Rumor detection method for selecting comments based on sentiment values

Abstract

The present application relates to the technical field of rumor detection, and disclose a rumor detection method based on emotion value selection comment, comprising the following steps: 1) data preprocessing of rumor and comment.The present application has the advantages that: the rumor text and emotional features are used as output features, the most advanced pre-training model is selected, and the model performance is effectively improved by combining emotional features, the individual score of each comment in the emotional dictionary is calculated, the existing emotional dictionary is used to match and calculate the value of negative words and degree words, the emotion value of a certain type in the text is calculated according to the emotional words, degree adverbs and negative words in each text, the emotion value of each comment is calculated according to the emotional dictionary, the first L1 comment with the maximum emotion value and the first L2 comment with the minimum emotion value are selected, on the one hand, the words with larger emotion value, on the other hand, can effectively explain the truth or falsehood of the rumor, and only from the selected comments, the explanatory of the truth or falsehood of the rumor is significantly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rumor detection, and in particular to a rumor detection method for selecting comments based on sentiment values. Background Art

[0002] In the information age, social media is an important way for us to obtain news and has become an integral part of our lives. However, in recent years, the rampant and rapid spread of various rumors has seriously misled public opinion, harmed the public interest, and polluted the online environment.

[0003] Some scholars consider rumor detection to be part of fake news detection, while others believe rumors are fake news. However, the academic community generally defines rumors as information that spreads on social networks and remains unverified, or has been officially confirmed to be false. Currently, there are a variety of methods based on manual feature construction, deep learning, propagation models, and sentiment analysis. However, most of these methods focus on effectively detecting rumors with underlying characteristics. Even when sentiment features are used, they are only used as enhanced features and cannot explain why a rumor is detected as a rumor. Online rumors often have unusual emotional characteristics such as exaggeration and incitement, so using sentiment features as part of detection is particularly effective. This paper proposes a rumor detection method that selects comments based on sentiment values. This method extracts sentiment features by selecting comments based on sentiment values ​​and combines them with rumor semantic features for rumor detection. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of the existing technology, the present invention provides a rumor detection method for selecting comments based on sentiment values, which has the advantages of good detection effect and solves the problems raised by the background technology.

[0006] (2) Technical solution

[0007] To achieve the above-mentioned object, the present invention provides the following technical solution: a rumor detection method for selecting comments based on sentiment value, comprising the following steps:

[0008] 1) Preprocessing of rumors and comments;

[0009] 2) Comment selection: Use the sentiment dictionary to calculate the number of comments with the largest and smallest sentiment values ​​in the rumor comments;

[0010] 3) Calculating the sentiment features of rumor comments from multiple perspectives, including sentiment vocabulary features, sentiment intensity features, and auxiliary features;

[0011] 4) Use different models to embed text and verify the effectiveness of the model;

[0012] 5) Perform feature fusion and input the model to obtain the detection results.

[0013] Preferably, the preprocessing of data in step 1) mainly includes the deletion of useless data, and the deleted data features are rumor content length less than five, rumor comments are empty or empty after comment data processing, only forwarded comments, comments containing only @ and the length of the following characters is less than 5, and comments with other URLs, telephone numbers, user names, ID numbers, Chinese and English punctuation marks and stop words.

[0014] Preferably, when selecting comments in step 2), it is necessary to first calculate the individual scores of the words in the sentiment dictionary for each comment, and use the existing sentiment dictionary to match and calculate the values ​​of negative words and degree words, specifically the following formula:

[0015] s(x i )=D(x i )*neg(x i ,w)*deg(x i ,w)

[0016] Where D is the sentiment dictionary, x i is the vocabulary in the text, w represents the scope of the context, neg(x i ,w),deg(x i ,w) are the corresponding negation words and degree adverb values:

[0017]

[0018] The sentiment value of a certain category in each text is calculated based on the sentiment words, degree adverbs, and negation words in each text. The sentiment e of a text is the sum of all the sentiment words. L is the length of each comment, and the final sentiment value of the comment is calculated as:

