Social network false message detection method based on large model and multi-modal fusion

Abstract

The invention discloses a social network false message detection method based on large model and multi-modal fusion. According to the method, based on a given multi-modal message, for an image and a text, an embedding vector and aligned embedding are obtained respectively; utilizing the large model to obtain a background knowledge embedding vector; splicing and inputting the two aligned embedding pieces into a full connection layer to obtain a multi-modal hybrid aligned embedding vector; inputting image embedding, text embedding, hybrid alignment embedding and background knowledge embedding into an N-layer stacked progressive Transform model, and outputting an embedded vector after multi-modal fusion; and inputting the fused embedded vector into a classifier, and setting a cross entropy loss function at the same time. According to the detection method, texts, images, text-image alignment characteristics and background knowledge assisted by a large model are fused, false message detection is assisted by the background knowledge generated by the large language model, and different modal information is effectively fused by a designed stacking progressive model; and the accuracy of false message detection is improved from the perspective of multi-modal fusion.

Description

Technical Field

[0001] The present invention relates to the field of false message detection in social networks, and in particular to a false message detection method in social networks based on a large model and multimodal fusion. Background Art

[0002] Social networking platforms play a vital role in facilitating the sharing of personal information in today’s society. However, the lack of verification checks on user-generated content on these platforms has led to the spread of false information. This phenomenon is worrisome because the popularity of social networks allows for rapid and widespread dissemination of content, including tweets on the platforms. As a result, this phenomenon has had an adverse impact on public events. For example, during the COVID-19 pandemic, false information has continued to spread on social media platforms. This false information has led to panic buying and the adoption of ineffective preventive measures by the public, thus endangering public safety. These incidents highlight the urgency of automatically detecting fake news and raising public awareness.

[0003] In recent years, many researchers have explored different methods for automatic false news detection. Early graph-based optimization methods treat tweets as nodes in a graph and evaluate the credibility of each node based on the credibility of its neighbors, whose authenticity has been verified. This method focuses more on the relevance of news rather than its content. Content-based methods for automatic false news detection place more emphasis on the text content of tweets. Traditional text-based features, including language patterns or knowledge for fact-checking, often fail to reveal the user's writing style. Recently, a large number of deep neural networks have been used for false news detection. For example, the University of California proposed using convolutional neural networks (CNNs) to extract text style features and use the powerful capabilities of deep neural networks to reveal the underlying writing style. Universities such as the State University of New York use deep neural networks to extract multimodal features from text and images, achieving better performance than single modal methods. However, these deep learning-based methods often lack the integration of external knowledge, which limits the ability to integrate these multimodal features with important external information.

[0004] In view of this, the inventor of this case focuses on how to use multimodal information such as images, texts, and external background knowledge to propose a deep learning model that effectively integrates the above multiple key factors to predict false news. By detecting and identifying false news, we can protect the public interest, maintain social order, help prevent false information from misleading the public, and reduce its potential negative impact. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for detecting false messages in social networks based on a large model and multimodal fusion.

[0006] The technical solution adopted by the present invention to solve the technical problem comprises the following steps:

[0007] Step 1: Given a multimodal message, the message is represented as a text-image pair. I represents image and T represents text. For image I, the embedding vector of the image is extracted by Masked Autoencoder (MAE) to obtain p i ; For image I, extract the image alignment embedding f through the image encoder in the CLIP (Contrastive Language-Image Pre-Training) pre-training model CLIP-I .

[0008] Step 2: For the text T, extract the text embedding vector p through the BERT model j ; For the text T, extract the text alignment embedding f through the text encoder in the CLIP pre-trained model CLIP-T .

[0009] Step 3: Use the ChatGLM model to design a prompt. Input the false message text T into ChatGLM, and output the relevant background knowledge T' of the false message T. Input T' into BERT to obtain the embedding vector p of the background knowledge. k .

[0010] Step 4: Align the image embedding f CLIP-I Alignment embedding with text f CLIP-T Concatenate and input into the fully connected layer to obtain the multimodal mixed alignment feature p m .

[0011] Step 5: vector p i 、p j 、p k and p m Input N layers of stacked progressive Transformer model, output multimodal fusion embedding vector p n .

[0012] Step 6: Multimodal fusion vector p n Input Sigmoid classifier and set cross entropy loss function The model is trained using the Adam optimization method. After learning, the model parameters are used to detect the message (text-image pair) that needs to be predicted, and the detection result of whether it is a false message is output.

[0013] The specific implementation process of step 1 is as follows:

[0014] 1-1. Given a multimodal message I stands for image and T stands for text.

