False information detection method and device based on plug-and-play anti-fact depolarization framework

Abstract

The invention relates to a false information detection method and device based on a plug-and-play anti-fact depolarization framework. The plug-and-play anti-fact framework proposed by the method can be integrated with various traditional detection models to form a flexible depolarization module. The plug-and-play anti-fact framework adopts anti-fact reasoning to eliminate false correlation among declarations, evidences and prediction labels, so that the causal reasoning capability of the model is enhanced; false correlation among declarations, evidences and prediction labels is eliminated by using anti-factual reasoning, so that the causal reasoning capability of a detection model is enhanced; and the detection accuracy of various basic detection models on false information detection is obviously improved.

Description

Technical Field

[0001] This application relates to the technical field of false information detection, and particularly to a false information detection method and device based on a plug-and-play counterfactual debiasing framework. Background Art

[0002] The large emergence of electronic products has prompted the way people obtain information to gradually shift from printed media to digital media. This shift has lowered the threshold of information dissemination, accelerated the speed of information dissemination, and imperceptibly influenced people's concepts and value systems. However, due to the inability to verify the large amount of information spread on online platforms in a timely and effective manner, social networks have gradually become a hotbed for fake news. Fake news is unfounded and fabricated information, usually lacking factual support and being unverified, full of inaccuracies, or deliberately deceptive. They may include fabricated facts, exaggerated descriptions, and misleading contexts. Fake news may pose a serious threat to social order and may also trigger widespread panic among the public. Therefore, it is urgent to take effective measures to curb the spread of fake news to maintain social harmony and stability.

[0003] In recent years, the task of fake news detection has been widely studied from different perspectives. Most fake news detection methods can be roughly divided into two categories: pattern-based methods and evidence-aware methods. Pattern-based methods rely only on the claim itself and do not involve any other external information. To detect fake news, different methods may emphasize different features, such as text content, multimodal content, and dissemination structure.

[0004] In contrast, evidence-aware methods introduce relevant external information (such as evidence) to assist in detecting fake news. They usually judge the authenticity of news by analyzing the correlation between evidence and the claim. This method involves extracting keywords from the claim and searching for highly relevant terms in external sources, using these keywords to assist in determining the authenticity of the news. Based on evidence, evidence-aware fake news detection models show higher reliability and interpretability compared to other types of models.

[0005] Different from direct text classification tasks, evidence-aware fake news detection methods need to understand the context and identify information that needs to be verified. After collecting evidence, causal reasoning will be used to verify the information. For example, the input data includes not only the text claim but also the top 30 relevant pieces of evidence. The claim will be input into a deep learning network (such as TextCNN) to extract features, and the key information that needs to be checked will be identified through an attention mechanism. Finally, by concatenating and mixing the feature vectors of the evidence and the claim, an accurate modeling of causal relevance is achieved. This feature fusion method can fully consider the interaction between the evidence and the claim, thus more accurately judging the authenticity of the news. In addition, this method also focuses on the interpretability of the detection process. By recording the process of evidence retrieval and verification and the specific way of feature fusion, the model can provide detailed explanations for the detection results, thereby improving reliability. This interpretability can not only help users understand the basis of the prediction but also provide a reference for subsequent model optimization and improvement.

[0006] In addition, in the actual deployment of fake news detection systems, a significant challenge is that the training and test datasets usually have different data distributions, which is the so-called Out-of-Distribution (OOD) problem. In practice, fake news detection models are usually trained only on limited platform data, but their applications are oriented towards various different general platforms. At the same time, the training dataset is usually limited to data on specific topics or events within a short time range, while the model needs to generalize to a wider range of scenarios.

[0007] Given that the labels and relevant evidence of the training dataset are manually collected, evidence-aware fake news detection models are prone to learning spurious correlations between the claim and the predicted label. In addition, due to the correlation between the claim and the evidence, spurious correlations may also occur between the evidence and the predicted label. These spurious correlations pose significant out-of-distribution challenges to fake news detection models. The above spurious correlations can be defined as claim bias and evidence bias respectively. Both biases hinder the fake news detection model from effectively performing causal reasoning between the claim and the evidence. For example, in the training dataset, if most events in a certain field are mostly labeled as "fake news", the model may learn that the topics and keywords in that field indicate fake news instead of performing causal reasoning. As expected, when the data distribution changes, this correlation does not hold.

[0008] To improve the accuracy and generalization ability of the fake news detection model, it is necessary to mitigate the influence of claim bias and evidence bias, and enhance the causal reasoning ability of the detection model. However, the out-of-distribution problem in fake news detection has not been fully studied, especially for evidence-aware tasks. Some methods are evidence-based, but only concatenate claim and evidence features without true causal reasoning. Subsequently, different attention mechanisms have been applied to simulate the causal reasoning ability of the human brain. However, these methods mainly focus on identifying common features between different events or topics, but fail to consider the specific reasoning paths required in evidence-aware fake news detection. At the same time, different debiasing methods have been proposed to address this issue. ReW reduces the model's dependence on statistical features by decreasing the weights of high-frequency n-gram samples highly correlated with the label. However, it also ignores the patterns of bias features and may introduce new data imbalance. PoE uses the prediction results of the "bias expert model" to suppress the influence of bias features on the main model. However, the independence assumption between the bias model and the main model in PoE is too strong, and dynamic weighting may lead to unstable training. CF eliminates the influence of bias features through counterfactual reasoning, explicitly decoupling bias features, but the computational complexity of the CF architecture significantly increases the computational cost and training difficulty. Summary of the Invention

[0009] Based on this, in view of the above technical problems, it is necessary to provide a fake information detection method and device based on a plug-and-play counterfactual debiasing framework.

[0010] A fake information detection method based on a plug-and-play counterfactual debiasing framework, the method comprising: Obtain a text information dataset to be detected, where each data information includes: a claim set, a corresponding evidence set, and a true information label.

