False news detection method based on emotional interaction and graph attention network

Abstract

The invention discloses a false news detection method based on emotion interaction and a graph attention network, and relates to the technical field of computers. Through the design of an emotion scoring formula and an emotion classification formula, emotion classification of user emotions, namely news comments, under a social news scene is completed, and after feature extraction of news contents and news comments is completed, the news comments are extracted. The news emotion interaction graph is constructed by fully considering the news content and comments and the emotion interaction relation between different comments, attention weight calculation and weighted feature combination of nodes are carried out through a graph attention network, the news emotion interaction graph is learned, and a news prediction result is finally given.

Description

Technical Field

[0001] The present invention belongs to the field of computer technology, and in particular to a false news detection method based on emotional interaction and graph attention network. Background Art

[0002] The fake news detection task is generally a binary classification task, that is, classifying news as true or false. According to the different extracted features, the existing fake news detection methods can be mainly divided into two categories, namely, the fake news detection methods based on text content and the fake news detection methods based on social context.

[0003] Among them, the text content-based fake news detection method mainly extracts the content features of fake news, such as news titles, news texts, and news tags, so as to indirectly evaluate the editorial intentions of news writers, and detect fake news based on the analysis of the editorial intentions of news writers. The social context-based fake news detection method mainly analyzes users' views and positions on related news through the interactive behaviors of users on social news platforms, such as likes, comments, and reposts, to complete the fake news detection task.

[0004] A method for detecting fake news by mining dual emotions is proposed in the paper Zhang X, Cao J, Li X, et al. Mining dual emotion for fake news detection [C] / / Proceedings of the web conference 2021. 2021: 3465-3476. This method introduces the dual emotion concept of publisher emotion and social emotion into the study of fake news detection. The method first extracts the features of publisher emotion and social emotion, and completes the modeling of the emotion gap. The dual emotion features can be used to model the dual emotion signals in fake news detection by utilizing the similarities and differences of publisher emotion, social emotion, and dual emotion, and then the dual emotion features are integrated into the existing fake news detector for fake news detection.

[0005] However, the method of detecting fake news based on mining dual emotions in the paper Zhang X, Cao J, Li X, et al. Mining dual emotion for fake news detection [C] / / Proceedings of the web conference 2021. 2021: 3465-3476. only considers the impact of the emotion gap between the publisher's emotion and the user's emotion on the fake news detection task when mining the dual emotion features in the news and modeling the emotion gap, while ignoring the role of the emotion gap between different users in the same news instance in the fake news detection task. If fake news detection is only assisted by the emotion gap between the publisher's emotion and the user's emotion, then when the publisher's emotion has no obvious emotional tendency, it will not be practical to detect fake news by mining dual emotions. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a fake news detection method based on emotional interaction and graph attention network. Through the design of emotional scoring formula and emotional classification formula, the emotional classification of user emotions, that is, news comments, in the context of social news is completed. After completing the feature extraction of news content and news comments, the emotional interaction relationship between news content and comments as well as between different comments is fully considered to complete the construction of the news emotional interaction graph. The graph attention network is used to calculate the attention weights of the nodes and combine the weighted features, the news emotional interaction graph is learned, and finally the prediction results of the news are given.

[0007] The technical solution of the present invention is:

[0008] A fake news detection method based on emotional interaction and graph attention network includes the following steps:

[0009] Step 1: Obtain a news data set and divide it into a training set and a test set according to a set ratio; the news data set includes a number of samples, each of which includes news content, news comments and news labels;

[0010] Specifically: A news instance is represented by N = {A, R}, where A represents the news content, R represents the set of comments on news content A, and R = {r1, r2, r3, ..., r u ,…,r n}, where r u represents a comment of news content A, n represents the number of comments of news content A, and each news content A has a label y∈{0,1}, where y=1 means that the news content is true, and y=0 means that the news content is false;

[0011] Step 2: Build a fake news detection model based on sentiment interaction and graph attention network;

[0012] The fake news detection model based on emotional interaction and graph attention network includes a text feature extraction module, an emotional classification module, an emotional interaction graph construction module, a graph feature learning module and a news detection module;

[0013] The text feature extraction module is used to encode the news content and comments and convert them into embedded vector feature representations to obtain feature representations of the news content and the comments respectively;