[0015] 1-2. Scale the image I to 224*224 and use the MAE module to get the embedding vector p of image I i .

[0016] 1-3. Scale the image I to 224*224 and use the image-encoder module in the pre-trained CLIP model to get the aligned embedding f of image I CLIP-I .

[0017] The specific implementation process of step 2 is as follows:

[0018] 2-1. Use the Chinese Bert pre-training model to extract the embedding vector of each word in the text T, and then use the Meanpooling method to output the mean of each word as the sentence vector of the text t. The present invention represents the embedding vector of the text T as p i =BERT(T).

[0019] 2-2. Use the text-encoder module in the pre-trained CLIP model to get the aligned embedding f of the text T CLIP-t .

[0020] The specific implementation process of step 3 is as follows:

[0021] 3-1. Based on ChatG LM The big model design prompt is as follows: "You are designed to be an auxiliary tool for ****; your main function is ****; your goal is to output ****; your explanation should be less than *** words."

[0022] 3-2. Enter the message text into ChatG LM In the output, the relevant background knowledge T′ of message t is output.

[0023] 3-3. Use the Chinese Bert pre-trained model to extract the embedding vector of each word in T', and then use the Meanpooling method to output the mean of each word as the embedding vector of text T', specifically expressed as p k =BERT(T').

[0024] The specific implementation process of step 4 is as follows:

[0025] 4-1. Alignment of images into f CLIP-I Alignment embedding with text f CLIP-T Concatenate and input into the fully connected layer to obtain the multimodal mixed feature p m The specific expression is as follows:

[0026] pm =W m ([f CLIP-I ,f CLIP-T ]) (1)

[0027] Among them, W m Represents training parameters.

[0028] The specific implementation process of step 5 is as follows:

[0029] 5-1. Settings Represents the multimodal fusion vector.

[0030] 5-2. If Figure 2 As shown, the embedding vector Input into the stacked progressive Transformer, and fuse multiple vectors according to the following formula:

[0031]

[0032] Where Trans_encoder represents the encoder module in Transformer. l represents the stacking level, l∈[0,N-1], and formulas (2)-(9) are executed sequentially.

[0033] 5-3. After stacking and passing it through the Transformer for N times, the obtained Set to p n .

[0034] The specific implementation process of step 6 is as follows:

[0035] 6-1. Multimodal fusion vector p n Input classifier, which consists of an MLP network and a sigmoid output layer. Predicted label

[0036]

[0037] 6-2. We define all parameters in the present invention as Θ, Γ represents the training set. K represents the number of training sets, y represents the true label in the training set, Represents the predicted label, and the model is trained using the cross entropy loss function. The false news detection of the present invention belongs to a binary classification problem, and the objective function calculation formula is as follows:

[0038]

[0039] The present invention utilizes the Adam optimizer to perform minimization learning on the above objective function.

[0040] 6-3. After learning, use the model parameters Θ to detect the messages that need to be detected and output the detection results.

[0041] The beneficial effects of the present invention are as follows:

[0042] The present invention detects false news by integrating text, images, text-image alignment characteristics and background knowledge assisted by a large model through multimodal fusion and large language model technology. The present invention not only improves the accuracy of false news detection from the perspective of multimodal fusion, but also effectively utilizes the additional background knowledge generated by the large language model to assist in false news detection. The present invention designs a stacked progressive Transformer model to effectively fuse different modal information. Given a specific message, the present invention can analyze the embedded features of text and images, as well as external background knowledge, to determine whether the message is false.

[0043] The spread of false information will undermine the public's trust in the source of information. The present invention can help the public improve their ability to identify false information, enhance society's immunity to false information, and reduce its spread and impact. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 False news detection in social networks based on multimodal fusion and large language model.

[0045] Figure 2 This is a diagram of the N-layer stacked progressive Transformer framework. DETAILED DESCRIPTION

[0046] The present invention will be further described below in conjunction with the accompanying drawings.

[0047] like Figure 1 As shown in the figure, the social network false information detection method based on large model and multimodal fusion includes the following steps:

[0048] Step 1: Given a multimodal message, the message is represented as a text-image pair. I represents image and T represents text. For image I, the embedding vector of the image is extracted by Masked Autoencoder (MAE) to obtain p i ; For image I, extract the image alignment embedding f through the image encoder in the CLIP (Contrastive Language-Image Pre-Training) pre-training model CLIP-I .

[0049] Step 2: For the text T, extract the text embedding vector p through the BERT model j; For the text T, extract the text alignment embedding f through the text encoder in the CLIP pre-trained model CLIP-T .

