Method for detecting fake news based on title content similarity and author implied sentiment

Abstract

The application provides a false news detection method based on title text similarity and author implied sentiment, which extracts semantic features of text through a text semantic feature extraction module, extracts semantic features of images through an image semantic feature extraction module, extracts potential sentiment features of the author in the text through an author implied sentiment feature extraction module, extracts similarity features between news titles and texts through a title text similarity extraction module, and predicts the authenticity of news by using the obtained feature vectors through a false news detection module. Through the joint work of the five modules, the semantic level features of the text and the images of the multi-modal news, the potential sentiment features containing the subjective intention of the author and the similarity features of the title and the text are well captured, so that the false news detection in a complex scene is more suitable than the existing model. The application shows great advantages compared with existing methods on two public data sets.

Description

Technical Field

[0001] This invention relates to the field of natural language processing, and in particular to a method for detecting fake news based on the similarity between the title and the main text and the author's implicit sentiment. Background Technology

[0002] Today, more and more people obtain news information through social media. However, due to the anonymity of the internet, a large amount of fake news spreads widely and is difficult for users to distinguish. The spread of fake news often causes irreparable and enormous losses and harm to individuals and the public. A study that thoroughly analyzed 126,000 verified true and fake news posts on Twitter from 2006 to 2017 indicated that fake news can spread faster and wider than fact-based news. Furthermore, according to psychological and social science perspectives, the more fake news spreads, the more likely users on social media are to spread and believe it due to repeated exposure or peer pressure. Therefore, to prevent the spread of fake news on social media, extensive research has been conducted on effective methods for identifying fake news.

[0003] Fake news detection methods can be broadly categorized into content-based methods and social context-based methods. The main difference lies in whether they rely on social context information: that is, information about how the news spreads on social media, including numerous auxiliary details about social media users and their connections / networks. While many existing methods utilize social context information, detecting fake news using these methods is often difficult when it has just been published and has not yet spread, as it is difficult to capture this social context. Therefore, most existing fake news detection research still focuses on content-based methods.

[0004] A news article typically contains both textual and visual information. To attract public attention, authors often express highly subjective emotions, which differ significantly from the emotions conveyed in objective statements in general news texts. However, existing content-based fake news detection methods only consider superficial features such as text and images, neglecting the underlying emotional characteristics of the author within fake news. Furthermore, in terms of writing style, some authors attempt to attract readers' attention by using headlines unrelated to the content; these characteristics are also overlooked by existing methods. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, this invention proposes a more general and effective method for detecting fake news based on the similarity between the headline and the body text and the author's implicit emotions. This method proposes a new perspective for detecting fake news by utilizing the subjective emotions of news authors and a new attention mechanism to fully explore the similarity between the headline and the body text. It solves the problems of existing fake news detection methods ignoring the potential emotions containing the author's subjective intentions and paying less attention to the similarity relationship between the news headline and the body text, which leads to low detection accuracy.

[0006] To achieve the above objectives, this invention proposes a fake news detection method based on title-body similarity and author's implicit sentiment, denoted as TieFake. This method includes a text semantic feature extraction module, an image semantic feature extraction module, an author's implicit sentiment feature extraction module, a title-body similarity extraction module, and a fake news detection module, comprising the following steps:

[0007] Step 1: Construct a text semantic feature extraction module to extract semantic features from the news title and body text respectively;

[0008] Step 2: Construct an image semantic feature extraction module to extract the semantic features of the image;

[0009] Step 3: Construct an author latent sentiment feature extraction module to extract the author's latent sentiment features from the text;

[0010] Step 4: Construct a title-body similarity extraction module to extract similarity features between news titles and body text;

[0011] Step 5: Construct a fake news detection module. Using the semantic features of the news headline, the news body, the image, the author's potential emotional features in the text, and the similarity features between the news headline and the body, extracted in the previous steps, multi-level semantic features are concatenated to obtain multi-level semantic features. These multi-level semantic features are used by the fake news detection module to predict the authenticity of the news.

