A Cross-Domain Fake News Detection Method and System Based on Domain and Instance-Level Transfer Learning

Abstract

This invention proposes a cross-domain fake news detection method and system based on domain and example-level transfer learning, comprising: acquiring news data from multiple source domains and one target domain as training data; training a generalization model based on the training data using meta-learning; performing a masked language training task based on the news data of the target domain to train a masked language model, thereby obtaining a language model for transferability measurement; calculating the transferability of the language model on the news data of the source domains, and weighting the news data of the source domains according to the transferability; using the news data of the target domain and the weighted news data of the source domains to guide the generalization model to adapt to the target domain, thereby obtaining a fake news detection model for the target domain; and using the fake news detection model to complete the fake news detection of news in the target domain.

Description

Technical Field

[0001] This invention relates to the field of fake news detection technology, and in particular to a cross-domain fake news detection method and system based on domain and example-level transfer. Background Technology

[0002] In real life, fake news permeates many fields, such as politics, health, and entertainment. While fake news in every field needs to be detected, some specific areas have a more serious social impact. Therefore, focusing on detecting fake news in these areas is particularly important.

[0003] Existing research on fake news detection includes methods targeting specific domains (e.g., health, politics), but these methods train models on a single domain (single-domain fake news detection), failing to address the problem of limited data in certain domains. Since data across different domains are correlated, some research uses data from multiple domains for modeling (multi-domain fake news detection). These methods improve overall fake news detection capabilities, but while improving detection in some domains, they also result in performance loss in others (a "seesaw effect"). Furthermore, because these multi-domain fake news detection methods lack a domain-oriented design, they struggle to specifically improve performance in a particular target domain. Summary of the Invention

[0004] To address the challenge of cross-domain fake news detection, we propose a method combining domain-level and sample-level transfer learning. To mitigate the aforementioned "seesaw" effect and enhance the model's generalization ability, we train the model from a meta-learning perspective, resulting in a generalizable model that encompasses common knowledge across multiple domains. To quantitatively assess the relevance between source domain data and the target domain, we train a language model for the target domain on the MLM (Mask Language Model) task and calculate its "perplexity" on source domain data to represent the relevance between the two. We then weight the source domain data based on this relevance. Finally, to guide the model to improve its fake news detection capability in the target domain, we use both target domain data and weighted source domain data to guide the generalizable model to adapt to the target domain.

[0005] Specifically, this invention proposes a cross-domain fake news detection method based on domain and example-level transfer, including:

[0006] Step 1: Obtain news data from multiple source domains and one target domain as training data. Based on this training data, train a generalization model using meta-learning.

[0007] Step 2: Based on news data in the target domain, perform a masked language training task to train the masked language model and obtain a language model for transferability measurement.

[0008] Step 3: Calculate the transferability of the language model on the news data of the source domain, and weight the news data of the source domain according to the transferability;

[0009] Step 4: Using news data from the target domain and weighted news data from the source domain, guide the generalization model to adapt to the target domain, and obtain the fake news detection model for the target domain;

[0010] Step 5: Use the fake news detection model to detect fake news in the target area.

[0011] The cross-domain fake news detection method based on domain and example-level transfer, wherein step 1 includes:

[0012] In the meta-learning approach, in addition to model parameters, each task has its own set of model parameters. During each model parameter update, a batch of data is divided into n task data. For each task data, it is divided into a support set and a query set. The model is first trained on the support set and the corresponding classification loss is calculated using cross-entropy.

[0013]

[0014] Where m s This is the current data volume of the support set, y i For the true value of the i-th data point, Given the predicted value of the i-th data point, update the model parameters θ corresponding to the current task data based on this loss. d :

[0015]

[0016] Then, the model for the current task is assigned the parameter θ. d And calculate the loss on the query set of the current task data:

[0017]

[0018] Where m q y is the data size of the current task query set. i For the true value of the i-th data point, For the predicted value of the i-th data point, the loss is calculated by summing the results of all tasks in a batch of data after iterating through all tasks. And update the model parameters based on this loss:

[0019] The aforementioned cross-domain fake news detection method based on domain and example-level transfer, wherein...

[0020] Step 2 includes:

[0021] For each piece of data in the target domain, after word segmentation, a sequence {w1,...,w...} containing n word elements is obtained. n The sequence is replaced with a mask [MASK]. The mask language model is then used to predict the word at the [MASK] position based on the other words in the sequence. During the training of the mask language model, the language model is updated by minimizing the cross-entropy loss to obtain a target domain-adaptive language model.

[0022] Step 3 includes:

[0023] P, each piece of data in the source domain s Word segmentation yields P s ={[CLS],w0,...,w m-1 [SEP]}, and then the masked data P is obtained by sequentially masking the tokens at each position. mask = {[CLS],...,w i-1 [MASK],w i+1 ,...,[SEP]};Use a target domain-adaptive language model to predict each word in the source domain data and calculate the probability of its correct prediction prob(w i ), to calculate the perplexity pp of the language model:

[0024] prob(w i ) = LM(P mask ,w i )

[0025]

[0026] The portability of the source domain data is obtained by the following formula: w = 1 / pp.