[0019]

[0020] Calculate the sentiment value of each comment based on the sentiment dictionary, select the top L1 comments with the largest sentiment value and the top L2 comments with the smallest sentiment value. M1 and M2 represent the comments with the largest and smallest sentiment values, respectively. The details are as follows:

[0021]

[0022] Preferably, the calculation of the emotional features in step 3) uses the expression of emotional types in the emotional dictionary:

[0023] E={e1,e2,…,e d}

[0024] The corresponding word for a certain emotion e is We ={w1,w2,…,w l};

[0025] The calculation formula for the sentiment vocabulary features of the comment text is as follows:

[0026]

[0027] neg(x i ,w),deg(x i ,w) is the calculation of the corresponding negation word and degree adverb value, and then the characteristics of a certain emotion of the comment text are calculated. The formula is as follows:

[0028]

[0029] Finally, all the features extracted from the emotional vocabulary are connected to obtain:

[0030]

[0031] When calculating the emotional features of emotional vocabulary, it is necessary to consider the intensity of the vocabulary. Given the emotion E, the dictionary and text corresponding to each emotion, the emotional intensity of a certain emotion is first calculated using the following formula:

[0032]

[0033] where int(x i ) indicates the intensity of the word, (x i ,e) is the vocabulary x i In the score of sentiment category e, if x i In the dictionary, calculate int(x i ), otherwise it is 0. Finally, the emotion intensity feature is obtained by connecting the intensities of various emotions:

[0034]

[0035] Finally, auxiliary features are used to capture emoticons, punctuation marks, emotional words, and personal pronouns. In the comments, emoticons are mostly used instead of words, and auxiliary features such as emojis are obtained by using multiple emotional dictionaries. f , then the sentiment feature of each comment can be obtained by connecting each feature:

[0036]

[0037] Where "emo" is defined as "emotional feature vector emo q ,emo d ,emo f A collection of .

[0038] Preferably, in step 4), the article is embedded by the Bert model to obtain h c = Bert(c), the output semantic feature is represented as h c , connect the sentiment features and semantic features, and train the model through the cross entropy loss function:

[0039]

[0040] Where c represents the comment text.

[0041] (3) Beneficial effects

[0042] Compared with the existing technology, the present invention provides a rumor detection method based on sentiment value selection of comments, which has the following beneficial effects:

[0043] This rumor detection method for selecting comments based on sentiment value adopts rumor text and sentiment features as output features, selects the most advanced pre-trained model, and combines sentiment features to effectively improve model performance. It also calculates the individual scores of each comment word in the sentiment dictionary, uses the existing sentiment dictionary to match and calculate the values ​​of negative words and degree words, calculates the sentiment value of a certain category in the text based on the sentiment words, degree adverbs and negative words in each text, calculates the sentiment value of each comment based on the sentiment dictionary, selects the top L1 comments with the largest sentiment value and the top L2 comments with the smallest sentiment value. On the one hand, they are words with larger sentiment values, and on the other hand, they can effectively explain the truth of rumors. From the perspective of selecting comments alone, the interpretability of the truth of rumors is significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a framework diagram of the rumor detection model of the present invention;

[0045] Figure 2 This is a model diagram of the Bert model used in the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] Example 1:

[0048] The rumor detection method for selecting comments based on sentiment value includes the following steps:

[0049] 1) Preprocessing of rumors and comments;

[0050] 2) Comment selection: Use the sentiment dictionary to calculate the number of comments with the largest and smallest sentiment values ​​in the rumor comments;

[0051] 3) Calculating the sentiment features of rumor comments from multiple perspectives, including sentiment vocabulary features, sentiment intensity features, and auxiliary features;

[0052] 4) Use different models to embed text and verify the effectiveness of the model;

[0053] 5) Perform feature fusion and input the model to obtain the detection results.