[0012] Preferably, in the text semantic feature extraction module, text semantic features are extracted from the penultimate layer before the classification layer of the BERT pre-trained model, and the feature length is reduced by a fully connected layer.

[0013] Preferably, in the image semantic feature extraction module, image semantic features are extracted from the last layer of the ResNeSt-50 pre-trained model, and the length of the image semantic features is reduced by using a fully connected layer.

[0014] Preferably, in the author implicit sentiment feature extraction module, the existing author sentiment extractor in reference [1]: Mining Dual Emotion for Fake News Detection is used to obtain multi-dimensional sentiment features such as the author's sentiment category, sentiment vocabulary, sentiment intensity, sentiment score and other auxiliary features from the news text content.

[0015] Preferably, in the title-text similarity extraction module, a scaled dot product attention mechanism is applied to the title and text of the news from two levels to capture the similarity between the title and text in the news.

[0016] Preferably, in the fake news detection module, features obtained through the text feature extraction module, image semantic feature extraction module, author implicit sentiment feature extraction module, and title-text similarity extraction module are concatenated and then input into the sigmoid function to obtain the predicted probability P of the news being true or false.

[0017] This invention proposes a fake news detection method based on headline-text similarity and author's implicit emotion. It utilizes a novel perspective and attention mechanism to fully explore the similarity between headlines and text, addressing the problems of existing fake news detection methods that neglect the potential emotions containing the author's subjective intentions and pay insufficient attention to the similarity between headlines and text, resulting in low detection accuracy. Furthermore, through the collaborative work of five sub-modules (text semantic feature extraction module, image semantic feature extraction module, author implicit emotion feature extraction module, headline-text similarity extraction module, and fake news detection module), it better mines the semantic information of different modalities in news data, thus achieving better fake news detection results than existing methods in various scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method of the present invention;

[0019] Figure 2 This is a schematic diagram of the fake news detection framework based on title-text similarity and author's implicit sentiment in this invention;

[0020] Figure 3 This is a schematic diagram of the title-text similarity extraction module in this invention (taking the similarity level of the title applied to the semantic features of the text as an example). Detailed Implementation

[0021] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0022] Please see Figure 1-3 This invention provides a method for detecting fake news based on the similarity between the title and the main text, and the author's implicit sentiment. This method is denoted as TieFake. Figure 1 As shown, the specific steps include the following:

[0023] Step 1: Construct a text semantic feature extraction module to extract semantic features from the news text and news title respectively;

[0024] Step 2: Construct an image semantic feature extraction module to extract the semantic features of the image;

[0025] Step 3: Construct an author latent sentiment feature extraction module to extract the author's latent sentiment features from the text;

[0026] Step 4: Construct a title-body similarity extraction module to extract similarity features between news titles and body text;

[0027] Step 5: Construct a fake news detection module. Using the semantic features of the news headline and body, the semantic features of the news body, the semantic features of the image, the author's potential emotional features in the text, and the similarity features between the news headline and body, extracted in the previous steps, multi-level semantic features are concatenated to obtain multi-level semantic features. These multi-level semantic features are used by the fake news detection module to predict the authenticity of the news.

[0028] like Figure 2 The diagram shown illustrates the structure of a fake news detection framework based on title-text similarity and author implicit sentiment used to implement the method of this invention. It includes a text semantic feature extraction module, an image semantic feature extraction module, an author implicit sentiment feature extraction module, a title-text similarity extraction module, and a fake news detection module.

[0029] In this embodiment, in the text semantic feature extraction module, for a news article containing text content (title + body text) and an image, the body text is input into the BERT pre-trained model, and the semantic features of the body text are extracted from the penultimate layer of the model. The extracted semantic features are then reduced to a final semantic feature R of length 32 using a fully connected layer 1. T Calculate using the following formula (1):

[0030]

[0031] Where: σ(*) is the ReLU activation function, W vt is the weight matrix of the fully connected layer 1, and × is the matrix multiplication operator.