[0027] The cross-domain fake news detection method based on domain and example-level transfer, wherein step 4 includes: updating the model parameters through cross-entropy loss to obtain the fake news detection model for the target domain:

[0028]

[0029]

[0030] Where y is the true value of the sample. For the sample predicted value, Let w(x) be the cross-entropy loss function, and w(x) be the sample weights.

[0031] This invention also proposes a cross-domain fake news detection system based on domain and example-level transfer, including:

[0032] Module 1 is used to acquire news data from multiple source domains and one target domain as training data. Based on this training data, a generalization model is trained through meta-learning.

[0033] Module 2 is used to perform masked language training tasks based on news data in the target domain, train a masked language model, and obtain a language model for transferability measurement.

[0034] Module 3 is used to calculate the transferability of the language model on the news data of the source domain, and to weight the news data of the source domain based on the transferability;

[0035] Module 4 is used to guide the generalization model to adapt to the target domain using news data from the target domain and weighted news data from the source domain, thereby obtaining a fake news detection model for the target domain.

[0036] Module 5 is used to perform fake news detection in the target field using this fake news detection model.

[0037] The aforementioned cross-domain fake news detection system based on domain and example-level transfer, wherein module 1 is used for:

[0038] In the meta-learning approach, in addition to model parameters, each task has its own set of model parameters. During each model parameter update, a batch of data is divided into n task data. For each task data, it is divided into a support set and a query set. The model is first trained on the support set and the corresponding classification loss is calculated using cross-entropy.

[0039]

[0040] Where m s This is the current data volume of the support set, y i For the true value of the i-th data point, Given the predicted value of the i-th data point, update the model parameters θ corresponding to the current task data based on this loss. d :

[0041]

[0042] Then, the model for the current task is assigned the parameter θ. d And calculate the loss on the query set of the current task data:

[0043]

[0044] Where m q y is the data size of the current task query set. i For the true value of the i-th data point, For the predicted value of the i-th data point, the loss is calculated by summing the results of all tasks in a batch of data after iterating through all tasks. And update the model parameters based on this loss:

[0045] The aforementioned cross-domain fake news detection system based on domain and example-level transfer, wherein...

[0046] This module 2 is used for:

[0047] For each piece of data in the target domain, after word segmentation, a sequence {w1,...,w...} containing n word elements is obtained. n The sequence is replaced with a mask [MASK]. The mask language model is then used to predict the word at the [MASK] position based on the other words in the sequence. During the training of the mask language model, the language model is updated by minimizing the cross-entropy loss to obtain a target domain-adaptive language model.

[0048] This module 3 is used for:

[0049] P, each piece of data in the source domain s Word segmentation yields P s ={[CLS],w0,...,w m-1 [SEP]}, and then the masked data P is obtained by sequentially masking the tokens at each position. mask = {[CLS],...,w i-1 [MASK],w i+1 ,...,[SEP]};Use a target domain-adaptive language model to predict each word in the source domain data and calculate the probability of its correct prediction prob(w i ), to calculate the perplexity pp of the language model:

[0050] prob(w i ) = LM(P mask ,w i )

[0051]

[0052] The portability of the source domain data is obtained by the following formula: w = 1 / pp.

[0053] The aforementioned cross-domain fake news detection system based on domain and example-level transfer, wherein module 4 includes: updating the model parameters through cross-entropy loss to obtain a fake news detection model for the target domain:

[0054]

[0055]

[0056] Where y is the true value of the sample. For the sample predicted value, Let w(x) be the cross-entropy loss function, and w(x) be the sample weights.

[0057] The present invention also proposes a storage medium for storing a program that executes any of the cross-domain fake news detection methods based on domain and example-level migration.

[0058] This invention also proposes a client for any of the aforementioned cross-domain fake news detection systems based on domain and example-level migration.

[0059] As can be seen from the above solutions, the advantages of the present invention are:

[0060] The meta-learning-based generalizable model training method can obtain a generalizable model that covers common knowledge in multiple domains; it can effectively measure the transferability of source domain data through the "perplexity" metric, thereby weighting the source domain samples; and it can guide the "generalizable model" to adapt to the target domain by using the weighted source domain data and target domain data, thus enabling it to have a good detection capability for the target domain. Attached Figure Description

[0061] Figure 1 This is a flowchart of a cross-domain fake news detection method based on domain-level and sample-level transfer. Detailed Implementation

[0062] The purpose of this invention is to address the problem of improving the detection capability of fake news in a target domain in cross-domain fake news detection. It proposes a cross-domain fake news detection method that combines domain-level and sample-level transfer learning, and its main key technical points include the following:

[0063] Key Point 1: Training of generalizable models based on meta-learning. This method trains the model on data from multiple domains using a meta-learning training strategy. By simulating the phenomenon of "different distributions of training and test data" during multiple rounds of training, the model acquires excellent generalization ability.