[0054] The data preprocessing in step 1) mainly includes the deletion of useless data. The deleted data features are rumors with a content length of less than five, rumor comments that are empty or empty after comment data processing, comments that are only forwarded, comments that only contain @ and the length of the following characters is less than 5, and comments with other URLs, phone numbers, user names, ID numbers, Chinese and English punctuation marks and stop words. In the comment data set, rumor comments often contain a lot of truth. It is undeniable that there are also a large number of useless comments, especially simple forwarding behaviors, which are displayed as comment content "forwarded" in the data set, as well as some online emoticons, all of which are displayed in text mode. Therefore, before inputting text processing, rumors and comments need to be preprocessed and some rumors and comments need to be deleted.

[0055] When selecting comments in step 2), you need to first calculate the individual scores of the words in the sentiment dictionary for each comment, and use the existing sentiment dictionary to match and calculate the values ​​of negative words and degree words. Specifically, the formula is as follows:

[0056] s(x i )=D(x i )*neg(x i ,w)*deg(x i ,w)

[0057] Where D is the sentiment dictionary, x i is the vocabulary in the text, w represents the scope of the context, neg(x i ,w),deg(x i ,w) are the corresponding negation words and degree adverb values:

[0058]

[0059] The sentiment value of a certain category in each text is calculated based on the sentiment words, degree adverbs, and negation words in each text. The sentiment e of a text is the sum of all the sentiment words. L is the length of each comment, and the final sentiment value of the comment is calculated as:

[0060]

[0061] Calculate the sentiment value of each comment based on the sentiment dictionary, select the top L1 comments with the largest sentiment value and the top L2 comments with the smallest sentiment value. M1 and M2 represent the comments with the largest and smallest sentiment values, respectively. The details are as follows:

[0062]

[0063] After preprocessing and physical screening, some relatively irrelevant comments are removed. However, for rumor comments, on the one hand, there are still a large number of comments that are irrelevant to the rumor itself, which may be advertisements or some hot news content that users are browsing. On the other hand, the number of comments is too large, and the embedding matrix is ​​relatively sparse, making it difficult to achieve the desired results. Therefore, it is necessary to design a control mode to select some "valid" comments and control the relative effectiveness of rumor comments entering the model. The sentiment value reflects to a certain extent the reviewer's understanding and recognition of the text content. The sentiment value selects comments with strong emotions as part of the input features and can explain the truth of the rumor. Because this article uses the calculated sentiment value as the basis for selecting rumor comments, for each comment, the sentiment value is calculated by aggregating each word according to the sentiment dictionary. However, for a word in the dictionary, we must consider not only its frequency of occurrence, but also its context words, specifically degree adverbs, affirmative words, and negative words. Therefore, it is necessary to first calculate the individual score of each word in the sentiment dictionary for each comment;

[0064] The calculation of the emotional features in step 3) uses the expression of emotional types in the emotional dictionary:

[0065] E={e1,e2,…,e d}

[0066] The corresponding word for a certain emotion e is W e ={w1,w2,…,w l};

[0067] The calculation formula for the sentiment vocabulary features of the comment text is as follows:

[0068]

[0069] neg(x i ,w),deg(x i ,w) is the calculation of the corresponding negation word and degree adverb value, and then the characteristics of a certain emotion of the comment text are calculated. The formula is as follows:

[0070]

[0071] Finally, all the features extracted from the emotional vocabulary are connected to obtain:

[0072]

[0073] When calculating the emotional features of emotional vocabulary, it is necessary to consider the intensity of the vocabulary. Given the emotion E, the dictionary and text corresponding to each emotion, the emotional intensity of a certain emotion is first calculated using the following formula:

[0074]

[0075] where int(x i ) indicates the intensity of the word, (x i ,e) is the vocabulary x i In the score of sentiment category e, if x i In the dictionary, calculate int(x i ), otherwise it is 0. Finally, the emotion intensity feature is obtained by connecting the intensities of various emotions:

[0076]

[0077] Finally, auxiliary features are used to capture emoticons, punctuation marks, emotional words, and personal pronouns. In the comments, emoticons are mostly used instead of words, and auxiliary features such as emojis are obtained by using multiple emotional dictionaries. f , then the sentiment feature of each comment can be obtained by connecting each feature:

[0078]

[0079] Where emo is defined as the emotional feature vector emo q ,emo d ,emo f A collection of .