[0032] Similarly, the headline content of the news is input into the BERT pre-trained model, and the semantic features of the headline are extracted from the penultimate layer of the model. The extracted title semantic features are then reduced to the final body text semantic features R with a length of 32 using a fully connected layer 2. Ti Calculate using the following formula (2):

[0033]

[0034] Where: σ(*) is the ReLU activation function, W vti is the weight matrix of the fully connected layer 2, and × is the matrix multiplication operator.

[0035] In this embodiment, in the image semantic feature extraction module, for a news article containing a title, body text, and an image, the news image is input into the ResNeSt-50 pre-trained model, and image features are extracted from the last layer of the model. The extracted image features are then reduced to a final image feature R of length 30 using a fully connected layer 3. V Calculate using the following formula (3):

[0036]

[0037] Where: σ(*) is the ReLU activation function, W vf is the weight matrix of the fully connected layer 3, and × is the matrix multiplication operator.

[0038] In this embodiment, in the author's implicit emotion feature extraction module, in order to comprehensively mine the author's implicit potential emotions in the text, this invention uses the existing author emotion extractor in reference [1]: Mining Dual Emotion for Fake News Detection to obtain the author's multidimensional emotion features from the news text content, including emotion category, emotional lexicon, emotional intensity, sentiment score, and other auxiliary features. The corresponding feature representations are as follows: and Of these, sentiment category, sentiment intensity, and sentiment score provide overall sentiment information, while the other two features provide sentiment information at the word and symbol levels. This invention connects all five features to form a final sentiment feature R of length 38. E Calculate using the following formula (4):

[0039]

[0040] Where ⊕ is the vector concatenation operator.

[0041] In this embodiment, in the title-body similarity extraction module, the title and body content of the news are respectively processed using the news text feature extraction module to obtain the title semantic feature R. Ti and semantic features of the main text content R T The scaling dot product attention mechanism is applied to the semantic features of the title and the semantic features of the body text from two levels: the similarity of the body text applied to the semantic features of the title and the similarity of the title applied to the semantic features of the body text, respectively, in order to fully capture the similarity between the title and the body text in the news.

[0042] When applying a scaled dot-product attention mechanism at the similarity level applied from the body text to the title semantic features, the body text semantic features R are used. T In the scaling dot product attention mechanism, the semantic features R of the query and title are calculated. Ti When calculating the Key and Value in the scaling dot product attention mechanism (for ease of description, Q, K, and V will be used as abbreviations for Query, Key, and Value respectively), Q, K, and V are calculated using the following formula (5):

[0043] Q = R T ×W Q K = R Ti ×W K V = R Ti ×WV (5)

[0044] Where: × is the matrix multiplication operator, W Q W K W V These are the weight matrices corresponding to Q, K, and V, respectively.

[0045] Similarity output matrix Att applied to scaled dot product attention mechanism on Q, K, and V T→Ti Calculate using the following formula (6):

[0046]

[0047] Where: K T It is the transpose of K. is the scaling factor, d has a value of 32, softmax(*) is the activation function, and × is the matrix multiplication operator.

[0048] The extracted similarity output matrix Att T→Ti Inputting a fully connected layer 4 yields a 32-bit similarity feature R applied to the semantic features of the title. T→Ti Calculate using the following formula (7):

[0049] R T→Ti =σ(Att) T→Ti ×W t→ti (7)

[0050] Where: σ(*) is the ReLU activation function, W t→ti is the weight matrix of the fully connected layer 4, and × is the matrix multiplication operator.

[0051] Similarly, when applying a scaled dot product attention mechanism at the similarity level applied from the title to the semantic features of the body text, using the title semantic features R... Ti Calculate Q and text semantic features R in the scaling dot product attention mechanism. t When calculating K and V in the scaled dot product attention mechanism, Q, K, and V are calculated using the following formula (8):

[0052] Q = R Ti ×W Q K = R T ×W K V = R T ×W V (8)

[0053] Where: × is the matrix multiplication operator, W Q W K W V These are the weight matrices corresponding to Q, K, and V, respectively.

