics of both datasets have decreased, indicating that the sentence-level news semantic learning module helps the model understand the associations between sentences in the news.

[0181] The second group is to remove the comment representation learning based on reading habits and directly use the initial comment representation as the input for the comment evolution learning part and the comment semantics and user sentiment fusion module. Name this sub-model SECS-rhcl. The results are shown in Table 4. The model performance has decreased on both datasets, and the performance has decreased more significantly on the RumourEval-19 dataset, with Mac-F1 and RMSE decreasing by 9.5% and 19.9% respectively. Therefore, considering multiple segments of news text and selectively enhancing comments according to their relevance to comments helps improve the performance of fake news detection.

[0182] The third group is to mask the comment semantic evolution learning part, and the news classification module concatenates the remaining feature vectors. Name this sub-model SECS-sem. The results are shown in Table 4. On the RumourEval-19 dataset, compared with SECS, the Mac-F1 of SECS-sem has decreased by 11.7%; on the Weibo-comp dataset, the F1 has decreased by 3.5%. This means that capturing the semantic changes of users from the perspective of the comment area helps improve the accuracy of fake news detection.

[0183] The fourth group is to remove the user sentiment evolution learning part, and the news classification module concatenates the remaining feature vectors. Name this sub-model SECS-eem. The results are shown in Table 4. Compared with SECA, the various metrics of SECA-eem have also decreased to varying degrees, indicating that building the association of emotional changes between different users participating in the news is beneficial to performance improvement.

[0184] The fifth group is to evaluate whether designing a gating mechanism to fuse user semantic and emotional information helps improve the model performance. Therefore, remove the comment semantics and sentiment fusion module, and concatenate the remaining features in the news classification stage. Name this sub-model SECS-fce. The results are shown in Table 4. Compared with the RumourEval-19 dataset, the performance of SECS-fce has decreased more significantly on the Weibo-comp dataset. This is related to the fact that the Weibo-comp dataset has more comments, and more semantic information and emotional signals can be captured compared with RumourEval-19.

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

[0186]

[0187] To evaluate the impact of different parameter settings on the performance of the present 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 in different news varies, setting different maximum comment numbers and maximum news sentence numbers may affect the model performance. In the configuration of these two parameters, we considered both truncation and padding strategies. According to the number of sentences and comments in each news in the dataset, four parameter settings were made: maximum value, minimum value, average value, and the average of the maximum and minimum values. Specifically, in the RumourEval-19 dataset, the maximum number of comments was set to 193, 1, 18, and 97 for experiments in sequence, and the maximum number of news sentences was set to 9, 1, 3, and 5 for experiments in sequence. In the Weibo-comp dataset, the maximum number of comments was set to 45, 5, 16, and 25, and the maximum number of news sentences was set to 63, 1, 3, and 32 for experiments in sequence

[0189] For the d ce parameter, we selected some common LSTM hidden layer dimensions for parameter sensitivity analysis. The d ce was set to 32, 64, 128, and 256 for experiments in sequence

[0190] Figure 4 And Figure 5 shows that for the two datasets, when the two parameters of the maximum number of comments and the maximum number of news sentences are set to the corresponding maximum numbers in the dataset, the model performance is optimal; meanwhile, the model performance improves as the values of these two parameters increase. This indicates that although the padding operation will introduce some data noise, compared with the truncation operation, padding can effectively avoid the omission of key information and contribute to the improvement of the model performance

[0191] In addition Figure 6 shows that when the hidden layer dimension d of the LSTM in the comment and user sentiment fusion module ce is set to 64, the model performance is optimal. When the hidden layer dimension is too small, it is difficult for the LSTM to fully capture the sequential relationship between comment semantics and sentiment; on the contrary, when the dimension is too large, redundant features will be introduced, increasing noise and interfering with the capture of key information instead. Setting d ce to 64 provides sufficient expressive power for the LSTM, enabling it to effectively learn the internal correlation between sentiment and semantics in the fusion features, thereby improving the overall performance of the model

[0192] These results indicate that maximizing data structure information and balancing feature expression ability are of great significance for improving the model performance in sentiment classification tasks.

[0193] To evaluate the effectiveness of the user sentiment evolution learning module, the present invention deeply analyzes the sentiment evolution features extracted by this module based on the data of the RumourEval-19 and Weibo-comp datasets. Specifically, the present invention reduces the sentiment evolution features to a three-dimensional space through 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 area. The sentiment changes in the comment area are drawn by visualizing the dominant sentiment of each comment, so as to intuitively present the sentiment evolution pattern in the comment area. The present invention selects two cases with strong comment sentiment consistency and significant comment sentiment reversal respectively for analysis.