[0011] Use a text embedding module to convert the claim set and the corresponding evidence set into a claim embedding feature vector and an evidence embedding feature vector.

[0012] According to the claim embedding feature vector and the evidence embedding feature vector, use a traditional evidence-aware fake information detection model to make a prediction, and obtain a traditional prediction result.

[0013] According to the claim embedding feature vector and the evidence embedding feature vector, use a debiasing model for processing to obtain a claim-based prediction result and an evidence-based prediction result; the debiasing model uses a counterfactual reasoning debiasing method.

[0014] Perform weighted summation on the claim-based prediction result and the evidence-based prediction result to obtain a prediction result under counterfactual conditions.

[0015] The feature fusion module is used to fuse the traditional prediction results and the prediction results under counterfactual conditions to obtain the debiased prediction results of false information.

[0016] A false information detection device based on a plug-and-play counterfactual debiasing framework, the device includes: A text information dataset acquisition unit to be detected, used to acquire a text information dataset to be detected, and each data information includes: a statement set, a corresponding evidence set, and a true information label.

[0017] A feature embedding unit, used to convert the statement set and the corresponding evidence set into a statement embedding feature vector and an evidence embedding feature vector by using a text embedding module.

[0018] A traditional prediction unit, used to make a prediction by using a traditional evidence-aware false information detection model according to the statement embedding feature vector and the evidence embedding feature vector to obtain a traditional prediction result.

[0019] A prediction unit under counterfactual conditions, used to process the statement embedding feature vector and the evidence embedding feature vector by using a debiasing model to obtain a statement-based prediction result and an evidence-based prediction result; the debiasing model adopts a counterfactual reasoning debiasing method; the statement-based prediction result and the evidence-based prediction result are weighted and summed to obtain a prediction result under counterfactual conditions.

[0020] A false information debiased prediction result determination unit, used to fuse the traditional prediction result and the prediction result under counterfactual conditions by using a feature fusion module to obtain a false information debiased prediction result.

[0021] The above-mentioned false information detection method and device based on a plug-and-play counterfactual debiasing framework, the plug-and-play counterfactual framework proposed by the method can be a flexible debiasing module integrated with various traditional detection models. The plug-and-play counterfactual framework uses counterfactual reasoning to eliminate the spurious correlations between statements, evidence, and prediction labels, thereby enhancing the causal reasoning ability of the model; using counterfactual reasoning to eliminate the spurious correlations between statements, evidence, and prediction labels, thereby enhancing the causal reasoning ability of the detection model; significantly improving the detection accuracy of various basic detection models for false information detection. Description of the Drawings

[0022] Figure 1 It is a causal graph of evidence-aware false information detection in an embodiment; Figure 2 It is a schematic flowchart of a false information detection method based on a plug-and-play counterfactual debiasing framework in an embodiment; Figure 3 It is an overview diagram of the DP2CF debiasing method in another embodiment; Figure 4Flow chart of an evidence-based misinformation detection model in another embodiment; Figure 5 General process of the DP2CF model in another embodiment; Figure 6 Sensitivity analysis comparison chart of DP2CF in another embodiment, where (a) is the sensitivity analysis comparison chart of BERT tested on Snopes, (b) is the sensitivity analysis comparison chart of BERT tested on Politifact, (c) is the sensitivity analysis comparison chart of DeClare tested on Snopes, (d) is the sensitivity analysis comparison chart of DeClare tested on Politifact, (e) is the sensitivity analysis comparison chart of MAC tested on Snopes, and (f) is the sensitivity analysis comparison chart of MAC tested on Politifact. Detailed implementation manners

[0023] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

[0024] Misinformation detection method based on a plug-and-play counterfactual debiasing framework, abbreviated as: DP2CF debiasing method.

[0025] In the present application, denotes a dataset to be detected with a length of k, where each news is composed of a corresponding claim , the corresponding evidence set , and the true news label . For each piece of evidence from the evidence set , they are similar in composition to the claim and are composed of words . The constituent elements of represent a single word. Considering that misinformation detection is a binary classification problem, its true label , where 0 and 1 represent fake news and real news respectively. The goal of misinformation is to develop a classification model to effectively calculate the predicted label of the test news , based on the input claim and the evidence set .

[0026] The input dataset of news to be verified will be directly obtained before the fake news detection task. From the perspective of causal inference, the authenticity of news should be verified through causal reasoning between claims and evidence. After the interaction between evidence and claims, a potential embedding vector is generated, denoted as the interaction feature. . To eliminate the specific biases between claims, evidence, and prediction labels, causal graphs are used to study the relationships among them. Based on the insights obtained from the causal graphs, appropriate debiasing strategies can be designed. The causal graph of the evidence-aware fake news detection method is shown in Figure 1 . In Figure 1 , "×" means truncation.

[0027] In addition, each news is composed of a claim and an evidence set . The researchers intend to make predictions based on the interaction feature. . However, in the causal graph of traditional methods, there are spurious correlations that interfere with causal inference. Since the original dataset is manually collected, specific keywords or fields of may have spurious correlations with the prediction label, denoted as . Since there is a strong correlation between and , may share the same keywords or specific fields with , denoted as . Therefore, there may also be spurious correlations between and , which can be denoted as . During the training phase, when using traditional fake news detection methods, the model may learn from spurious correlations, resulting in . Therefore, it is necessary to use counterfactual reasoning to mitigate the effects of and .

[0028] In one embodiment, as shown in Figure 2 , a fake information detection method based on a plug-and-play counterfactual debiasing framework is provided, which includes the following steps: Step 100: Obtain a dataset of text information to be detected, where each data information includes: a claim set, a corresponding evidence set, and a true information label.

[0029] Specifically, in the field of natural language processing (NLP), word embedding techniques play a key role in capturing semantic relationships between words. The core challenge in the NLP field is to efficiently represent the semantics of words. Based on existing fake news detection methods, GloVe and BERT models are mainly used as embedding modules.