[0080] In order to obtain the emotional features of the comments, we extracted emotional vocabulary, emotional intensity, emotional value, auxiliary features and other features. The emotional intensity and emotional value provide overall information, while the emotional vocabulary and auxiliary features provide information at the vocabulary and symbol level. The emotion of the text is usually expressed by the vocabulary. Looking at different words from different perspectives, different features are extracted to reflect the overall emotional characteristics, which are usually expressed by several key words. The extraction of all features depends on the emotional dictionary annotated by experts. The difference is that different features are extracted from different dictionaries. Given a text of length L, the text composed of words T = [x1, x2, ... x L ], calculate the sentiment features of the text;

[0081] In step 4), the Bert model is used to embed the text to obtain h c = Bert(c), the output semantic feature is represented as h c, connect the sentiment features and semantic features, and train the model through the cross entropy loss function:

[0082]

[0083] Where c represents the comment text.

[0084] The BERT model is a natural language processing model based on deep learning. Its primary model structure is the Transformer encoder. Through self-attention modeling, it can directly capture the global information of the text. Furthermore, because it lacks a forget gate, all word information is retained. Therefore, BERT can better express the complete semantic information of a sentence and directly identify correlations between words from global word features. Compared to traditional NLP models such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), it has more powerful text encoding capabilities. Given these advantages of the BERT model and its success in the NLP field, we are working to extract rumor semantic representation vectors based on the BERT model, addressing the need to rely on contextual information to better extract semantic information when processing Chinese text.

[0085] Experimental example:

[0086] To eliminate the influence of rumor detectors, this experiment uses only sentiment features to detect fake news using a 5-layer MLP model. The results are shown in Table 2. Publisheremotion is the sentiment feature extracted from rumor text, Dualemotion is the representation method used by Zhang, and Select100s is the sentiment feature obtained by screening comments in this article.

[0087] Table 2 Results of sentiment features

[0088]

[0089] In order to prove the effectiveness of the current model, several advanced rumor detection models were selected. The experimental results are shown in Table 3:

[0090] Table 3 Comparative model experimental results

[0091]

[0092] During the experiment, the dEFEND model uses rumor text and comments as input, while other models only use rumor text as input. The model proposed in this paper uses rumor text and sentiment features as output features. Table 3 shows that the Bert model shows superior performance because when only using rumor text as input, the Bert model can better embed text to represent semantics. The CNN model and the RNN model have similar effects. The dEFEND model performs poorly because it uses the co-attention mechanism to learn the correlation between rumor text and comments, which cannot capture the close correlation between comments and rumors. The comments rarely directly mention the content involved in the rumor text. The experimental results of the model proposed in this paper on two Weibo datasets show good performance, proving that selecting the most advanced pre-training model and combining sentiment features can improve model performance to a certain extent. Compared with the Bert model, the accuracy on the two datasets increased by 4.2% and 3.1% respectively.

[0093] Without employing an attention mechanism, we attempt to explain rumors using the comments selected by the minimum comment value. It's noteworthy that selecting only the comments corresponding to the top three largest and smallest sentiment values ​​can explain the truth of the rumor. The selected comments are highly correlated with the news content, and the correlation is higher for the comments with the largest sentiment values. For the comments with the smallest sentiment values, the content relevance is lower from a direct semantic perspective, but the sentiment contained in these comments is inversely proportional to the rumor itself. Black font is used for annotations; on the one hand, it indicates words with high sentiment values, and on the other hand, it can effectively explain the truth of the rumor. Judging from the selected comments alone, it has a significant effect on explaining the truth of the rumor.

[0094] Judgment criteria: Common standards include accuracy (acc), precision (pre), recall (rec) and f1.

[0095] The parameters are set as follows: the maximum length of rumor sentences is 128, the maximum length of comment sentences is 32, and the number is 100; Bert's hide_size = 768.