[0194] In the case of the Weibo-comp dataset, such as Figure 7 the sentiment in the comment area shows a highly neutral tendency, and the sentiment distribution is mainly concentrated in one area, which indicates that the comment sentiment consistency is strong and there is no obvious sentiment fluctuation. This feature corresponds to the Figure 8 sentiment change graph, clearly showing the stability of the sentiment in the comment area. This result indicates that the user sentiment evolution learning module can accurately capture the stable pattern of the sentiment in the comment area and effectively learn the sentiment consistency features.

[0195] Relatively speaking, in the case of sentiment reversal in the Weibo-comp dataset, it shows a significant turning trend of the sentiment in the comment area. Figure 9 The distribution of the sentiment evolution features reveals that the emotion gradually changes from the initial "neutral" to the negative emotion of "pessimistic", and then gradually returns to "neutral". This alternating evolution feature between "neutral" and "pessimistic" is consistent with the Figure 10 sentiment fluctuations in the comment sentiment change graph. In addition, in another case of the RumourEval-19 dataset, such as Figures 11 - 13 shown, it can be observed that the sentiment in the comment area gradually transitions from negative emotion to positive emotion, and the corresponding sentiment change graph also shows a pattern of the sentiment gradually developing in a positive direction over time. To more intuitively observe the sentiment changes in the comment area, the present invention draws a sentiment win-loss graph based on the sentiment change graph for this sentiment reversal case from the RumourEval-19 dataset. See Figure 13 .

[0196] According to Figure 13From the results shown, it can be observed that before the 97th comment, although the cumulative sentiment score of the comments fluctuated, the score was always negative, indicating that most of the comments before that were negative. However, with the appearance of the 97th comment, the curve crossed zero, and the cumulative score changed from negative to positive, marking the sentiment reversal. After the sentiment reversal, the cumulative sentiment score increased rapidly and by a large margin, indicating that the sentiment reversal was relatively strong. The continuous accumulation of the positive sentiment score also reflects the process of the public sentiment gradually recovering from negative to positive.

[0197] These cases fully demonstrate that the user sentiment evolution learning module can accurately capture the patterns of sentiment changes in the comment area. Whether in the scenarios of sentiment consistency or sentiment reversal, it can effectively reflect the flow and evolution of sentiment in the comment area. Through the application of this module, we can gain a deeper understanding of the evolution process of user sentiment and its actual performance in social interactions.

[0198] Generally speaking, the present invention comprehensively utilizes comment data and studies the problem of fake news detection from the perspective of the comment area, and innovatively designs a fake news detection method to explore the evolution patterns of comment semantics and user sentiment. The experimental results show that the model has good performance on both Chinese and English datasets and can effectively improve the efficiency of fake news detection.

[0199] The embodiment of the present invention also provides a fake news detection device; the specific device may include:

[0200] An initial feature representation module, configured to obtain a target news and its comment set, and extract news text embedding, comment set text embedding, and comment set sentiment embedding;

[0201] A news encoding module, configured to perform context-related encoding representation on the news text embedding to obtain a news encoding representation;

[0202] A comment representation learning module, configured to perform news enhancement learning on the comment set text embedding according to the news text embedding to obtain a personalized representation of the comment set text;

[0203] A semantic sentiment fusion module, configured to fuse the personalized representation of the comment set text and the comment set sentiment embedding to obtain a comment semantic sentiment fusion vector representation;

[0204] A news classification module, configured to obtain semantic evolution features and user sentiment evolution features according to the personalized representation of the comment set text and the comment set sentiment embedding, and perform fake news classification prediction in combination with the news text embedding and the comment semantic sentiment fusion vector representation.

[0205] The fake news detection device in this embodiment is used to implement the aforementioned fake news detection method. Therefore, the specific implementation in the fake news detection device can be seen in the embodiment part of the previous fake news detection method. For example, the initial feature representation module, news encoding module, comment representation learning module, semantic sentiment fusion module, and news classification module are respectively used to implement steps S101, S102, S103, S104, and S105 in the above fake news detection method. Therefore, its specific implementation can refer to the descriptions of the corresponding individual part embodiments and will not be elaborated here.

[0206] A specific embodiment of the present invention also provides a fake news detection device, including: a memory for storing a computer program; a processor for implementing the steps of the above-mentioned fake news detection method when executing the computer program.

[0207] A specific embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned fake news detection method are implemented.

[0208] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented 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] The present application is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the specified functions in Figure 1 one or more of the processes or multiple processes and / or blocks Figure 1 one or more of the blocks or multiple blocks.

[0210] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including instruction means, and the instruction means implements the processes in Figure 1One process or multiple processes and / or boxes Figure 1 The functions specified in one box or multiple boxes.