[0054] Similarity output vector Att applied to scaled dot product attention mechanism on Q, K, and V Ti→T Calculate using the following formula (9):

[0055]

[0056] Where: K T It is the transpose of K. is the scaling factor, d has a value of 32, softmax(*) is the activation function, and × is the matrix multiplication operator.

[0057] The extracted similarity output matrix Att Ti→T Inputting a fully connected layer 5 yields a 32-bit similarity feature R applied from the body text to the title semantic features. Ti→T Calculate using the following formula (10):

[0058] R Ti→T =σ(Att) Ti→T ×W ti→t (10)

[0059] Where: σ(*) is the ReLU activation function, W ti→t is the weight matrix of the fully connected layer 5, and × is the matrix multiplication operator.

[0060] In this embodiment, in the fake news detection module, the features obtained from the news through the text semantic feature extraction module, image semantic feature extraction module, author implicit sentiment feature extraction module, and title-text similarity extraction module are concatenated to obtain the final representation feature R of the news. F It is calculated using the following formula (11):

[0061] R F =R F ⊕R V ⊕R E ⊕R T→Ti ⊕R Ti→T (11)

[0062] Where: ⊕ is the concatenation operator.

[0063] The final representation feature R after splicing F The dimension is reduced to 1 through a fully connected layer 6, and then the sigmoid function is input to obtain the predicted probability P of the news true / false category, which is calculated using the following formula (12):

[0064] P = sigmoid(R) F ×W f (12)

[0065] Where: sigmoid(*) is the activation function, W f is the weight matrix of the fully connected layer 6, and × is the matrix multiplication operator.

[0066] If P ≥ 0.5, the predicted category of the news is true; if P < 0.5, the predicted category of the news is false.

[0067] This invention is implemented using the PyTorch deep learning framework. To illustrate the rationality and interpretability of this invention, it is implemented on two different domain sub-datasets, PolitiFact and GossipCop, contained in the FakeNewsNet public dataset.

[0068] In addition, to eliminate noise, this invention preprocessed two subsets of the dataset, removing news articles without any text or images. The final dataset information is shown in Table 1.

[0069] Table 1 Dataset Information

[0070]

[0071] To improve the accuracy of the TieFake model in detecting fake news, it needs to be trained multiple times. This invention divides each subset of the dataset into training and test sets in an 8:2 ratio. The pre-trained BERT and ResNeSt-50 weights are kept unchanged and used as feature extractors. During training, the TieFake model weights are updated using stochastic gradient descent; Adam is used as the optimizer with an initial learning rate of 0.0001; and StepLR is used to adjust the learning rate, reducing it to 0.1 every 7 epochs. Dropout is applied to every fully connected layer in the TieFake model to avoid overfitting, and the number of iteration epochs is set to 10. Training the TieFake model should be stopped prematurely when the accuracy obtained after three consecutive epochs does not exceed the current maximum value.

[0072] To illustrate the importance of title-text similarity and author's implicit sentiment in this invention, the TieFake method of this invention is first compared with other fake news detection methods, and then an ablation experiment is conducted to demonstrate the advantages of this invention.

[0073] This invention selects the classic evaluation metrics of accuracy, precision, recall, and F1 score as the evaluation criteria for performance comparison.

[0074] The TieFake method of this invention was compared with three other methods on two subset datasets, and the results are shown in Table 2.

[0075] 1. Fake news detection methods based on text features, such as CNN, GRU, BERT, etc.

[0076] 2. Image feature-based fake news detection methods, such as VGG19 and RESNET50.

[0077] 3. Fake news detection methods based on multimodal features, such as MVAE, att-RNN, SpotFake, EANN, SpotFake+, SAFE, etc.