[0014] The sentiment classification module is used to calculate the sentiment score of the comment, and classify the comment according to the sentiment score of the comment to determine the sentiment type of the comment, S = {s1, s2, s3, ..., s u ,...,s n} is the sentiment score set corresponding to the comment set R, s u Comment u The sentiment score, sentiment type set T = {t1, t2, t3, ..., t n}, t u For comments u Emotion type, where t u ∈{Negative, Neutral, Positive}, Negative means negative, Neutral means neutral, and Positive means support;

[0015] The emotional interaction graph construction module is used to model the emotional interaction between news content and comments and construct an emotional interaction graph;

[0016] The graph feature learning module is used to learn the feature representation of news content and comments corresponding to nodes in the sentiment interaction graph to obtain a graph representation of the sentiment interaction graph; the graph feature learning module is two graph attention network layers connected in sequence;

[0017] The news detection module is used to classify the graph representation of the emotional interaction graph generated in the graph feature learning module and output the detection result of the news;

[0018] Furthermore, the text feature extraction module is a pre-trained BERT model;

[0019] Furthermore, the coding process of news content is:

[0020]

[0021] in, is the feature representation of news content, d represents Dimension, EncodeText represents the process of text encoding;

[0022] Furthermore, the coding process of the comments is:

[0023]

[0024] in, is the feature representation of news comments and u=1,2…,n;

[0025] Furthermore, the process of determining the sentiment type of the comment is as follows:

[0026] For Chinese reviews, we use the natural language processing tool AipNLP as a sentiment analyzer:

[0027]

[0028] For English reviews, the sentiment analysis model VADER is used as the sentiment analyzer:

[0029]

[0030] in, is the result dictionary corresponding to Chinese comments, and is the result dictionary corresponding to English comments, and O u It indicates the probability that the text in the comment belongs to the positive category, and its value range is [0,1]. u It indicates the probability that the text in the comment belongs to the negative category, and its value range is [0,1]. u exist and where are the text sentiment polarity value and composite score in the comment, C u Indicates confidence level;

[0031] Comments u Sentiment scores u Calculated as follows:

[0032]

[0033] Among them, ω o Indicates O u The weight coefficient, ω p Indicates P u The weight coefficient of z represents the comprehensive sentiment weight coefficient. For Chinese comments, when O u ≥P u Time o =C u ,ω p =(1-C u ); when Ou <P u Time o =(1-C u ),ω p =C u ; For English comments, ω p =1-ω o ;

[0034] After calculating the sentiment scores of news comments, the sentiment of the comment text is classified as follows:

[0035]

[0036] In the formula, Analysis represents the sentiment classification operation, μ is the mean of the sentiment score set S, and σ is the standard deviation of the sentiment score set S, and k is the sentiment region parameter;

[0037] Furthermore, the sentiment interaction graph uses the feature representation of news content and the feature representation of comments as nodes of the sentiment interaction graph, which are news content nodes and comment nodes, respectively, and determines whether there is a connected edge between the news content node and the comment node and between two comment nodes according to the sentiment interaction relationship;

[0038] The method of determining whether there is a connected edge between a news content node and a comment node and between two comment nodes according to the emotional interaction relationship is as follows: when the emotional type of the comment is positive, there is an edge of type support between the comment node and the news content node; when the emotional type of the comment is neutral, the type of the edge is neutral; when the emotional type of the comment is negative, the type of the edge is opposition; for news comment nodes and news comment nodes, there is an edge of type support between the two news comment nodes only when the emotional types of the two comments belonging to the same news content are both positive;

[0039] The emotional interaction graph G=(V,E,T), where V={v0,v1,v2,...,v n} represents the set of nodes in the emotion interaction graph, vi represents the nodes in the emotion interaction graph and i = 0, 1, 2, ..., n, E = {e1, e2, e3, ..., e m} represents the edge set in the emotional interaction graph, e q represents the edge in the emotion interaction graph, and m is the number of edges in the emotion interaction graph; the nodes and edges in the emotion interaction graph have their own type mapping:

[0040] Type mapping for nodes:

[0041] φ(v):V→T V

[0042] Edge type mapping:

[0043]

[0044] Among them, φ(v):V represents the type mapping of the node, Represents the type mapping of the edge, T V = {news content, comments} represents the node type set, v∈T v Indicates the node type, T E = {support, neutral, oppose} represents the edge type set, e∈T E Indicates the edge type;

[0045] Furthermore, the process of obtaining the graph representation of the emotion interaction graph is as follows:

[0046] The input of the graph attention network layer is the feature matrix H of the node:

[0047]