[0050] Step 3: Use the ChatGLM model to design a prompt. Input the false message text T into ChatGLM, and output the relevant background knowledge T' of the false message T. Input T' into BERT to obtain the embedding vector p of the background knowledge. k .

[0051] Step 4: Align the image embedding f CLIP-I Alignment embedding with text f CLIP-T Concatenate and input into the fully connected layer to obtain the multimodal mixed alignment feature p m .

[0052] Step 5: vector p i , p j , p k and p m Input N layers of stacked progressive Transformer model, output multimodal fusion embedding vector p n .

[0053] Step 6: Multimodal fusion vector p n Input Sigmoid classifier and set cross entropy loss function The model is trained using the Adam optimization method. After learning, the model parameters are used to detect the message (text-image pair) that needs to be predicted, and the detection result of whether it is a false message is output.

[0054] The specific implementation process of step 1 is as follows:

[0055] 1-1. Given a multimodal message I stands for image and T stands for text.

[0056] 1-2. Scale the image I to 224*224 and use the MAE module to get the embedding vector p of image I i In the specific implementation, the patch value in the MAE model is set to 16 and the embedding dimension of the image is set to 768.

[0057] 1-3. Scale the image I to 224*224 and use the image-encoder module in the pre-trained CLIP model to get the aligned embedding f of image I CLIP-I In the specific implementation, the alignment embedding dimension is set to 512.

[0058] The specific implementation process of step 2 is as follows:

[0059] 2-1. Use the Chinese Bert pre-training model to extract the embedding vector of each word in the text T, and then use the Meanpooling method to output the mean of each word as the sentence vector of the text T. The present invention represents the embedding vector of the text T as p i = BERT(T). In the specific implementation, the BERT model uses the Chinese-bert-wwm-ext version, and p i The dimension is set to 768.

[0060] 2-2. Use the text-encoder module in the pre-trained CLIP model to get the aligned embedding f of the text T CLIP-T In specific implementation, f CLIP-T The dimension is set to 512.

[0061] The specific implementation process of step 3 is as follows:

[0062] 3-1. Based on ChatG LM The specific design is as follows: "You are designed as an auxiliary tool for a fake news detection system. Your main function is to provide an explanation of relevant background knowledge based on a given message. Your goal is to output a concise piece of relevant background knowledge so that people can assist in judging the authenticity of the message when reading it without the need for additional background knowledge. The length of your explanation should be less than 300 words." In terms of specific implementation, ChatGLM chooses the ChatGlm-4-Air version.

[0063] 3-2. Enter the message text T to ChatG LM In the example below, the relevant background knowledge T' of the message T is output. In the specific implementation, a message text-image pair of the following form is input, and the corresponding background knowledge explanation is output, as shown in the following example:

[0064]

[0065] 3-3. Use the Chinese Bert pre-trained model to extract the embedding vector of each word in T', and then use the Meanpooling method to output the mean of each word as the embedding vector of text T', specifically expressed as p k = BERT(T'). In the specific implementation, the BERT model uses the Chinese-bert-wwm-ext version, and p k The dimension is set to 768.

[0066] The specific implementation process of step 4 is as follows:

[0067] 4-1. Alignment of images into f CLIP-I Alignment embedding with text f CLIP-TConcatenate and input into the fully connected layer to obtain the multimodal mixed feature p m The specific expression is as follows:

[0068] p m =W m ([f CLIP-I ,f CLIP-T ]) (1)

[0069] Among them, W m Represents the training parameters. In specific implementation, p m The dimension of is set to 768.

[0070] The specific implementation process of step 5 is as follows:

[0071] 5-1. Settings Represents the multimodal fusion vector.

[0072] 5-2. If Figure 2 As shown, the embedding vector Input into the stacked progressive Transformer, and fuse multiple vectors according to the following formula:

[0073]

[0074] Where Trans_encoder represents the encoder module in Transformer. l represents the stacking level, l∈[0,N-1], and formulas (2)-(9) are executed sequentially.

[0075] 5-3. After stacking and passing the Transformer N times, the obtained Set to p n In the specific implementation, N is set to 4.

[0076] The specific implementation process of step 6 is as follows:

[0077] 6-1. Multimodal fusion vector p n Input classifier, which consists of an MLP network and a sigmoid output layer. Predicted label

[0078]

[0079] In the specific implementation, the number of layers of the MLP network is set to 768, 384, and 1.