[0078] Table 2 Comparison of Fake News Detection Results

[0079]

[0080] As shown in Table 2, the TieFake method of this invention has a significant advantage over fake news detection methods based solely on text features or image features in the four metrics of the two subset datasets. This is a direct result of the invention's simultaneous mining of multiple modal features.

[0081] The TieFake method of this invention outperforms (existing) state-of-the-art methods on both the PolitiFact and GossipCop datasets. Compared to SAFE, the state-of-the-art method on the PolitiFact dataset, TieFake improves accuracy by 4.0%, precision by 4.8%, recall by 1.2%, and F1 score by 3.0%. Compared to SpotFake+, the state-of-the-art method on the GossipCop dataset, TieFake improves accuracy by 2.8%, precision by 2.8%, recall by 2.0%, and F1 score by 2.4%. This verifies that the proposed TieFake method can jointly capture multimodal (text and image) and sentiment information, which can better characterize the underlying representation of news content and improve the performance of fake news detection. The TieFake method of this invention achieves good results on both small sample datasets (PolitiFact dataset) and large sample datasets (GossipCop dataset), which also demonstrates the versatility of the proposed TieFake method.

[0082] Table 3 shows the ablation experiments of the present invention, where TieFake-T, TieFake-V, TieFake-E, and TieFake-S are variations of the TieFake method in the present invention by removing the text semantic feature extraction module, image semantic feature extraction module, author implicit sentiment feature extraction module, and title-text similarity extraction module, respectively.

[0083] Table 3 Ablation Experiment Results

[0084]

[0085] Table 3 shows that integrating textual and image information from news articles with the author's latent emotional features can improve the accuracy of identifying fake news. The TieFake method outperforms TieFake-E, demonstrating the effectiveness of incorporating the author's latent emotional features in detecting fake news. The TieFake method also outperforms TieFake-S, demonstrating the effectiveness of mining the similarity between the headline and body text in detecting fake news. Furthermore, the TieFake method outperforms TieFake-V, demonstrating the effectiveness of mining the image semantic information contained in the news content in detecting fake news. Compared to other methods, the TieFake method shows the largest difference in results compared to TieFake-T, indicating that the textual semantic information contained in the news content plays a crucial role in detecting fake news.

