Rumor detection method and system based on timing sentiment features

Abstract

A rumor detection method based on time sequence emotion features, comprising: processing data obtained from a social network to obtain source posts, reply posts and propagation paths of an event; performing text feature extraction on the texts of the source and the reply to convert them into vector form; performing emotion feature extraction on the texts of the source and the reply to obtain emotion scores of the texts; performing time sequence feature extraction on the texts of the posts to obtain one-hot time sequence vectors; multiplying the extracted emotion features and the extracted time sequence features to obtain emotion time sequence feature vectors of the posts; performing vector splicing on the emotion time sequence feature vectors and the text feature vectors to obtain post representation vectors; inputting the post representation vectors and the propagation paths of the posts into an RvNN neural network to obtain a comprehensive representation of the event; inputting the obtained comprehensive representation of the event into a neural network in a rumor classification module to judge the event and output a rumor classification result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rumor detection, in particular to a rumor detection method and system based on time sequence emotional features. BACKGROUND

[0002] The development of social media has changed the way the public consumes information. However, a large amount of information has also brought many rumors. Rumors are defined as unverified information, and most rumors have adverse effects. At the same time, studies have shown that ordinary people have difficulty identifying whether an event is a rumor. Traditional rumor identification requires trained rumor discriminators to identify events, but with the development of social networks, the number of events has exploded, and traditional human-based rumor detection methods have been in a dilemma. Therefore, it is necessary to develop an effective rumor detection method.

[0003] Current rumor detection methods focus on text features. Recently, with the development of social media, more and more features have been investigated and integrated into the complex representation of rumors, such as visual features. However, malicious users can still easily fake text or images through image generation or text generation techniques. In addition, compared with the text and visual features of social media, the propagation feature is difficult for malicious users to manipulate. Therefore, more and more multi-modal detection models integrate the propagation feature into their methods. However, on events with relatively inconspicuous propagation features and text features, traditional multi-modal methods have difficulty identifying whether the event is fake news. At the same time, during the propagation of an event, the emotional features of the posts will usually change, but the current multi-modal rumor detection method does not fully utilize the time sequence emotional features of the rumor. SUMMARY

[0004] The present application overcomes the problem that the existing rumor detection algorithm has low performance when the text and propagation structure information are missing, and proposes to use time sequence emotions as features for rumor detection, and use time sequence emotional features as information supplements.

[0005] The technical concept of the present application is that emotion, as a representation of people's psychological state, is also attracting more and more attention from researchers. Most existing emotion-based rumor detection methods use emotion as a supplementary attribute of text features, but ignore the emotional changes during information propagation. Therefore, the present application proposes a rumor detection system using time sequence emotional features, which uses time sequence emotional features as event information supplements to effectively improve the generalization performance of the model.

[0006] The technical scheme adopted by the present application to achieve the above application purpose is as follows:

[0007] A rumor detection method based on time sequence emotional features, comprising the following steps:

[0008] S1: data extraction and format conversion on data obtained from social network, obtaining source post, reply post and propagation path of event;

[0009] S2: text feature extraction on source and reply text processed by S1, converting to vector form;

[0010] S3: sentiment feature extraction on source and reply text processed by S1, converting text sentiment feature to sentiment score;

[0011] S4: time sequence feature extraction on text of post processed by S1, converting post sequence to one-hot time sequence vector;

[0012] S5, multiplying the sentiment feature extracted by S3 and the time sequence feature extracted by S4 to obtain the sentiment time sequence feature vector of the post;

[0013] S6, vector splicing the sentiment time sequence feature vector extracted by S5 and the text feature vector extracted by S2 to obtain the representation vector of the post;

[0014] S7, inputting the post representation vector processed by S6 and the propagation path of the post into RvNN neural network to obtain the comprehensive representation of the event;

[0015] S8, inputting the event comprehensive representation obtained by S7 into the neural network in the rumor classification module to judge the event and output the rumor classification result.

[0016] Preferably, the step S1 specifically comprises:

[0017] defining the training data as wherein represents the number of training data, C i is one of the events. The event is represented as a tree structure C i = <V i ,E i >, wherein is a node set, each element in the node set is the source post is the total number of posts is k i . The forwarding relationship between posts is defined as wherein represents is the reply of .