[0030] Step 102: Use a text embedding module to convert the claim set and the corresponding evidence set into a claim embedding feature vector and an evidence embedding feature vector.

[0031] Specifically, the text embedding module converts the claims and evidence in the original text data into embedding feature vectors.

[0032] Step 104: Use a traditional evidence-aware misinformation detection model to make predictions based on the claim embedding feature vector and the evidence embedding feature vector, and obtain a traditional prediction result.

[0033] Step 106: Process the claim embedding feature vector and the evidence embedding feature vector using a debiasing model to obtain a claim-based prediction result and an evidence-based prediction result; the debiasing model uses a counterfactual reasoning debiasing method.

[0034] Specifically, due to data bias in the original dataset, traditional detection models may learn spurious correlations between specific fields or keywords and prediction labels. To mitigate this bias, counterfactual reasoning is used as a debiasing method. Counterfactual reasoning refers to the process of inferring potential outcomes under the assumption that certain variables are disabled, thereby evaluating the contribution of the remaining variables to the prediction result. In the field of fake news detection, counterfactual reasoning requires the detection model to make predictions based only on the claim or evidence.

[0035] Step 108: Perform a weighted sum of the claim-based prediction result and the evidence-based prediction result to obtain a prediction result under counterfactual conditions.

[0036] Step 110: Use a feature fusion module to fuse the traditional prediction result and the prediction result under counterfactual conditions to obtain a debiased misinformation prediction result.

[0037] Specifically, the feature fusion module fuses different feature vectors to obtain a final debiased prediction result.

[0038] The feature fusion module aims to enhance accuracy, stability, and generalization ability by integrating multi-dimensional data features. This module uses the fuzzy logic principle to map each output feature vector to a fuzzy membership space. Therefore, this method more effectively captures the inherent features of the underlying patterns while reducing the impact of noise.

[0039] The false information detection method based on the plug-and-play counterfactual debiasing framework proposed in this application is a novel counterfactual debiasing method, namely the "Plug-and-Play Counterfactual Framework (DP2CF)". Although previous studies have applied CF in the field of fake news detection, some of these studies assume that the average features of evidence are essential when calculating counterfactual predictions, rather than relying only on the claim without external sources. In contrast, DP2CF, as a flexible plug-and-play module, can be easily attached to various traditional evidence-aware fake news detection models and avoid the above strong assumptions. Specifically: 1) For the embedded representation of claims or evidence, a multi-scale convolutional feature extraction module is used for feature extraction. The one-dimensional convolutional layer is used to convolve the input embedded sequence, and then the convolutional feature map is pooled into a fixed-length feature vector through the adaptive max-pooling layer. Finally, the feature vectors of different sizes are concatenated to obtain a comprehensive feature representation. 2) To enhance the accuracy and adaptability of the detection model when testing out-of-distribution datasets, a gradient reversal layer is introduced for feature processing. 3) To eliminate spurious correlations, hyperparameters α and β are introduced to adjust the calculation process. Since counterfactual reasoning requires subtracting the logical value of the traditional detection model from the logical value in the counterfactual situation, the hyperparameters may help balance this process, thus ensuring that the detection model retains important semantic information while removing spurious correlations.

[0040] Different from existing methods that may be restricted by strong assumptions or high computational complexity, the DP2CF framework in this application is a flexible plug-and-play module that can be easily integrated into various traditional detection models, providing a more efficient and effective debiasing solution and reducing the impact of claim bias and evidence bias in evidence-aware fake news detection.

[0041] The overall structure of the DP2CF debiasing method is as Figure 3 shown. This framework consists of three modules. The first module is the text embedding module, which converts the claims and evidence in the original text data into embedded feature vectors. The second module is the feature processing module, which includes a traditional detection model and a debiasing model for claims and evidence. In addition to calculating the traditional prediction results, counterfactual predictions can also be obtained after this process. The last module is the feature fusion module, in which different feature vectors are fused to obtain the final debiased prediction result.

[0042] In the above false information detection method based on the plug-and-play counterfactual debiasing framework, for the false information detection method and device based on the plug-and-play counterfactual debiasing framework, the proposed plug-and-play counterfactual framework of the method can be a flexible debiasing module integrated with various traditional detection models. The plug-and-play counterfactual framework uses counterfactual reasoning to eliminate the spurious correlations between claims, evidence, and prediction labels, thereby enhancing the causal reasoning ability of the model; using counterfactual reasoning to eliminate the spurious correlations between claims, evidence, and prediction labels, thus enhancing the causal reasoning ability of the detection model; significantly improving the detection accuracy of various basic detection models for false information detection.

[0043] In one embodiment, if the BERT model is selected as the basis for the traditional evidence-aware false information detection model, then step 102 selects the BERT model as the text embedding module.

[0044] Specifically, BERT is a novel pre-training method for language representation that has achieved excellent performance in a wide range of natural language processing tasks. The main contributions of BERT are bidirectional context representation and the pre-training and fine-tuning paradigm. BERT is based on the Transformer architecture and specifically utilizes the encoder part of the Transformer. The encoder structure of BERT enables it to capture bidirectional context, which means that each token in the input sequence can attend to all other tokens during the encoding process, whether before or after it.

[0045] The input of BERT consists of three types of embeddings: (1) Token Embeddings: Each token in the input sequence is mapped to a vector in a high-dimensional space through the WordPiece tokenization method. (2) Segment Embeddings: These embeddings are used to distinguish different sentences in the input. For example, in a pair of sentences A and B, the tokens in sentence A are assigned a segment ID of 0, while the tokens in sentence B are assigned a segment ID of 1. (3) Position Embeddings: These embeddings provide the position information of each token in the sequence, and the position embeddings of BERT are learnable. The final input representation is obtained by adding these three embeddings for each token.