[0080] 6-2. We define all parameters in the present invention as Θ, Γ represents the training set. K represents the number of training sets, y represents the true label in the training set, Represents the predicted label, and the model is trained using the cross entropy loss function. The false news detection of the present invention belongs to a binary classification problem, and the objective function calculation formula is as follows:

[0081]

[0082] The present invention uses the Adam optimizer to minimize the objective function. In specific implementation, the learning rate is set to 2e-5, the momentum parameters beta1 and beta2 are set to 0.9 and 0.999 respectively, and epsilon is set to the default value.

[0083] 6-3. After learning, use the model parameters Θ to detect the message to be detected and output the detection results. In specific implementation, given a new message to be detected (including text-image pairs), first, extract the embedding vectors of each modality based on the text and image in the message, p i , p j , p k and p m ; Then, vector p i , p j , p k and p m Input N layers of stacked progressive Transformer model to get p n Finally, according to formula (11) The value of .

[0084] Experimental data:

[0085] The present invention collects a database of false news mainly based on the Sina platform. Each message contains a text-image pair. The database information is shown in the following table:

[0086] Table 1 Information of Sina Weibo false news database

[0087] database Number of texts Number of images Number of false news 2970 2970 Normal message count 3124 3124 total 6094 6094

[0088] The positive-negative sample ratio of the database is set to 1.05:1. Using the above information, the text, image, text-image alignment embedding, and background knowledge generated by the large model are analyzed. The present invention is compared with the unimodal BERT method and the multimodal method. In the BERT method, the representation of [CLS] is used to determine whether it is a false message. The multimodal benchmark models are the CLIP model, the CAFE model, and the Spotfake model. In the CLIP model, the image and text of a specific message are input into CLIP to obtain a joint representation of the visual and textual content, and then this joint representation vector is used to predict whether it is a false message. The CAFE model proposed by Yixuan Chen et al. is a fake news detection method with ambiguity perception. Specifically, CAFE integrates the text unimodal features generated by BERT and the image unimodal features generated by ResNet, and combines cross-modal correlation. When the cross-modal ambiguity is weak, it relies on unimodal features, and when the cross-modal ambiguity is strong, it relies on cross-modal correlation. The Spotfake model proposed by Shivangi Singhal et al. uses BERT and VGG to generate text and images respectively, and fuses text and image features for message detection. The present invention uses the two indicators of Accuracy and F1-score to analyze the experimental results. The specific comparison results are shown in Table 2.

[0089] Table 2 Comparison of experimental results

[0090] method Accuracy F1-score BERT 0.873 0.872 CLIP 0.883 0.886 CAFE 0.882 0.885 SpotFake 0.851 0.828 The present invention 0.919 0.919

[0091] It can be seen from the data in this table that the method proposed in the present invention is better than other unimodal and multimodal methods. Compared with the unimodal BERT method, the present invention is 4.7% higher in F1-score and 4.6% higher in Accuracy. Compared with the multimodal method SpotFake, the present invention is 9.1% higher in F1-score and 6.8% higher in Accuracy. In addition, the present invention has obvious improvements over both CLIP and CAFE. The present invention believes that the main reason for the performance improvement is that the present invention not only takes text and images into consideration, but also the designed stacked progressive Transformer model can effectively fuse multimodal information. In addition, the present invention incorporates the background knowledge of the large language model, which effectively expands the semantic content of the background knowledge of false messages.

[0092] In addition, the present invention conducted an ablation experiment. We removed the stacked progressive Transformer module from the method proposed in the present invention and marked it as R1; at the same time, we removed the large language model module from the method proposed in the present invention and marked it as R2. In the R1 method, we replaced the stacked progressive Transformer model by performing a simple splicing operation on different features. We analyzed the impact of the stacked progressive Transformer module and the large language model module on the experimental results, and the specific effects are shown in Table 3. It can be seen from the data in Table 3 that if the stacked progressive Transformer module or the large language model module is removed, the final effect is not as good as that of the present invention. It can be seen that the stacked progressive Transformer module or the large language model module designed in the present invention has a certain improvement in the detection performance.

[0093] Table 3. Comparison of ablation experiments

[0094] method Accuracy F1-score AUC R1 0.907 0.907 0.974 R2 0.911 0.908 0.970 The present invention 0.919 0.919 0.975

[0095] In summary, the present invention detects false news by integrating text, image, text-image alignment characteristics and background knowledge assisted by a large model through multimodal fusion and large language model technology. The present invention not only improves the accuracy of false news detection from the perspective of multimodal fusion, but also effectively utilizes the additional background knowledge generated by the large language model to assist in false news detection. The stacked progressive Transformer model designed by the present invention effectively fuses different modal information, which has a certain improvement effect on detection performance.