[0048] in, Represents node v i The feature representation of

[0049] For each edge type e, using the k'th attention head in the graph attention network layer, calculate the node v i and neighbor node v j The attention coefficient e ij :

[0050]

[0051] In the formula, is the attention weight vector, || represents the vector concatenation operation, W is the weight matrix, Represents the neighbor node v j The feature representation is LeakyReLU as the activation function;

[0052] Complete the node v i and neighbor node v j After calculating the attention coefficient between them, all neighbor nodes are normalized to obtain the normalized attention coefficient α ij :

[0053]

[0054] Where N (i) Represents node v i The set of neighbor nodes, v k v i The neighbor node set of v is different from j Neighbor nodes, eik Represents node v i and neighbor node v k The attention coefficient between

[0055] Each attention head is normalized according to the attention coefficient α ij Update the feature representation of the node, and finally K attention heads obtain K sets of new feature representations;

[0056]

[0057] In the formula, is the new feature representation, σ′ is the activation function ReLU;

[0058] Then the feature representations of all attention heads are concatenated together to form the concatenated feature representation

[0059]

[0060] Then, the feature representation corresponding to all edge types of the sentiment interaction graph is Aggregate and get node v i The final feature representation

[0061]

[0062] Stack two graph attention network layers to learn graph features, and finally calculate the average value of the final feature representation of each node in the sentiment interaction graph Graph representation as a sentiment interaction graph:

[0063]

[0064] Furthermore, the process of outputting the detection result of the news is as follows:

[0065] The average value of the final feature representation of each node in the sentiment interaction graph Perform a linear transformation to obtain the unnormalized score vector z′:

[0066]

[0067] Where W f is the network learning parameter matrix, b f is the bias term;

[0068] The label is then predicted through the softmax layer When it is 0, the news is false. When it is 1, the news is true;

[0069]

[0070] Step 3: Use the training set to train the graph feature learning module in the fake news detection model based on emotional interaction and graph attention network to obtain a trained fake news detection model based on emotional interaction and graph attention network;

[0071] The model is trained by minimizing the cross entropy loss function:

[0072]

[0073] Where L(Θ) is the cross entropy loss function, Θ is the learning parameter, y is the true label of the news, The model predicts the probability that the news is true, Predict the probability that the news is false for the model;

[0074] Step 4: Input the news in the test set into the trained fake news detection model based on sentiment interaction and graph attention network to obtain the news detection results.

[0075] Compared with the prior art, the beneficial effects of the present invention are:

[0076] The present invention proposes a fake news detection method based on emotional interaction and graph attention network. The method designs a sentiment classification algorithm, which can accurately complete the sentiment classification of comment text in the context of social news, and fully considers the emotional interaction relationship between news content and comments as well as different comments according to the classification results, and constructs a news sentiment interaction graph. After learning the sentiment interaction graph through the graph attention network model, the final model is obtained. Compared with the existing model, this model significantly improves the accuracy of the fake news detection task. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is a flow chart of a false news detection method based on emotional interaction and graph attention network in an embodiment of the present invention;

[0078] Figure 2 This is a comparison chart of the ablation experiment results of the Weibo16 dataset in an embodiment of the present invention;

[0079] Figure 3 This is a comparison chart of the ablation experiment results of the Weibo20 dataset in an embodiment of the present invention;

[0080] Figure 4 This is a comparison chart of the ablation experiment results of the PolitiFact dataset in an embodiment of the present invention. DETAILED DESCRIPTION

[0081] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0082] A fake news detection method based on sentiment interaction and graph attention network, such as Figure 1 As shown, the following steps are included:

[0083] Step 1: Obtain a news data set and divide it into a training set and a test set according to a set ratio; the news data set includes a number of samples, each of which includes news content, news comments and news labels;

[0084] Specifically: A news instance is represented by N = {A, R}, where A represents the news content, R represents the set of comments on news content A, and R = {r1, r2, r3, ..., r u ,…,r n}, where r u represents a comment of news content A, n represents the number of comments of news content A; each news content A has a label y∈{0,1}, where y=1 means the news content is true, and y=0 means the news content is false;

[0085] In this implementation, Weibo16, Weibo20 and PolitiFact datasets are used;

[0086] Step 2: Build a fake news detection model based on emotional interaction and graph attention network (EIGN-FND);

[0087] The fake news detection model based on emotional interaction and graph attention network (EIGN-FND) includes a text feature extraction module, an emotional classification module, an emotional interaction graph construction module, a graph feature learning module and a news detection module;