[0046] The basic architecture of BERT relies on a bidirectional Transformer encoder. The input text is processed sequentially through a series of sub-layers, including multi-head self-attention layers, fully connected feed-forward layers, etc. Each sub-layer integrates normalization and residual connections, and the output of each sub-layer is calculated according to the following formula.

[0047] ; Among them, sublayer represents the specific operations performed by each sublayer. In the self-attention module, the query matrix (q), key matrix (k), and value matrix (v) are used to evaluate the similarity between words and assign weights accordingly. This process helps to aggregate the relevant information between words, enabling the model to capture the dependencies between different positions. The calculation formula of this attention layer is as follows: ; where represents the dimension of the query and key matrices. To improve the performance of the model, the multi-head attention mechanism is usually implemented. In this method, the input is linearly transformed into q, k, and v matrices. After generating multiple attention matrices, they are horizontally concatenated and then multiplied by a weight matrix to produce a unified matrix. The specific calculation process is as follows: ; ; where the weight vectors , and correspond to the query q, key k, and value v matrices of the i-th attention head respectively. The operation Cocat refers to the horizontal concatenation of the outputs of multiple attention heads, and represents the weight matrix applied in the concatenation step. After the multi-head attention layer, its output is passed to a fully connected feed-forward network, which processes the data through several activation functions to generate the final output, as follows: ; where RELU represents the Rectified Linear Unit activation function, and and represent the weights and bias parameters of the multi-head attention layer respectively, and correspond to the weights and bias of the fully connected feed-forward network. In this application, the hidden state vector extracted from the last layer of the BERT model is used as the text embedding vector. In an application, if the BERT model is selected as the basis of a traditional detection model, this method will be used as the text embedding module.

[0048] In one of the embodiments, when the traditional evidence-aware false information detection model is MAC or DeClare, then in step 102, the GloVe model is selected as the text embedding module.

[0049] Specifically, GloVe (Global Vectors for Word Representation) is a powerful word embedding method that combines the advantages of both count-based and prediction-based methods. By leveraging global co-occurrence statistics, the word vectors generated by GloVe can preserve meaningful linear substructures in the vector space. The core idea of GloVe is to capture the semantic relationships between words through a co-occurrence matrix, which records the co-occurrence frequencies of words in a large corpus. The objective function of GloVe is formulated as: ; where is the size of the vocabulary, denotes the co-occurrence frequency of word and word . and are the word vectors of word and word respectively, and are the additional bias terms of word and word . is a weight function that adjusts the impact of co-occurrence frequencies on the loss function, defined as: ; where is a threshold used to prevent overly high co-occurrence frequencies from having an adverse impact on the model during the training phase. And is a hyperparameter. When is set to 0.75, the model will achieve the best performance. The main advantages of GloVe are its ability to capture global context, scale to large vocabularies, and preserve linear structures in the vector space. These features make GloVe a perfect tool for natural language processing tasks, such as detecting fake news. If DP2CF (assumed to be a detection framework or method) selects MAC or DeClare as its traditional detection model, then the GloVe method will operate as a text embedding module.

[0050] In one embodiment, the debiasing model includes: a claim debiasing model and an evidence debiasing model; the claim debiasing model is used to extract the relationship between the claim embedding feature vector and the predicted label to obtain a claim-based prediction result; the evidence debiasing model is used to extract the relationship between the evidence embedding feature vector and the predicted label to obtain an evidence-based prediction result; step 106 includes: Inputting the claim embedding feature vector into the claim debiasing model to obtain a claim-based prediction result; inputting the evidence embedding feature vector into the evidence debiasing model to obtain an evidence-based prediction result.

[0051] In one embodiment, the claim debiasing model includes: a TextCNN module, a max pooling layer, a gradient reversal layer, and a fully feed-forward network layer; in the claim debiasing model: the claim embedding feature vector is subjected to feature extraction using the TextCNN module to obtain a convolutional vector; the convolutional vector is activated by the ReLU function and then processed by the max pooling layer to obtain |K| first intermediate feature vectors; where K is a set of kernels for convolution; all intermediate feature vectors are concatenated through the Cocat operation to obtain a concatenated vector; the concatenated vector is processed through the gradient reversal layer and then input into the fully feed-forward network layer to obtain a claim-based prediction result.

[0052] In one embodiment, the evidence debiasing model has the same structure as the claim debiasing model; in the evidence debiasing model: the evidence embedding feature vector is subjected to feature extraction using the TextCNN module to obtain a first convolutional vector; the first convolutional vector is activated by the ReLU function and then processed by the max pooling layer to obtain |K| first intermediate feature vectors; where K is a set of kernels for convolution; all first intermediate feature vectors are concatenated through the Cocat operation to obtain a first concatenated vector; the first concatenated vector is processed through the gradient reversal layer and then input into the fully feed-forward network layer to obtain an evidence-based prediction result.

[0053] In one embodiment, a feature fusion module is used to fuse the traditional prediction result and the prediction result under counterfactual conditions to obtain a false information debiasing prediction result. The expression of the false information debiasing prediction result is: ; where, represents the false information debiasing prediction result, represents the prediction result of the traditional detection model, represents the prediction result under counterfactual conditions.

[0054] Specifically, for the method for detecting false information based on evidence, its general flowchart is as Figure 4 shown.