Claims

1. A method for detecting fake news on social networks based on a large model and multimodal fusion, characterized by It includes the following steps: Step 1: Given a multimodal message, the message is represented as a text-image pair. I stands for image, T stands for text; For image I, the masked autoencoder MAE is used to extract the image embedding vector p i ; For image I, extract the image alignment embedding f through the image encoder in the CLIP pre-trained model CLIP-I ; Step 2: For the text T, extract the text embedding vector p through the BERT model j ; For the text T, extract the text alignment embedding f through the text encoder in the CLIP pre-trained model CLIP-T ; Step 3: Use the ChatGLM model to design a prompt; input the false message text T into ChatGLM, and output the relevant background knowledge T' of the false message T; input T' into the BERT model to obtain the embedding vector p of the background knowledge k ; Step 4: Align the image into f CLIP-I Alignment embedding with text f CLIP-T Concatenate and input into the fully connected layer to obtain the multimodal mixed alignment feature p m ; Step 5: vector p i 、p j 、p k and p m Input N layers of stacked progressive Transformer model, output multimodal fusion embedding vector p n ; Step 6: Embedding vector p after multimodal fusion n Input Sigmoid classifier and set cross entropy loss function The model is trained using the Adam optimization method; After learning, the model parameters are used to detect the message to be predicted and output the detection result of whether it is false news.

2. The method for detecting fake messages on social networks based on a large model and multimodal fusion according to claim 1 is characterized in that: The specific implementation process of step 1 is as follows: 1-1. Given a multimodal message I stands for image, T stands for text; 1-2. Scale the image I to 224*224 and use the MAE module to get the embedding vector p of image I i ; 1-3. Scale the image I to 224*224 and use the image-encoder module in the pre-trained CLIP model to get the aligned embedding f of image I CLIP-I .

3. The method for detecting fake messages on social networks based on large models and multimodal fusion according to claim 1 is characterized in that: The specific implementation process of step 2 is as follows: 2-1. Use the Chinese Bert pre-trained model to extract the embedding vector of each word in the text T, and then use the Meanpooling method to output the mean of each word as the sentence vector of the text T; the embedding vector of the text T is represented as p j =BERT(T); 2-2. Use the text-encoder module in the pre-trained CLIP model to get the aligned embedding f of the text T CLIP-T .

4. The method for detecting fake messages on social networks based on large models and multimodal fusion according to claim 2 is characterized in that: The specific implementation process of step 3 is as follows: 3-1. Based on ChatG LM Large model design prompt; 3-2. Enter the message text T to ChatG LM In the output, the relevant background knowledge T' of the message T is output; 3-3. Use the Chinese Bert pre-trained model to extract the embedding vector of each word in T', and then use the Meanpooling method to output the mean of each word as the embedding vector of text T', specifically expressed as p k =BERT(T').

5. The method for detecting fake messages on social networks based on large models and multimodal fusion according to claim 4 is characterized in that: The prompt designed in step 3-1 is as follows: "You are designed to be an auxiliary tool for ****; your main function is ****; your goal is to output ****; the length of your explanation should be less than *** words." 6. The method for detecting fake messages on social networks based on large models and multimodal fusion according to claim 1, characterized in that: The specific implementation process of step 4 is as follows: 4-1. Alignment of images into f CLIP-I Alignment embedding with text f CLIP-T Concatenate and input into the fully connected layer to obtain the multimodal mixed feature p m ; The specific expression is as follows: p m =W m ([f CLIP-I ,f CLIP-T ]) (1) Among them, W m Represents training parameters.

7. The method for detecting fake messages on social networks based on large models and multimodal fusion according to claim 1, characterized in that: The specific implementation process of step 5 is as follows: 5-1. Settings Represents the multimodal fusion vector; 5-2. Embedding vector Input into the stacked progressive Transformer, and fuse multiple vectors according to the following formula: Where Trans_encoder represents the encoder module in Transformer; l represents the stacking level, l∈[0,N-1], and formulas (2)-(9) are executed sequentially; 5-3. After stacking and passing it through the Transformer for N times, the obtained Set to p n .

8. The method for detecting fake messages on social networks based on large models and multimodal fusion according to claim 1, characterized in that: The specific implementation process of step 6 is as follows: 6-1. Multimodal fusion vector p n Input classifier, which consists of an MLP network and a sigmoid output layer; predicted label 6-2. We define all parameters in the present invention as Θ, Γ represents the training set. K represents the number of training sets, y represents the true label in the training set, Represents the predicted label, and the model is trained using the cross entropy loss function; false news detection belongs to a binary classification problem, so the objective function calculation formula is as follows: 6-3. After learning, use the model parameters Θ to detect the messages that need to be detected and output the detection results.

Images