[0088] The text feature extraction module is used to encode the news content and comments and convert them into embedded vector feature representations to obtain feature representations of the news content and the comments respectively;

[0089] The bidirectional encoder representation model based on transformer (BERT) is widely used as a text feature extractor in the task of false news detection. Therefore, the pre-trained BERT model is selected in the present invention to encode the news content and comments. Due to the differences between the word structures of different languages, the present invention uses the Chinese-Roberta-WWM-EXT model to encode the Chinese text data in the news content and comments, and uses the Bert-Base-Uncased model to encode the English text data in the news content and comments;

[0090] The coding process of news content is as follows:

[0091]

[0092] in, is the feature representation of news content, the value is the embedding vector of the [CLS] tag in the pre-trained BERT model, and d represents Dimension, EncodeText represents the process of text encoding;

[0093] The coding process of the comments is as follows:

[0094]

[0095] in, is the feature representation of news comments and u=1,2…,n, and the value is the embedding vector of the [CLS] tag in the pre-trained BERT model;

[0096] The sentiment classification module is used to calculate the sentiment score of the comment, and classify the comment according to the sentiment score of the comment to determine the sentiment type of the comment, S = {s1, s2, s3, ..., s u ,...,s n} is the sentiment score set corresponding to the comment set R, s u Comment u The sentiment score, sentiment type set T = {t1, t2, t3, ..., t n}, t u For comments u Emotion type, where t u ∈{Negative,Neutral,Positive};

[0097] Also, due to the differences in the language structure between Chinese text and English text, there are also differences when calculating sentiment scores.

[0098] For Chinese reviews, we use the natural language processing tool AipNLP as a sentiment analyzer:

[0099]

[0100] For English reviews, the sentiment analysis model VADER is used as the sentiment analyzer:

[0101]

[0102] in, is the result dictionary corresponding to Chinese comments, and is the result dictionary corresponding to English comments, and O u It indicates the probability that the text in the comment belongs to the positive category, and its value range is [0,1]. u It indicates the probability that the text in the comment belongs to the negative category, and its value range is [0,1]. u exist and where are the text sentiment polarity value and composite score in the comment, C u Indicates confidence level;

[0103] Comments u Sentiment scores u Calculated as follows:

[0104]

[0105] Among them, ω o Indicates O u The weight coefficient, ω p Indicates P u The weight coefficient of z represents the comprehensive sentiment weight coefficient, which is 0.95 in this implementation. For Chinese comments, when O u ≥P u Time o =C u ,ω p =(1-C u ); when O u <P u Time o =(1-C u ),ω p =C u ; For English comments, ω p =1-ω o ;

[0106] After calculating the sentiment scores of news comments, the sentiment of the comment text is classified as follows:

[0107]

[0108] In the formula, Analysis represents the sentiment classification operation, μ is the mean of the sentiment score set S, and σ is the standard deviation of the sentiment score set S, k is the sentiment region parameter, which is set to 1.1 in this implementation. Negative means negative, Neutral means neutral, and Positive means support;

[0109] The emotional interaction graph construction module is used to model the emotional interaction between news content and comments and construct an emotional interaction graph;

[0110] The present invention performs sentiment scoring and classification on comments, and then obtains the sentiment interaction relationship between news content and comments, and between comments. Then, the feature representation of news content and the feature representation of comments are used as nodes of the graph, and whether there are connected edges between news content nodes and comment nodes, and between two comment nodes, is determined based on the sentiment interaction relationship between texts. For example, when the sentiment type of a comment is positive, there is an edge of type "support" between the comment and the news content; when the sentiment type of the comment is neutral, the edge type is "neutral"; and when the sentiment type of the comment is negative, the edge type is "opposition". For news comments and comments, there is an edge of type "support" between the two comments only when the sentiment types of the two comments belonging to the same news are both positive, so as to avoid reducing the model performance due to the overly complex structure of the sentiment interaction graph.