[0211] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 One process or multiple processes and / or boxes Figure 1 The steps of the functions specified in one box or multiple boxes.

[0212] Obviously, the above embodiments are only examples given for clear illustration and are not limitations on the implementation manners. For those of ordinary skill in the art, other different forms of changes or variations can be made based on the above description. It is not necessary and impossible to list all the implementation manners here. And the obvious changes or variations derived therefrom are still within the protection scope of the present invention.

Claims

1. A fake news detection method, characterized in that, Including: Obtain the target news and its comment set, and extract the news text embedding, the comment set text embedding, and the comment set sentiment embedding; Perform context-related encoding representation on the news text embedding to obtain a news encoding representation; Perform news reinforcement learning on the comment set text embedding according to the news text embedding to obtain a personalized representation of the comment set text; Fuse the personalized representation of the comment set text and the comment set sentiment embedding to obtain a comment semantic sentiment fusion vector representation; According to the personalized representation of the comment set text and the comment set sentiment embedding, obtain semantic evolution features and user sentiment evolution features, and combine the news text embedding and the comment semantic sentiment fusion vector representation for fake news classification prediction.

2. The fake news detection method according to claim 1, wherein, The extraction of the news text embedding, the comment set text embedding, and the comment set sentiment embedding includes: Obtain the text embedding of each sentence in the target news through a first pre-trained model to obtain a news text embedding; Obtain the text embedding of each comment in the target news comment set through a second pre-trained model to obtain a comment set text embedding; Perform sentiment classification on each comment in the target news comment set to obtain a comment set sentiment embedding.

3. The false news detection method according to claim 1, wherein The performing context-related encoding representation on the news text embedding to obtain a news encoding representation includes: Use a bidirectional long short-term memory network to encode the news text embedding to obtain pre-order state sequence information and post-order state sequence information; Concatenate the hidden layer state of the last time step in the pre-order state sequence information and the hidden layer state of the last time step in the post-order state sequence information to obtain a news encoding representation.

4. The false news detection method according to claim 1, wherein The performing news reinforcement learning on the comment set text embedding according to the news text embedding to obtain a personalized representation of the comment set text includes: Based on the attention mechanism, perform news reinforcement learning on each comment text embedding according to the news text embedding to obtain a personalized representation of the comment set text.

5. The false news detection method according to claim 1, characterized in that, The fusing the personalized representation of the comment set text and the comment set sentiment embedding to obtain a comment semantic sentiment fusion vector representation includes: Concatenate the personalized representation and the sentiment embedding corresponding to each comment text and generate a gating vector; Perform a linear transformation on the personalized representation and the sentiment embedding corresponding to each comment text to map them to the same space, and generate a comment semantic vector and a sentiment vector with the same dimension; Generate a fused comment feature according to the gating vector, the comment semantic vector, and the sentiment vector; Input the fused comment feature sequence of the comment set into a long short-term memory network to generate state sequence information, and select the hidden layer state of the last time step in the state sequence information as the comment semantic sentiment fusion vector.

6. The false news detection method according to claim 1, wherein The obtaining semantic evolution features and user sentiment evolution features according to the personalized representation of the comment set text and the comment set sentiment embedding includes: Input the personalized representation of the comment set text and the comment set sentiment embedding into a long short-term memory network respectively to generate corresponding state sequence information, and select the hidden state layer of the last time step to generate semantic evolution features and user sentiment evolution features.

7. The false news detection method according to claim 1, wherein The performing fake news classification prediction includes: Concatenate the semantic evolution features, user sentiment evolution features, news text embedding, and comment semantic sentiment fusion vector, and calculate the correlation relationship between features through transformation by a fully connected network and a softmax function, and finally output a prediction vector.

8. A false news detection device, characterized in that, It includes: An initial feature representation module, configured to obtain a target news and its comment set, and extract news text embedding, comment set text embedding, and comment set sentiment embedding; A news encoding module, configured to perform context correlation encoding representation on the news text embedding to obtain a news encoding representation; A comment representation learning module, configured to perform news enhanced learning on the comment set text embedding according to the news text embedding to obtain a personalized representation of the comment set text; A semantic sentiment fusion module, configured to fuse the personalized representation of the comment set text and the comment set sentiment embedding to obtain a comment semantic sentiment fusion vector representation; A news classification module, configured to obtain semantic evolution features and user sentiment evolution features according to the personalized representation of the comment set text and the comment set sentiment embedding, and perform false news classification prediction by combining the news text embedding and the comment semantic sentiment fusion vector representation.

9. A false news detection device, characterized in that, It includes: A memory, configured to store a computer program; A processor, configured to implement the steps of a false news detection method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps of a false news detection method according to any one of claims 1 to 7 are implemented.

Images