[0018] Preferably, the step S2 specifically comprises:

[0019] using a pre-trained Word2Vec model to convert the text in the post into a word vector [e1, e2,..., en].l ] and convert the word vectors into the text feature vector of the post using the following formula:

[0020]

[0021] Preferably, the step S3 specifically comprises:

[0022] The text sentiment feature is extracted using a general sentiment analysis tool. The general sentiment analysis tool can extract the sentiment feature in the post. Finally, the sentiment score of the post is obtained

[0023] Preferably, the step S4 specifically comprises:

[0024] The post sequence is compressed and encoded, and a fixed-dimensional vector is used to represent the position of the post in all posts of the current event, as follows:

[0025]

[0026] Where [ ] represents the floor function, I (l=k) represents the function value of 1 if l=k, otherwise 0, j is the event index, d sen is the position vector dimension, k i is the total number of posts in the current event.

[0027] Preferably, the step S5 specifically comprises:

[0028] The features obtained in step S3 and the features obtained in step S4 are fused using the following formula to obtain the timing sentiment feature

[0029]

[0030] Where is the post sentiment feature obtained in step S3, is the post timing feature obtained in step S4, and the final output is the timing sentiment feature of the post.

[0031] Preferably, the step S6 specifically comprises:

[0032] The features extracted in S5 and S2 are fused according to the following formula to obtain the comprehensive representation of the post:

[0033]

[0034] Where is the text feature obtained in step S2, is the timing sentiment feature obtained in step S5.

[0035] Preferably, the step S7 specifically comprises:

[0036] S7.1: Capture the propagation features of the post integrated representation using RvNN, the formula is as follows:

[0037]

[0038] Wherein represents the parent node of the node , GRU(*) represents the GUR hidden unit, and the formula is as follows:

[0039]

[0040]

[0041]

[0042]

[0043]

[0044] In the formula, E is a transformation matrix, {W * ,U *} is a weight matrix, σ(*) and tanh(*) represent activation functions, x t represents the input h t represents the output hidden state, h t-1 represents the hidden state of the previous node. The hidden state of all leaf nodes can be obtained

[0045] S7.2: After obtaining the hidden state of all leaf nodes, use the maximum pooling layer to obtain the integrated representation of the event, which is represented as follows:

[0046]

[0047] Preferably, the step S8 specifically comprises:

[0048] Using MLP as a classifier, classify according to the propagation features of the post integrated representation, and the calculation formula is represented as follows:

[0049]

[0050] Wherein {W * ,b *} are the weights and biases of the neural network, Relu(*) and σ(*) are activation functions, and the final output is the prediction of the model on the event C i .

[0051] A rumor detection method based on time sequence emotional features is implemented, comprising a post representation module, an event representation module and a rumor classification module.

[0052] The post representation module represents social network posts using text features and time sequence emotional features.

[0053] The event representation module uses RvNN to capture event post representation features according to news propagation paths to obtain comprehensive event representation.

[0054] The rumor classification module judges whether an event is a rumor according to the comprehensive event representation.

[0055] The post representation module, the event representation module and the rumor classification module are sequentially connected.

[0056] The present application has the following advantages:

[0057] (1) The rumor detection method based on time sequence emotional features can use time sequence emotional features in event propagation processes, and can significantly improve recognition accuracy.

[0058] (2) The rumor detection method based on time sequence emotional features is superior to existing methods. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 The structural diagram of the present application. DETAILED DESCRIPTION

[0060] The specific embodiments of the present application will be further described in detail below with reference to the accompanying drawings.

[0061] Referring to Figure 1 The rumor detection method based on time sequence emotional features comprises the following steps:

[0062] S1: Data extraction and format conversion are performed on data obtained from a social network to obtain source posts, reply posts and propagation paths of events, specifically comprising:

[0063] The data obtained from the original social network platform API is preprocessed, and "forward microblog" and "forward" are removed. The original data is repackaged according to the format of post belonging to an event, post id, post text, post time, reply id, wherein the post text is text information after text preprocessing.

[0064] S2: Text features are extracted from the source and reply texts processed in S1, and are converted into vector form, specifically comprising:

[0065] S2.1 Transform the words in the text into vectors using the Word2Vec pre-training model. Train the text words using Skip-gram in Word2Vec to obtain a pre-training word vector model, and the embedding dimension of the word vector is 200. Then use the pre-training word vector model to obtain the word vector [e1, e2,..., en] of the post. l ]。

[0066] S2.2 Convert the word vector into a text representation of the post using the average pooling layer, as shown in the following process:

[0067]

[0068] S3: Extract the sentiment features of the source and reply text processed in S1, and convert the text sentiment features into sentiment scores, including:

[0069] Use the NLTK public API to analyze the sentiment of the text and obtain the text sentiment features The result ranges from (-1, 1), and the closer the result is to 1, the more positive the text sentiment is.