[0055] (1) Traditional detection model: Given a claim and its corresponding evidence dataset , the detection model first encodes the text content into word-level embedding vectors. The calculation process is described as follows: ; ; where, and are the word-level embedding matrices of the claim and the evidence respectively. In addition, denotes the dimension of word-level embeddings, while denotes the learnable parameters of the embedding module. Subsequently, the detection model conducts word-level interactions between the claim and the evidence to generate sentence-level embeddings. Sentence-level embeddings are typically obtained through various pooling operations, as follows: ; ; where and . While denotes the dimension of the sentence-level embeddings. The learnable parameters of the word-level interaction module are denoted as . Finally, the detection model conducts sentence-level interactions between the claim and the evidence to obtain the prediction output . The specific formula for this process can be expressed as follows: ; If the input data includes the publisher embedding vector and the speaker embedding vector to represent the credibility of the claim and the evidence respectively, then the above formula can be rewritten as: ; (2) Debiased Model Due to data bias in the original dataset, traditional detection models may learn spurious correlations between specific fields or keywords and prediction labels. To mitigate this bias, counterfactual reasoning is adopted as a debiasing method. Counterfactual reasoning refers to the process of inferring potential outcomes by assuming the failure of certain variables, thereby evaluating the contribution of the remaining variables to the prediction result. In the field of fake news detection, counterfactual reasoning requires the detection model to make predictions based only on the claim or the evidence. Specifically, all incoming links of the intermediate variable o are eliminated through causal intervention. In this case, the detection model must rely on spurious correlations. In this method, the prediction under this counterfactual condition is defined as: ; where is the prediction result under the counterfactual condition, and ]>represent the predictions based only on the claim and the evidence respectively. While and are two introduced hyperparameters to avoid underestimation or overestimation. Therefore, the DP2CF model is introduced to calculate and .

[0056] Such as Figure 5As shown in Figure 2, the main purpose of the DP2CF debiasing method is to extract the relationship between the statement or evidence and the predicted label. Considering the efficiency and generalization ability of the DP2CF debiasing method, TextCNN is used to extract the main features. Before entering the DP2CF model, each sentence is processed in the text embedding module to form a dimension of The embedding vector of represents the length of each sentence, and D represents the dimension of the embedding. After the convolution operation, an embedding of dimension The convolution vector of represents the number of filters, and the kernel is the core set In addition, each convolution vector corresponds to a kernel Subsequently, the convolution vector passes through an activation function (ReLU) and a max pooling layer to identify the most important features, forming K intermediate feature vectors, which are then merged into a concatenated vector through the Cocat operation.

[0057] To more effectively mitigate the impact of bias on causal prediction output in a cross-platform setting, a gradient reversal layer (GRL) is employed. GRL works by reversing the gradient flow, enabling the model to learn both information related to the target variable and features unrelated to bias during training. This dual learning process enhances the model's generalization across different platforms and improves the accuracy of causal inference. This is followed by a fully feedforward network (FFN) layer. The FFN layer is the final stage of this application's model architecture. It takes the learned feature vector as input and performs a series of nonlinear transformations to generate the final prediction.

[0058] In one embodiment, the feature fusion module maps each output feature vector to a fuzzy membership space using fuzzy logic principles; the membership function expression is: ; in, is the membership function, m is the parameter that determines the center position of the curve, is the parameter that determines the degree of diffusion of the curve. x is the independent variable of the membership function.

[0059] The instance The fuzzy membership values of all features are averaged to obtain the instance For categories Membership value of For categories The membership value expression is: ; in, For an instance For a category the membership value represents the dimension of the feature vector is the i th feature of the k th instance and are two parameters of the membership function corresponding to the feature belonging to the category respectively

[0060] Specifically, the feature fusion module aims to enhance accuracy, stability, and generalization ability by integrating multi-dimensional data features. This module uses the principle of fuzzy logic to map each output feature vector to a fuzzy membership space. Therefore, this method can more effectively capture the inherent features of the underlying patterns while reducing the impact of noise. In the feature fusion module, the Gaussian membership function is represented by and m, which determine the shape of the function. Each membership function is as shown in the above membership function expression

[0061] To calculate the membership value of the instance for the category it is necessary to average the fuzzy membership values of all features. The calculation method is as shown in the membership value expression of the above instance for the category

[0062] Therefore, according to the above description, the debiased prediction can be calculated as follows ; where represents the debiased prediction result of false information represents the prediction result of the traditional detection model represents the prediction result under counterfactual conditions

[0063] It should be understood that although Figure 2 the steps in the flowchart of Figure 2 are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreoverat least a part of the steps in may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or sub-steps or stages of other steps

[0064] In a validation example, the Snopes and PolitiFact datasets collected by Hansen were used for experimental analysis. In the preprocessing stage, the Snopes and PolitiFact datasets are characterized by binary labels ("True" or "False"), which are naturally suitable for binary classification tasks. The ratio of real news to fake news in the Snopes dataset is approximately 1:3. In the PolitiFact dataset, the ratio of real news to fake news is roughly 1:1. In addition, these datasets are divided into a training set, a validation set, and a test set. The detailed statistics of the above datasets are shown in Table 1.

[0065] Table 1 Detailed data of Snopes and Politifact datasets

[0066] (1) Baseline models To more effectively evaluate the performance of the DP2CF method, DP2CF and CF debiasing methods were implemented on the following state-of-the-art evidence-aware fake news detection methods.

[0067] DeClare evaluates the credibility of claims through a neural network, uses external evidence, and avoids artificial features.

[0068] Due to its hierarchical multi-head attention mechanism, MAC can understand explanations at the word level and the evidence level.

[0069] BERT captures language patterns and relationships through unsupervised learning on large-scale text data. The advantage of BERT lies in its bidirectional Transformer architecture, which can fully consider context information to generate accurate word embeddings.

[0070] According to previous research, the experiments were conducted on two different datasets (PolitiFact and Snopes). When the training dataset and the test dataset are the same, the performance of each baseline model on the PolitiFact and Snopes datasets is shown in Table 2.

[0071] Table 2 Running results of baseline models under the condition that the training set and the test set are the same

[0072] In the cross-platform setting, each baseline model will be applied to the PolitiFact and Snopes datasets and different debiasing strategies will be adopted.

[0073] (2) Experimental settings Select AUC, F1-Macro, and F1-Micro as evaluation metrics. According to the cross-platform setting proposed by Hansen, each model is trained and validated on one dataset and tested on another dataset. Since F1-Macro can effectively handle the class distribution imbalance and consider both precision and recall simultaneously, it can achieve a balance among the key metrics in fake news detection. During the training phase, if F1-Macro does not improve within 10 consecutive epochs, each model will stop training.