[0086] In summary, analyzing the similarity between the title and the main text, as well as the author's implicit emotions, plays a significant role in this invention. Those skilled in the art will recognize that the examples described herein are intended to help readers understand the principles of the invention and should be understood as not limiting the scope of protection of the invention to such specific statements and examples. Various modifications and variations can be made to this invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for detecting fake news based on title-text similarity and the author's implicit sentiment, characterized in that, The method includes the following steps: Step 1: Construct a text semantic feature extraction module to extract semantic features from the news title and body text respectively; Step 2: Construct an image semantic feature extraction module to extract the semantic features of the image; Step 3: Construct an author latent sentiment feature extraction module to extract the author's latent sentiment features from the text; Step 4: Construct a title-body similarity extraction module to extract similarity features between news titles and body text; Step 5: Construct a fake news detection module. Using the semantic features of the news headline, the semantic features of the news body, the semantic features of the image, the author's potential emotional features in the text, and the similarity features between the news headline and the body, which are extracted in the above steps, multi-level semantic features are spliced ​​together to obtain multi-level semantic features. These multi-level semantic features are used by the fake news detection module to predict the authenticity of the news. In the author implicit sentiment feature extraction module, an existing author sentiment extractor is used to obtain multidimensional sentiment features of the author from the news text content, including sentiment category, sentiment vocabulary, sentiment intensity, sentiment score, and other auxiliary features. The corresponding feature representations are as follows: , , , and Connect all five features mentioned above to form a final sentiment feature with a length of 38. Calculate using the following formula (4): （4） in: It is a vector concatenation operator; In the title-text similarity extraction module, the title and text content of the news are analyzed using... and The scaling dot product attention mechanism is applied to the semantic features of the title and the semantic features of the body text from two levels: the similarity of the body text applied to the semantic features of the title and the similarity of the title applied to the semantic features of the body text, respectively, in order to fully capture the similarity between the title and the body text in the news. When applying a scaled dot-product attention mechanism at the similarity level applied from the body text to the title's semantic features, the body text semantic features are used. Query and title semantic features in the scaling dot product attention mechanism When calculating the Key and Value in the scaling dot product attention mechanism, for ease of expression, Q, K, and V will be used as abbreviations for Query, Key, and Value respectively. Q, K, and V are calculated using the following formula (5): （5） in: , , These are the weight matrices corresponding to Q, K, and V, respectively; Similarity output matrix of scaled dot product attention mechanism applied to Q, K, and V Calculate using the following formula (6): （6） in: The transpose of K is the scaling factor, d has a value of 32, and softmax(*) is the activation function; The extracted similarity output matrix Inputting into fully connected layer 4 yields a similarity feature of length 32 applied to the semantic features of the title within the main text. Calculate using the following formula (7): （7） in: It is the weight matrix of the fully connected layer 4; When applying a scaled dot-product attention mechanism at the similarity level applied to the semantic features of the title on the body text, the semantic features of the title are used. Calculating Q-values ​​and semantic features of the text in the scaling dot product attention mechanism When calculating K and V in the scaled dot product attention mechanism, Q, K, and V are calculated using the following formula (8): （8） Similarity output vectors applied to scaled dot product attention mechanisms on Q, K, and V Calculate using the following formula (9): （9） The extracted similarity output matrix Inputting a fully connected layer 5 yields a similarity feature of length 32 applied to the semantic features of the title within the main text. Calculate using the following formula (10): （10） in: It is the weight matrix of the fully connected layer 5; In the fake news detection module, features obtained from the news through the text semantic feature extraction module, image semantic feature extraction module, author implicit sentiment feature extraction module, and title-text similarity extraction module are concatenated to obtain the final representation features of the news. It is calculated using the following formula (11): （11） in: It is a concatenation operator; The final representation features after splicing The dimension is reduced to 1 through a fully connected layer 6, and then input into the sigmoid function to obtain the predicted probability of the news being true or false. Calculate using the following formula (12): （12） Where: sigmoid(*) is the activation function. It is the weight matrix of the fully connected layer 6; like If so, the news prediction category is true; if false.

2. The method for detecting fake news based on title-text similarity and author's implicit sentiment as described in claim 1, characterized in that, In the text semantic feature extraction module, for a news article containing text content and images, where the text content includes a title and body text, the body text is input into the BERT pre-trained model, and the semantic features of the body text are extracted from the penultimate layer of the model. The extracted semantic features are then reduced to a final semantic feature length of 32 through a fully connected layer 1. The process is calculated using the following formula (1): （1） in: It is the ReLU activation function. It is the weight matrix of fully connected layer 1. It is a matrix multiplication operator; The news headline content is input into the BERT pre-trained model, and the semantic features of the headline are extracted from the penultimate layer of the model. The extracted title semantic features are then reduced to a final title semantic feature of length 32 through a fully connected layer 2. The process is calculated using the following formula (2): （2） in: It is the weight matrix of the fully connected layer 2.

3. The method for detecting fake news based on title-text similarity and author's implicit sentiment as described in claim 2, characterized in that, In the image semantic feature extraction module, for a news article containing a title, body text, and an image, the news image is input into the ResNeSt-50 pre-trained model, and image features are extracted from the last layer of the model. The extracted image features are then reduced to a final image feature length of 30 using a fully connected layer 3. The process is calculated using the following formula (3): （3） in: It is the weight matrix of the fully connected layer 3.

4. The method for detecting fake news based on title-text similarity and author's implicit sentiment as described in claim 3, characterized in that, In the author's implicit sentiment feature extraction module, the three features of sentiment category, sentiment intensity and sentiment score provide overall sentiment information, while the two features of sentiment vocabulary and auxiliary features provide sentiment information at the word and symbol levels.

Images