[0096] The beneficial effects of the present invention are: by adopting rumor text and sentiment features as output features, selecting the most advanced pre-trained model, and combining with sentiment features, the model performance is effectively improved, and the individual scores of the words in the sentiment dictionary of each comment are calculated, and the existing sentiment dictionary is used to match and calculate the values ​​of negative words and degree words. The sentiment value of a certain category in the text is calculated according to the sentiment words, degree adverbs and negative words in each text, the sentiment value of each comment is calculated according to the sentiment dictionary, and the top L1 with the largest sentiment value and the top L2 with the smallest sentiment value are selected. On the one hand, they are words with larger sentiment values, and on the other hand, they can effectively explain the truth of rumors. From the perspective of selecting comments alone, the interpretability of the truth of rumors is significantly improved.

[0097] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A rumor detection method based on sentiment value selection of comments, characterized in that: The following steps are involved: 1) Data preprocessing of rumors and comments; 2) Comment selection: Use the sentiment dictionary to calculate the number of comments with the largest and smallest sentiment values ​​among the rumor comments; 3) Calculate the sentiment characteristics of rumor comments from multiple perspectives; 4) Use different models to embed text and verify the effectiveness of the model; 5) Perform feature fusion and input the model to obtain the detection results; The calculation of the emotional features in step 3) uses the expression of emotional types in the emotional dictionary: ； One of the emotions The corresponding vocabulary is ; The calculation formula for the sentiment vocabulary features of the comment text is as follows: ； , To calculate the corresponding negation word and degree adverb values, and then calculate the characteristics of a certain emotion of the comment text, the formula is as follows: ； Finally, all the features extracted from the emotional vocabulary are connected to obtain: ； When calculating the emotional features of emotional words, it is necessary to consider the intensity of the words. , the dictionary and text corresponding to each emotion, first calculate the emotional intensity of a certain emotion, the formula is as follows: in Indicates the intensity of the word, if In the dictionary, it is calculated based on the sentiment dictionary , otherwise it is 0. Finally, the emotion intensity feature is obtained by connecting the intensities of various emotions: ； Finally, auxiliary features are used to capture emoticons, punctuation marks, emotional words and personal pronouns. In the comments, emoticons are mostly used instead of words, and auxiliary features are obtained by using multiple emotional dictionaries. , then the sentiment feature of each comment can be obtained by connecting each feature: 。 2. The rumor detection method for selecting comments based on sentiment value according to claim 1 is characterized in that: The data preprocessing in step 1) mainly includes the deletion of useless data. The deleted data features are rumors with a content length of less than five, empty rumor comments or empty after comment data processing, only forwarded comments, comments containing only @ and the length of the following characters is less than 5, and comments with other URLs, phone numbers, user names, ID numbers, Chinese and English punctuation marks and stop words.

3. The rumor detection method based on sentiment value selection of comments according to claim 1 is characterized in that: When selecting comments in step 2), it is necessary to first calculate the individual scores of the words in the sentiment dictionary for each comment, and use the existing sentiment dictionary to match and calculate the values ​​of negative words and degree words. Specifically, the formula is as follows: ； in For the sentiment dictionary, is the vocabulary in the text, Indicates the scope of the context, , For the corresponding negation and degree adverb values: ； ； The sentiment value of a certain category in the text is calculated based on the sentiment words, degree adverbs and negation words in each text. is the sum of all the sentiment words, is the length of each comment, and the final sentiment value of the comment is calculated as follows: ； Calculate the sentiment value of each comment according to the sentiment dictionary and select the top comment with the largest sentiment value. The first one with the smallest sentiment value strip, , Represents the comments with the largest and smallest sentiment values, respectively, as follows: ； 。 4. The rumor detection method based on sentiment value selection of comments according to claim 1 is characterized in that: In step 4), the article is embedded using the Bert model. , the output semantic feature is expressed as , connect the sentiment features and semantic features, and train the model through the cross entropy loss function: 。

Images