[0111] Emotional interaction graph G = (V, E, T), where V = {v0, v1, v2, ..., v n} represents the set of nodes in the emotional interaction graph, v i represents a node in the emotional interaction graph and i = 0, 1, 2, ..., n, E = {e1, e2, e3, ..., e m} represents the edge set in the emotional interaction graph, e q represents the edge in the emotion interaction graph, and m is the number of edges in the emotion interaction graph; the nodes and edges in the emotion interaction graph have their own type mapping:

[0112] Type mapping for nodes:

[0113] φ(v):V→T V

[0114] Edge type mapping:

[0115]

[0116] Among them, φ(v):V represents the type mapping of the node, Represents the type mapping of the edge, T V = {news content, comments} represents the node type set, v∈T v Indicates the node type, T E = {support, neutral, oppose} represents the edge type set, e∈T E Indicates the edge type;

[0117] The graph feature learning module is used to learn the feature representation of news content and comments corresponding to nodes in the sentiment interaction graph to obtain a graph representation of the sentiment interaction graph; the graph feature learning module is two graph attention network layers connected in sequence;

[0118] The input of the graph attention network layer is the feature matrix H of the node:

[0119]

[0120] in, Represents node v i The feature representation of

[0121] For each edge type e, using the k'th attention head in the graph attention network layer, calculate the node v i and neighbor node v j The attention coefficient e ij :

[0122]

[0123] In the formula, is the attention weight vector, || represents the vector concatenation operation, W is the weight matrix, Represents the neighbor node v j The feature representation is LeakyReLU as the activation function;

[0124] Complete the node v i and neighbor node v j After calculating the attention coefficient between them, all neighbor nodes are normalized to obtain the normalized attention coefficient α ij :

[0125]

[0126] Where N (i) Represents node v i The set of neighbor nodes, v k v i The neighbor node set of v is different from j Neighbor nodes, e ik Represents node v i and neighbor node v k The attention coefficient between

[0127] Each attention head is normalized according to the attention coefficient α ij Update the feature representation of the node, and finally K attention heads obtain K sets of new feature representations;

[0128]

[0129] In the formula, is the new feature representation, σ′ is the activation function ReLU;

[0130] Then the feature representations of all attention heads are concatenated together to form the concatenated feature representation

[0131]

[0132] Then, the feature representation corresponding to all edge types of the sentiment interaction graph is Aggregate and get node v i The final feature representation

[0133]

[0134] Stack two graph attention network layers to learn graph features, and finally calculate the average value of the final feature representation of each node in the sentiment interaction graph Graph representation as a sentiment interaction graph:

[0135]

[0136] The news detection module is used to classify the graph representation of the emotional interaction graph generated in the graph feature learning module and output the detection result of the news;

[0137] The average value of the final feature representation of each node in the sentiment interaction graph Perform a linear transformation to obtain the unnormalized score vector z′:

[0138]

[0139] Where W f is the network learning parameter matrix, b f is the bias term;

[0140] The label is then predicted through the softmax layer When it is 0, the news is false. When it is 1, the news is true;

[0141]

[0142] Step 3: Use the training set to train the graph feature learning module in the fake news detection model based on emotional interaction and graph attention network (EIGN-FND) to obtain the trained fake news detection model based on emotional interaction and graph attention network (EIGN-FND);

[0143] The model is trained by minimizing the cross entropy loss function:

[0144]

[0145] Where L(Θ) is the cross entropy loss function, Θ is the learning parameter, y is the true label of the news, The model predicts the probability that the news is true, Predict the probability that the news is false for the model;

[0146] Step 4: Input the news in the test set into the trained fake news detection model based on emotional interaction and graph attention network (EIGN-FND) to obtain the news detection results;

[0147] The present invention conducts experiments on three datasets, including two Chinese datasets and one English dataset.

[0148] (1) Chinese dataset:

[0149] 1) Weibo16: After being proposed, Weibo16 has become a benchmark dataset for Chinese fake news detection. Since repetition may increase the risk of data leakage and make the model tend to learn some event-specific features, the version without duplicate news provided by Zhang et al. is used in the experiment.

[0150] 2) Weibo20: This dataset was collected by Zhang et al. from the Sina Weibo Community Management Center. The real news collected was verified by NewsVerify7. There is a very small amount of text data in the Weibo20 dataset that cannot be encoded. Considering that the proportion of samples with encoding errors in the dataset is extremely small, these news data were manually cleared from the dataset.

[0151] (2) English dataset: The data was collected from the fact-checking platform PolitiFact. The dataset contains labeled news content and news comments. Since the original dataset contained some noise data during the crawling process, the noise data was removed in the present invention. At the same time, since the distribution of the two categories of news in the dataset is uneven, the data was oversampled at a ratio of 1:1.3 during training.

[0152] Since the standard deviation of comment sentiment scores needs to be calculated in the sentiment classification module, news with less than 2 comments in the dataset are filtered out. The detailed statistical results of the dataset are shown in Table 1.