[0074] (3)Implementation Details The maximum lengths of the claim and evidence are set to 30 and 100 respectively. Adam is used as the optimizer. Each claim is paired with at most 30 pieces of evidence from the original dataset. The final output of each evaluation metric is aggregated based on five different random seeds. This paper uses hyperparameters α and β, and their values are from {0.001, 0.005, 0.01, 0.05, 0.1, 0.5}. All experiments mentioned in this paper are conducted through Python 3.9.19 and Pytorch 2.0.0, using an NVIDIA RTX 2080 Ti GPU.

[0075] (4)Experimental Results and Analysis 1)Overall Performance In Tables 2 and 3, the performance of three baseline models on two datasets (PolitiFact and Snopes) is compared, and there are significant differences in data distribution between these two datasets. Specifically, Table 2 shows the AUC, F1-Macro, and F1-Micro scores of the baseline models when the training and test datasets are the same. Meanwhile, Table 3 shows the performance metrics of the baseline models under the cross-platform setting. Each baseline model is evaluated using a different debiasing method. Among them, None means no debiasing method is applied in the corresponding experiment, CF represents the counterfactual inference debiasing method, and DP2CF represents the new debiasing method proposed in this paper. Through Tables 2 and 3, the following observations can be made.

[0076] First, from the results in Tables 2 and 3, it can be seen that without using debiasing methods, the evaluation metrics of the baseline models tend to decline when the test dataset is different from the training dataset. In particular, when the training dataset is Snopes, the F1-Macro score of BERT significantly drops from 61.34% to 45.34%, DeClare drops from 69.87% to 49.05%, and MAC drops from 78.66% to 51.75%. Additionally, other evaluation metrics, including AUC and F1-Micro, also show an obvious downward trend. If the training dataset is switched to PolitiFact, although the performance still declines in terms of AUC and F1-Macro metrics, this downward trend is no longer obvious in the F1-Micro metric.

[0077] Second, in Table 3, compared with the slight improvements brought by the CF debiasing method, DP2CF achieved significant improvements for each baseline model in all evaluation metrics. This indicates that DP2CF can more effectively reduce the bias in news and evidence content. Specifically, compared with the case without using debiasing methods, BERT improved by 8% to 14% in AUC, 10% to 11% in F1-Macro, and 5% to 10% in F1-Micro. For DeClare, AUC improved by 2% to 7%, F1-Macro improved by 3% to 6%, and F1-Micro improved by 3% to 10%. In contrast, MAC improved by 1% to 2% in F1-Macro and 1% to 5% in F1-Micro. These results show that compared with the baseline models with better initial performance, the DP2CF debiasing method can bring more significant performance improvements to the baseline models with poorer initial performance.

[0078] Based on the above results, the following conclusions can be drawn. First, the decline in the performance of the baseline models when the training and test datasets are different highlights the importance of addressing data bias. This indicates that a model trained on one dataset may not generalize well to another dataset, especially when there are significant differences in their data distributions. Second, the significant improvements achieved by the DP2CF debiasing method in all baseline models and evaluation metrics (superior to CF) demonstrate its effectiveness in reducing bias in claim and evidence content. Finally, among the three baseline models, DP2CF achieved the most significant improvements on BERT. This is because the text embedding module of BERT is different from the other two baseline models, and BERT has the strongest ability to capture semantic information, which may also make it more vulnerable to the spurious correlations between claims, evidence, and predicted labels. Therefore, the DP2CF debiasing method has the most significant improvements on the BERT model.

[0079] 2) Ablation experiment To further clarify the effectiveness of the DP2CF debiasing method in reducing claim bias and evidence bias, an ablation study was conducted on three baseline models in a cross-platform setting. The results are shown in Table 4. As Figure 4 shown, the DP2CF debiasing method is applied to two separate parts. Therefore, the debiasing strategies listed in Table 4 are designed as follows: (1) None: The DP2CF debiasing method is not applied. (2) Evidence only: The DP2CF debiasing method is limited to the evidence part, excluding the embedded claim feature vectors and focusing on evidence-related features; (3) Claim only: The embedded evidence feature vectors are removed from the DP2CF model, thus highlighting claim-related features while ignoring evidence features; (4) DP2CF debiasing method: The feature vectors of both claims and evidence are incorporated into the DP2CF model. Based on the results, the following key insights can be drawn.

[0080] First, compared with the "None" debiasing method that does not use the DP2CF debiasing method, all other strategies that include the DP2CF debiasing method show consistent improvements in all evaluation metrics. However, there are also a few exceptions. When the "Evidence only" debiasing method is applied to DeClare using the Snopes training dataset, a slight decrease in the AUC metric is observed. Similarly, when the "Evidence only" debiasing method is adopted by the MAC baseline model using the Snopes training dataset, a slight decrease in the F1-Macro metric occurs. This may be because the evidence bias in the Snopes dataset is not as obvious as that in the PolitiFact dataset. Therefore, the debiasing method applied to the evidence part may unexpectedly affect the evidence-aware causal reasoning method.

[0081] Second, compared with the "Evidence only" and "Claim only" debiasing methods, the DP2CF debiasing method has achieved significant improvements in all evaluation metrics. Especially when focusing on the F1-Macro metric, BERT can achieve an improvement of 2% to 4%, DeClare can obtain a boost of 1% to 3%, and MAC can achieve an enhancement of 0.6% to 3%. This result fully demonstrates that when the DP2CF debiasing method effectively reduces both claim bias and evidence bias, the fake news detection model can achieve better performance. Both the claim part and the evidence part are useful for the DP2CF debiasing method.