[0070] S4: Extract the timing features of the text of the post processed in S1, and convert the post sequence into a one-hot timing vector, including:

[0071] Use the post information in the dataset to obtain the post sequence information of the same event, and then obtain the post timing features according to the formula, as shown below:

[0072]

[0073] Where [ ] represents rounding down, I(l=k) represents the function value of 1 if l=k, and 0 otherwise, j is the event index, d sen is the position vector dimension, which is 100, and k i is the total number of posts in the current event.

[0074] S5, multiply the sentiment features extracted in S3 and the timing features extracted in S4 to obtain the sentiment timing feature vector of the post, including:

[0075] Multiply the text sentiment features obtained in S3 and the post timing features obtained in S4 to obtain the timing sentiment features of the post, i.e.

[0076]

[0077] S6, concatenate the sentiment timing feature vector extracted in S5 and the text feature vector extracted in S2 to obtain the representation vector of the post, including:

[0078] The vector splicing of the emotion time sequence feature vector extracted by S5 and the text feature vector extracted by S2 is performed according to the following formula:

[0079]

[0080] The comprehensive representation of the post is obtained.

[0081] S7, inputting the post representation vector obtained by S6 processing and the propagation path of the post into the RvNN neural network to obtain the comprehensive representation of the event, specifically comprising:

[0082] S7.1 uses RvNN to capture the propagation features of the post comprehensive representation, and the formula is as follows:

[0083]

[0084] Wherein represents the parent node of the node , GRU(*) represents the GUR hidden unit, and the formula is as follows:

[0085]

[0086]

[0087]

[0088]

[0089]

[0090] In the formula, E is a transformation matrix, {W * ,U *} is a weight matrix, sigma(*) and tanh(*) represent activation functions, x t represents the input h t represents the output hidden state, h t-1 represents the hidden state of the previous node. The hidden states of all leaf nodes

[0091] S7.2, after obtaining the hidden states of all leaf nodes, uses the maximum pooling layer to obtain the comprehensive representation of the event, and the representation is as follows:

[0092]

[0093] Wherein is the comprehensive representation of the event, and Maxpooling(*) represents the maximum pooling operation.

[0094] S8, input the event comprehensive representation obtained in S7 into the neural network in the rumor classification module, judge the event, and output a rumor classification result, specifically comprising:

[0095] Using MLP as a classifier, the propagation features of the post comprehensive representation are classified, and the calculation formula is as follows:

[0096]

[0097] Where {W * ,b *} are the weights and biases of the neural network, Relu(*) and sigma(*) are activation functions, and the final output is the prediction of the model on the event C i .

[0098] In order to prove that the rumor detection system and method based on time sequence emotion proposed in the application have relatively improved effect compared with other existing models, we test the effect of the method on the microblog rumor data set, and use the cross-entropy loss function and the ADAM optimization method to train the model. From the effect of the experiment, it can be seen that the performance obtained by the method proposed in the application is better than that of the baseline method. Among them, BERT represents using BERT to extract text features of the event; BiGRU represents using BiGRU to extract text features; Emo-BiGRU represents using BiGRU to extract text features, and using emotion features as enhancement; RNN represents using RNN to extract text sequence features of the event; RvNN represents using RvNN to extract text propagation features; BiGCN represents using BiGCN to extract propagation features of the event. In addition, Accuracy (correct rate), macro-F1 (macro average F-Score), Precision (precision), Recall (recall) and F1 (F-Score) are used as evaluation indexes.

[0099] Table 1 compares with baseline methods on microblog data set.

[0100]

[0101] As can be seen from Table 1, the model proposed in the application improves the macro average F1 value of all events to 0.939, and the correct rate is also improved to 0.939, which is higher than all baseline methods.

[0102] In order to verify the effectiveness of the features in the rumor detection system and method based on time sequence emotion proposed in the application, this section uses the control model variable method to remove the text representation features and the emotion time sequence features in the news representation module, and the comparison results are shown in the following table.

[0103] Table 2. Performance of the model with different features removed on the microblog dataset.

[0104]

[0105] where w / o Text means removing the text feature in the news representation module, and w / o Sen means removing the time-seniority feature in the news representation module. When removing a feature in a news representation module, the performance of the model proposed in the embodiments will decrease.

[0106] The content of the embodiments of the present specification is merely a list of implementation forms of the inventive concept, and the protection scope of the present application should not be regarded as being limited to the specific forms stated in the embodiments, and the protection scope of the present application also extends to equivalent technical means that can be thought of by those skilled in the art according to the inventive concept.