[0153] Table 1 Experimental data set statistics

[0154] Dataset Weibo16 Weibo20 PolitiFact Total news 3700 6164 361 Number of real news 2345 3183 234 Number of fake news 1355 2981 127 Total number of comments 1874672 1982772 82801 Average number of reviews 506 321 229

[0155] Parameter settings:

[0156] The present invention uses accuracy, precision, recall and F1 score as model evaluation indicators to calculate the performance of the news prediction model. In the experiment, 75% of the data is taken as the training set and 25% of the data is taken as the test set. The learning rate of the model on the Chinese data set is 1e-3, and the learning rate on the English data set is 2e-5. The upper limit of the number of comments for each news is set to 200, the text feature dimension is 768, the output dimension of the graph attention network is 256, the number of attention heads is 4, and the number of model training rounds is 200.

[0157] Compare models:

[0158] In order to verify the effect of the model proposed in the present invention, the model proposed in the present invention is compared with multiple fake news detection models. The selected comparison models are as follows:

[0159] LSTM: This model takes a sequence of news words as input and obtains the final news text representation through network learning to detect fake news.

[0160] ML-GRU: This model uses a multi-layer general GRU network to model news events as time series of variable length, which is very effective for early rumor detection.

[0161] TextCNN: This model uses convolutional neural networks for sentence-level classification tasks and can complete multiple text classification tasks.

[0162] HSA-BLSTM: This method proposes a hierarchical neural network that combines social information for rumor detection. First, a hierarchical bidirectional long short-term memory model is constructed for representation learning. Then, the social context is integrated into the network through the attention mechanism.

[0163] Dual Emotion+HSA-BLSTM: This method uses dual emotion features to model dual emotion signals in fake news detection by leveraging the similarities and differences of publisher emotions, social emotions, and dual emotions. The dual emotion features are then integrated into the HSA-BLSTM model for fake news detection.

[0164] When conducting comparative experiments, the original parameters of the comparative model are used.

[0165] Performance comparison and analysis:

[0166] In the experiment, the model was trained and tested 5 times to ensure the fairness of the test results. The detailed comparison results of the performance of the present invention and other comparison models on three data sets are shown in Table 2, Table 3, and Table 4 respectively.

[0167] Table 2 Experimental results of the model on the Weibo16 dataset

[0168]

[0169]

[0170] Table 3 Experimental results of the model on the Weibo20 dataset

[0171]

[0172] Table 4 Experimental results of the model on the PolitiFact dataset

[0173]

[0174] From the data in the table, we can see that among the fake news detection methods based on text content, the TextCNN model performs best, probably because the TextCNN model can learn text representation better than the other two methods based on text content. Among the fake news detection methods based on social context, the Dual Emotion+HSA-BLSTM model performs better than the original HSA-BLSTM model because it mines the dual emotional features in the news for modeling. In general, the fake news detection methods based on social context are generally better than the fake news detection methods based on text content in terms of data performance because they can consider the comprehensive features of news from multiple aspects.

[0175] The EIGN-FND model proposed in the present invention outperforms other comparison models in classification accuracy on the three data sets. Compared with the model with the best overall performance among the comparison models, the F1 score of the EIGN-FND model is improved by 2.2% and 2.5% on the Chinese data set, and by 1.4% on the English data set. This shows that the fake news detection method based on emotional interaction proposed in the present invention can effectively improve the performance of the fake news detection task. The EIGN-FND model can assist in fake news detection by learning the emotional interaction relationship between news content and comments and between different comments, further improving the accuracy of fake news detection. The experimental results on the three data sets show that the model proposed in the present invention performs well in the fake news detection task.

[0176] Ablation experiment and analysis:

[0177] In order to verify the effectiveness of the fusion components of the present invention, independent ablation analysis was performed to evaluate the effectiveness of each component in the EIGN-FND model, such as Figure 2 , Figure 3 and Figure 4 shown.

[0178] (1) Effectiveness of the sentiment classification module

[0179] In order to verify the effectiveness of the sentiment classification module, the present invention named the sub-model without the sentiment classification module EIGN-FND-t. After removing the sentiment classification module, the sentiment classification of news comments was performed in a random manner. The experimental results are as follows: Figure 1 As shown in the figure, the performance of the EIGN-FND-t model on the three data sets has declined to varying degrees. Since the random sentiment classification of news comments still leads to a portion of news comments being correctly classified, the performance of the EIGN-FND-t model is not significantly reduced compared to the EIGN-FND model, and it still has a high accuracy for the false news detection task. Therefore, it is believed that the sentiment classification module of the present invention can improve the accuracy of the false news detection task.