[0082] Table 3 Comparison results in cross-platform setting

[0083] Table 4 Ablation experiments in cross-platform setting

[0084] 3) Hyperparameter sensitivity analysis In this embodiment, the sensitivity of DP2CF to the hyperparameters α and β is discussed in detail on two datasets. As Figure 6 shown, and take values from the set {0.001, 0.005, 0.01, 0.05, 0.1, 0.5}. The vertical axis represents the F1-Macro score, and each point on the surface corresponds to a specific set of and values, representing a discrete combination set rather than a continuous surface. Specifically, in the cross-platform setting, the best performance of BERT during the Snopes test occurs at , while the best performance during the PolitiFact test occurs at . The best performance of DeClare during the Snopes test occurs at , while the best performance during the PolitiFact test occurs at , . The best performance of MAC during the Snopes test occurs at , while the best performance during the PolitiFact test occurs at . These seemingly different results may be due to differences in the dataset distribution and variations between the baseline models.

[0085] Furthermore, the experiment is conducted under two settings, and the DP2CF debiasing method is applied to three evidence-aware fake news detection models. The DP2CF debiasing method is significantly better than the original detection model and the CF debiasing method. This result highlights the potential of the DP2CF debiasing method as a powerful tool for developing more robust and debiased models to combat the spread of fake news.

[0086] In one embodiment, a misinformation detection device based on a plug-and-play counterfactual debiasing framework is provided, including: a text information dataset acquisition unit to be detected, a feature embedding unit, a traditional prediction unit, a prediction unit under counterfactual conditions, and a misinformation debiasing prediction result determination unit, where: The text information dataset acquisition unit to be detected is used to acquire the text information dataset to be detected, and each data information includes: a claim set, a corresponding evidence set, and a true information label.

[0087] The feature embedding unit is used to convert the claim set and the corresponding evidence set into a claim embedding feature vector and an evidence embedding feature vector by using a text embedding module.

[0088] The traditional prediction unit is used to make a prediction using a traditional evidence-aware misinformation detection model based on the claim embedding feature vector and the evidence embedding feature vector to obtain a traditional prediction result.

[0089] A prediction unit under counterfactual conditions is configured to process a claim embedding feature vector and an evidence embedding feature vector using a debiasing model to obtain a claim-based prediction result and an evidence-based prediction result; the debiasing model adopts a debiasing method of counterfactual reasoning; a weighted sum of the claim-based prediction result and the evidence-based prediction result is obtained to get the prediction result under counterfactual conditions.

[0090] A misinformation debiasing prediction result determination unit is configured to fuse the traditional prediction result and the prediction result under counterfactual conditions using a feature fusion module to obtain a misinformation debiasing prediction result.

[0091] In one embodiment, if the BERT model is selected as the basis for the traditional evidence-aware misinformation detection model, the BERT model is selected as the text embedding module in the feature embedding unit.

[0092] In one embodiment, when the traditional evidence-aware misinformation detection model is MAC or DeClare, the GloVe model is selected as the text embedding module in the feature embedding unit.

[0093] In one embodiment, the debiasing model includes: a claim debiasing model and an evidence debiasing model; the claim debiasing model is configured to extract the relationship between the claim embedding feature vector and the prediction label to obtain a claim-based prediction result; the evidence debiasing model is configured to extract the relationship between the evidence embedding feature vector and the prediction label to obtain an evidence-based prediction result; the prediction unit under counterfactual conditions is further configured to input the claim embedding feature vector into the claim debiasing model to obtain a claim-based prediction result; and input the evidence embedding feature vector into the evidence debiasing model to obtain an evidence-based prediction result.

[0094] In one embodiment, the claim debiasing model includes: a TextCNN module, a max pooling layer, a gradient reversal layer, and a fully feed-forward network layer; the prediction unit under counterfactual conditions is further configured to, in the claim debiasing model: extract features from the claim embedding feature vector using the TextCNN module to obtain a convolutional vector; activate the convolutional vector through a ReLU function and then process it using the max pooling layer to obtain |K| first intermediate feature vectors; where K is a set of kernels for convolution; splice all the intermediate feature vectors through a Cocat operation to obtain a spliced vector; process the spliced vector through the gradient reversal layer and then input it into the fully feed-forward network layer to obtain a claim-based prediction result.

[0095] In one embodiment, the evidence debiasing model has the same structure as the claim debiasing model; the prediction unit under counterfactual conditions is further configured to, in the evidence debiasing model: perform feature extraction on the evidence embedding feature vector using a TextCNN module to obtain a first convolutional vector; activate the first convolutional vector through a ReLU function and then process it using a max pooling layer to obtain |K| first intermediate feature vectors, where K is a set of kernels for convolution; splice all the first intermediate feature vectors through a Cocat operation to obtain a first spliced vector; process the first spliced vector through a gradient reversal layer and then input it into a fully feed-forward network layer to obtain an evidence-based prediction result.

[0096] In one embodiment, the false information debiasing prediction result determination unit is further configured to use a feature fusion module to fuse the traditional prediction result and the prediction result under counterfactual conditions to obtain the false information debiasing prediction result as described in the false information debiasing prediction result expression above.

[0097] In one embodiment, the feature fusion module maps each output feature vector to a fuzzy membership space using the fuzzy logic principle; the membership function is as shown in the membership function expression above.

[0098] Average the fuzzy membership values of all features of the instance to obtain the membership degree value of the instance for the category The value is as that of the instance i for the category j as shown in the membership degree value expression.

[0099] For the specific limitations of the false information detection device based on the plug-and-play counterfactual debiasing framework, reference may be made to the limitations of the false information detection method based on the plug-and-play counterfactual debiasing framework in the above text, which will not be elaborated here. Each module in the above false information detection device based on the plug-and-play counterfactual debiasing framework can be implemented in whole or in part through software, hardware, and their combination. The above modules can be embedded in the processor of the computer device in hardware form or be independent of it, or can be stored in the memory of the computer device in software form for the processor to call and execute the operations corresponding to the above respective modules.