Claims

1. A rumor detection method based on temporal sentiment features, characterized in that: Includes the following steps: S1: Extract and convert the data obtained from social networks to obtain the source post, reply post, and propagation path of the event; S2: Extract text features from the source and response texts processed by S1 and convert them into vector form; S3: Extract sentiment features from the source and response texts processed in S1, and convert the text sentiment features into sentiment scores; S4: Extract temporal features from the text of the posts processed in S1 and convert the post sequence into a one-hot temporal vector; S5. Multiply the sentiment features extracted in S3 with the temporal features extracted in S4 to obtain the sentiment temporal feature vector of the post. S6. Concatenate the sentiment time-series feature vector extracted in S5 with the text feature vector extracted in S2 to obtain the post representation vector. S7. Input the post representation vector and post propagation path obtained from S6 into the RvNN neural network to obtain a comprehensive representation of the event. S8. Input the comprehensive representation of the events obtained in S7 into the neural network of the rumor classification module, judge the events, and output the rumor classification result.

2. The rumor detection method based on temporal sentiment features as described in claim 1, characterized in that: Step S1 specifically includes: The training data is defined as C = {C1, C2, ..., C}. |C| }, where |C| represents the number of training data, C i One of the events; the event is represented as a tree structure C. i = <V i E i >, among which Let be a set of nodes, and each element in the set of nodes is... The source post is The total number of posts is k i Define the forwarding relationship between posts as in express for The reply.

3. The rumor detection method based on temporal sentiment features as described in claim 1, characterized in that: Step S2 specifically includes: Use a pre-trained Word2CVec model to process posts The text in the document is converted into word vectors [e1, e2, ..., e]. l The word vectors are then converted into text feature vectors for the posts using the following formula:

4. The rumor detection method based on temporal sentiment features as described in claim 1, characterized in that: Step S3 specifically includes: A general sentiment analysis tool is used to extract sentiment features from the text; this tool can extract sentiment features from posts, ultimately yielding a sentiment score for the post.

5. The rumor detection method based on temporal sentiment features as described in claim 1, characterized in that: Step S4 specifically includes: The post sequence is compressed and encoded, and a fixed-dimensional vector is used to represent the position of each post among all posts in the current event, as shown in the following formula: Where [*] denotes rounding down, I(l=k) means the function value is 1 if l=k and 0 otherwise, j is the event index, and d sen k is the dimension of the position vector. i This represents the total number of posts related to the current event.

6. The rumor detection method based on temporal sentiment features as described in claim 1, characterized in that: Step S5 specifically includes: The features obtained in steps S3 and S4 are fused using the following formula to obtain the temporal sentiment features. in The sentiment features of the posts obtained in step S3, Step S4 obtains the temporal features of the posts, and the final output is the temporal sentiment features of the posts.

7. The rumor detection method based on temporal sentiment features as described in claim 1, characterized in that: Step S6 specifically includes: The features extracted from S5 and S2 are fused according to the following formula to obtain a comprehensive representation of the post: in The text features obtained in step S2, The temporal sentiment features obtained in step S5.

8. The rumor detection method based on temporal sentiment features as described in claim 1, characterized in that: Step S7 specifically includes: S7.1: Use RvNN to capture the propagation features of the post comprehensive representation, as shown in the following formula: in Represents a node The parent node, GRU(*), represents the GUR hidden unit, as shown in the following formula: In the formula, E is the transformation matrix, {W * U * } represents the weight matrix, σ(*) and tanh(*) represent the activation functions, and x t Indicates input h t Indicates the hidden state of the output, h t-1 This indicates the hidden state of the previous node; it allows you to obtain the hidden states of all leaf nodes. S7.2: After obtaining the hidden states of all leaf nodes, use the max pooling layer to obtain a comprehensive representation of the events, as follows:

9. The rumor detection method based on temporal sentiment features as described in claim 1, characterized in that: Step S8 specifically includes: Using MLP as a classifier, classification is performed based on the propagation features of the comprehensive representation of posts. The calculation formula is as follows: Where {W * ,b * } represents the weights and biases of the neural network, ReLU(*) and σ(*) are the activation functions, and the final output is... For the model of event C i The prediction.

10. A rumor detection method based on temporal sentiment features as described in claim 1, comprising: This includes a post display module, an event display module, and a rumor classification module; The post representation module uses text features and temporal sentiment features to represent social network posts; The event representation module uses RvNN to capture the features of post representations according to the post propagation path, and obtains a comprehensive representation of the event; The rumor classification module determines whether an event is a rumor based on a comprehensive representation of the event. The post representation module, event representation module, and rumor classification module are connected in sequence.

Images