[0180] (2) Effectiveness of the network model

[0181] In order to verify the effectiveness of the graph attention network model used in the graph feature learning module, the network model in the graph feature learning module is replaced with a graph convolutional network, and the replaced sub-model is named EIGN-FND-m. The experimental results are shown in Figure 2. Figure 1 As shown in the figure, since the replaced sub-model does not assign different weights to each neighbor node through the attention mechanism and cannot flexibly capture the complex relationships in the graph, the performance of the EIGN-FND-m model on the three datasets is significantly lower than that of the EIGN-FND model. The F1 scores on the Chinese dataset dropped by 6.8% and 6.3% respectively, and the F1 score on the English dataset dropped by 5.7%. Therefore, the graph attention network is of great significance to the graph feature learning module in the present invention, and can significantly improve the performance of the fake news detection model.

Claims

1. A fake news detection method based on emotional interaction and graph attention network, characterized in that: The steps include: Step 1: Obtain a news data set and divide it into a training set and a test set according to a set ratio; the news data set includes a number of samples, each of which includes news content, news comments and news labels; Step 2: Build a fake news detection model based on sentiment interaction and graph attention network; Step 3: Use the training set to train the graph feature learning module in the fake news detection model based on emotional interaction and graph attention network to obtain a trained fake news detection model based on emotional interaction and graph attention network; Step 4: Input the news in the test set into the trained fake news detection model based on sentiment interaction and graph attention network to obtain the news detection results.

2. According to claim 1, a false news detection method based on emotional interaction and graph attention network is characterized in that: In step 1, a news instance is represented by N = {A, R}, where A represents the news content, R represents the set of comments on news content A, and R = {r1, r2, r3, ..., r u ,…,r n }, where r u represents a comment on news content A, n represents the number of comments on news content A, and each news content A has a label y∈{0,1}, where y=1 indicates that the news content is true, and y=0 indicates that the news content is false.

3. According to claim 1, a false news detection method based on emotional interaction and graph attention network is characterized in that: The fake news detection model based on emotional interaction and graph attention network described in step 2 includes a text feature extraction module, an emotional classification module, an emotional interaction graph construction module, a graph feature learning module and a news detection module; The text feature extraction module is used to encode the news content and comments and convert them into embedded vector feature representations to obtain feature representations of the news content and the comments respectively; The sentiment classification module is used to calculate the sentiment score of the comment, and classify the comment according to the sentiment score of the comment to determine the sentiment type of the comment, S = {s1, s2, s3, ..., s u ,...,s n } is the sentiment score set corresponding to the comment set R, s u Comment u The sentiment score, sentiment type set T = {t1, t2, t3, ..., t n }, t u For comments u Emotion type, where t u ∈{Negative, Neutral, Positive}, Negative means negative, Neutral means neutral, and Positive means support; The emotional interaction graph construction module is used to model the emotional interaction between news content and comments and construct an emotional interaction graph; The graph feature learning module is used to learn the feature representation of news content and comments corresponding to nodes in the sentiment interaction graph to obtain a graph representation of the sentiment interaction graph; the graph feature learning module is two graph attention network layers connected in sequence; The news detection module is used to classify the graph representation of the emotional interaction graph generated in the graph feature learning module and output the detection result of the news; The text feature extraction module is a pre-trained BERT model.

4. According to claim 3, a false news detection method based on emotional interaction and graph attention network is characterized in that: The coding process of news content is as follows: in, is the feature representation of news content, d represents Dimension, EncodeText represents the process of text encoding; The coding process of the comments is: in, is the feature representation of news comments and u=1,2…,n.

5. According to claim 3, a false news detection method based on emotional interaction and graph attention network is characterized in that: The process of determining the sentiment type of a comment is as follows: For Chinese reviews, we use the natural language processing tool AipNLP as a sentiment analyzer: For English reviews, the sentiment analysis model VADER is used as the sentiment analyzer: in, is the result dictionary corresponding to Chinese comments, and is the result dictionary corresponding to English comments, and O u It indicates the probability that the text in the comment belongs to the positive category, and its value range is [0,1]. u It indicates the probability that the text in the comment belongs to the negative category, and its value range is [0,1]. u exist and where are the text sentiment polarity value and composite score in the comment, C u Indicates confidence level; Comments u Sentiment scores u Calculated as follows: Among them, ω o Indicates O u The weight coefficient, ω p Indicates P u The weight coefficient of z represents the comprehensive sentiment weight coefficient. For Chinese comments, when O u ≥P u Time o =C u ,ω p =(1-C u ); when O u <P u Time o =(1-C u ),ω p =C u ; For English comments, ω p =1-ω o ; After calculating the sentiment scores of news comments, the sentiment of the comment text is classified as follows: In the formula, Analysis represents the sentiment classification operation, μ is the mean of the sentiment score set S, and σ is the standard deviation of the sentiment score set S, and k is the sentiment region parameter.