[0100] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity in description, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combinations of these technical features do not conflict, they should all be considered as the scope recorded in this specification.

[0101] The above-described embodiments merely represent several implementation manners of the present application. The description thereof is relatively specific and detailed, but it should not be construed as a limitation to the scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all fall within the protection scope of the present application.

Claims

1. A misinformation detection method based on a plug-and-play counterfactual debiasing framework, characterized in that, The method includes: Obtain a text information dataset to be detected, where each data information includes: a claim set, a corresponding evidence set, and a true information label; Use a text embedding module to convert the claim set and the corresponding evidence set into a claim embedding feature vector and an evidence embedding feature vector; Perform prediction using a traditional evidence-aware false information detection model based on the claim embedding feature vector and the evidence embedding feature vector to obtain a traditional prediction result; Process the claim embedding feature vector and the evidence embedding feature vector using a debiasing model to obtain a claim-based prediction result and an evidence-based prediction result; the debiasing model uses a debiasing method based on counterfactual reasoning; Perform weighted summation on the claim-based prediction result and the evidence-based prediction result to obtain a prediction result under counterfactual conditions; Use a feature fusion module to fuse the traditional prediction result and the prediction result under counterfactual conditions to obtain a false information debiased prediction result.

2. The false information detection method based on the plug-and-play counterfactual debiasing framework according to claim 1, wherein If the BERT model is selected as the basis for the traditional evidence-aware false information detection model, then select the BERT model as the text embedding module.

3. The false information detection method based on the plug-and-play counterfactual debiasing framework according to claim 1, wherein When the traditional evidence-aware false information detection model is MAC or DeClare, then select the GloVe model as the text embedding module.

4. The false information detection method based on the plug-and-play counterfactual debiasing framework according to claim 1, wherein The debiasing model includes: a claim debiasing model and an evidence debiasing model; the claim debiasing model is used to extract the relationship between the claim embedding feature vector and the prediction label to obtain a claim-based prediction result; the evidence debiasing model is used to extract the relationship between the evidence embedding feature vector and the prediction label to obtain an evidence-based prediction result; Process the claim embedding feature vector and the evidence embedding feature vector using a debiasing model to obtain a claim-based prediction result and an evidence-based prediction result, including: Input the claim embedding feature vector into the claim debiasing model to obtain a claim-based prediction result; Input the evidence embedding feature vector into the evidence debiasing model to obtain an evidence-based prediction result.

5. The false information detection method based on the plug-and-play counterfactual debiasing framework according to claim 4, wherein The claim debiasing model includes: a TextCNN module, a max pooling layer, a gradient reversal layer, and a fully feed-forward network layer; In the claim debiasing model, Use the TextCNN module to extract features from the claim embedding feature vector to obtain a convolutional vector; Process the convolutional vector through the ReLU function and then use the max pooling layer to obtain |K| intermediate feature vectors; where K is the set of kernels used for convolution; Concatenate all the intermediate feature vectors through a Cocat operation to obtain a concatenated vector; Process the concatenated vector through the gradient reversal layer and then input it into the fully feed-forward network layer to obtain a claim-based prediction result.

6. The false information detection method based on the plug-and-play counterfactual debiasing framework according to claim 4, wherein The evidence debiasing model has the same structure as the claim debiasing model; In the evidence debiasing model: Use the TextCNN module to extract features from the evidence embedding feature vector to obtain a first convolutional vector; After activating the first convolution vector through the ReLU function, it is processed by a max pooling layer to obtain |K| first intermediate feature vectors; where K is the set of kernels for convolution; All the first intermediate feature vectors are concatenated through a Cocat operation to obtain a first concatenated vector; The first concatenated vector is processed through a gradient reversal layer and then input into a fully feed-forward network layer to obtain an evidence-based prediction result.

7. The false information detection method based on the plug-and-play counterfactual debiasing framework according to claim 1, wherein A feature fusion module is used to fuse the traditional prediction result and the prediction result under counterfactual conditions to obtain a false information debiased prediction result: ； Among them, represents the debiased prediction result of false information, represents the prediction result of the traditional detection model, represents the prediction result under counterfactual conditions.

8. The false information detection method based on the plug-and-play counterfactual debiasing framework according to claim 1, wherein The feature fusion module maps each output feature vector to a fuzzy membership space using the fuzzy logic principle; the membership function is: ； Among them, is the membership function, is the parameter that determines the center position of the curve, is the parameter that determines the diffusion degree of the curve, is the independent variable of the membership function; Average the fuzzy membership values of all features of the instance to obtain the instance The membership degree value for the category is as follows: ； Among them, is an instance For the category membership value, represents the dimension of the feature vector, is the th feature of the th instance, k and j are two parameters of the membership function corresponding to the feature of the category 9. A misinformation detection device based on a plug-and-play counterfactual debiasing framework, characterized in that, The device includes: A text information dataset acquisition unit to be detected, which is used to acquire a text information dataset to be detected, and each data information includes: a claim set, a corresponding evidence set, and a true information label; A feature embedding unit, which is used to convert the claim set and the corresponding evidence set into a claim embedding feature vector and an evidence embedding feature vector by using a text embedding module; A traditional prediction unit, which is used to predict by using a traditional evidence-aware false information detection model according to the claim embedding feature vector and the evidence embedding feature vector to obtain a traditional prediction result; A prediction unit under counterfactual conditions, which is used to process the claim embedding feature vector and the evidence embedding feature vector by using a debiasing model to obtain a claim-based prediction result and an evidence-based prediction result; the debiasing model adopts a debiasing method of counterfactual reasoning; the claim-based prediction result and the evidence-based prediction result are weighted and summed to obtain a prediction result under counterfactual conditions; A false information debiased prediction result determination unit, which is used to fuse the traditional prediction result and the prediction result under counterfactual conditions by using a feature fusion module to obtain a false information debiased prediction result.

Images