6. According to claim 3, a false news detection method based on emotional interaction and graph attention network is characterized in that: The emotional interaction graph uses the feature representation of news content and the feature representation of comments as nodes of the emotional interaction graph, which are news content nodes and comment nodes respectively, and determines whether there is a connected edge between the news content node and the comment node and between two comment nodes according to the emotional interaction relationship; The method of determining whether there are connected edges between a news content node and a comment node and between two comment nodes based on the emotional interaction relationship is specifically as follows: when the emotional type of the comment is positive, there is an edge of type support between the comment node and the news content node; when the emotional type of the comment is neutral, the type of the edge is neutral; when the emotional type of the comment is negative, the type of the edge is opposition; for between news comment nodes and news comment nodes, there is an edge of type support between the two news comment nodes only when the emotional types of the two comments belonging to the same news content are both positive.

7. The method for detecting fake news based on emotional interaction and graph attention network according to claim 6 is characterized in that: The emotional interaction graph G=(V,E,T), where V={v0,v1,v2,...,v n } represents the set of nodes in the emotional interaction graph, v i represents a node in the emotional interaction graph and i = 0, 1, 2, ..., n, E = {e1, e2, e3, ..., e m } represents the edge set in the emotional interaction graph, e q represents the edge in the emotion interaction graph, and m is the number of edges in the emotion interaction graph; the nodes and edges in the emotion interaction graph have their own type mapping: Type mapping of nodes: φ(v):V→T V Edge type mapping: Among them, φ(v):V represents the type mapping of the node, Represents the type mapping of the edge, T V = {news content, comments} represents the node type set, v∈T v Indicates the node type, T E = {support, neutral, oppose} represents the edge type set, e∈T E Indicates the edge type.

8. According to claim 3, a false news detection method based on emotional interaction and graph attention network is characterized in that: The process of obtaining the graph representation of the emotion interaction graph is as follows: The input of the graph attention network layer is the feature matrix H of the node: in, Represents node v i The feature representation of For each edge type e, using the k'th attention head in the graph attention network layer, calculate the node v i and neighbor node v j The attention coefficient e ij : In the formula, is the attention weight vector, || represents the vector concatenation operation, W is the weight matrix, Represents the neighbor node v j The feature representation is LeakyReLU as the activation function; Complete the node v i and neighbor node v j After calculating the attention coefficient between them, all neighbor nodes are normalized to obtain the normalized attention coefficient α ij : Where N (i) Represents node v i The set of neighbor nodes, v k v i The neighbor node set of v is different from j Neighbor nodes, e ik Represents node v i and neighbor node v k The attention coefficient between Each attention head is normalized according to the attention coefficient α ij Update the feature representation of the node, and finally K attention heads obtain K sets of new feature representations; In the formula, is the new feature representation, σ′ is the activation function ReLU; Then the feature representations of all attention heads are concatenated together to form the concatenated feature representation Then, the feature representation corresponding to all edge types of the sentiment interaction graph is Aggregate and get node v i The final feature representation Stack two graph attention network layers to learn graph features, and finally calculate the average value of the final feature representation of each node in the sentiment interaction graph Graph representation as a sentiment interaction graph: 。 9. According to claim 3, a false news detection method based on emotional interaction and graph attention network is characterized in that: The process of outputting the detection result of the news is as follows: The unnormalized score vector z′ is obtained by linear transformation: Where W f is the network learning parameter matrix, b f is the bias term; Then, the average value of the final feature representation of each node in the sentiment interaction graph is Perform softmax layer prediction label When it is 0, the news is false. When it is 1, the news is true; 。 10. The method for detecting fake news based on emotional interaction and graph attention network according to claim 1, characterized in that: In step 3, the model is trained by minimizing the cross entropy loss function: Where L(Θ) is the cross entropy loss function, Θ is the learning parameter, y is the true label of the news, The model predicts the probability that the news is true, Predict the probability that the news is false for the model.

Images