[0064] Key Point 2: Sample Transferability Assessment Based on Perplexity. This method first trains a language model for the target domain using the MLM (Masked Language Model) task. Then, it calculates the perplexity of the language model on all source domain data, which serves as a representation of the relevance between the data in the source domain and the target domain. That is, the lower the perplexity, the higher the relevance between the corresponding data and the target domain.

[0065] Key point 3: A cross-domain fake news detection method combining domain-level and sample-level transfer learning. This method uses both target domain data and source domain data weighted according to transferability to guide a "generalizable model" to adapt to the target domain, thereby enabling it to have a good detection capability for fake news in the target domain.

[0066] To make the above features and effects of the present invention clearer and easier to understand, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings.

[0067] The overall process of this invention's cross-domain fake news detection method, which combines domain-level and sample-level transfer learning, is as follows: Figure 1 As shown, the method involves several steps: First, a generalizable model is trained using data from N source domains (e.g., social life, culture and entertainment) and one target domain (politics or health) through meta-learning. Second, a language model for measuring transferability is trained using the target domain data through an MLM (masked language model) task. Third, the perplexity of the language model is calculated on the source domain data to measure the transferability of the source domain data, and the source domain data is weighted based on this transferability. Finally, the generalizable model is guided to adapt to the target domain using the target domain data and the weighted source domain data. The following sections provide a detailed description of each module of this method.

[0068] I. Training of generalizable models

[0069]

[0070] To achieve domain-level transfer learning, we train a generalizable model on data from all domains using meta-learning, as shown in Algorithm 1. During training, we divide a batch of data into n parts, each belonging to one of n tasks. In the meta-learning training approach, in addition to the parameters of the generalizable model, we record an extra set of model parameters for each task. Each update of the model parameters involves two steps: first, training and updating the model parameters for each task; then, updating the generalizable model parameters. Specifically, we divide the data for each task into two non-overlapping parts—the support set and the query set. The model is first trained on the support set, and the corresponding classification loss is calculated using cross-entropy.

[0071]

[0072] Here m s This is the current data volume of the support set, y i For the true value of the i-th data point, Let be the predicted value of the i-th data point. Update the parameter θ corresponding to the current task based on this loss. d :

[0073]

[0074] Then, the model for the current task is assigned the parameter θ. d And calculate the loss on the query set of the current task:

[0075]

[0076] Here m q y is the data size of the current task query set. i For the true value of the i-th data point, Let be the predicted value for the i-th data point. After completing all tasks for a batch of data, we sum the losses calculated on the query set of each task and update the generalizable model parameters θ based on these losses:

[0077]

[0078] Because the phenomenon of "different distribution of training and test data" was simulated during the training process, the final model was endowed with good generalization ability.

[0079] II. Transferability Measurement

[0080] In this process, we measure the transferability of data in the source domain.

[0081] First, we train an MLM (masked language model) on the target domain data to obtain a target domain-adaptive language model. For each piece of data in the target domain, we perform word segmentation to obtain a sequence {w1,...,w} containing n words. n In this sequence, we replace 15% of the word units with [MASK] or any other word unit with a certain probability, or leave them unchanged. We use a language model to predict the masked word unit (at the [MASK] position) based on the other word units in the sequence. During this training process, we update the language model by minimizing the cross-entropy loss.

[0082] Then we use the "perplexity" of the language model to measure the transferability of the source domain data. Perplexity is typically used to measure the predictive ability of a language model for a sentence; a lower perplexity indicates better predictive ability. First, we process each data point P from the source domain... s Word segmentation yields P s = {[CLS],w0,...,w m-1 [SEP]}, and then the masked data P is obtained by sequentially masking the tokens at each position. mask ={[CLS],...,w i-1 [MASK],w i+1 We use a target-domain adaptive language model to predict each word in the source-domain data and calculate the probability of its correct prediction. Finally, we calculate the perplexity of the language model using these probabilities:

[0083] prob(w i ) = LM(P mask ,w i )

[0084]

[0085] Since perplexity is negatively correlated with the transferability of the corresponding data, we represent the transferability of the source domain data and weight it using the following calculation method.

[0086] w = 1 / pp

[0087] III. Target Domain Adaptation

[0088] Finally, to adapt the generalizable model to the target domain, we further guide the training of the generalizable model using data from the target domain and weighted data from the source domain, updating the model's parameters through cross-entropy loss:

[0089]

[0090]

[0091] The following are system embodiments corresponding to the above method embodiments. This embodiment can be implemented in conjunction with the above embodiments. The relevant technical details mentioned in the above embodiments are still valid in this embodiment, and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiments.

[0092] This invention also proposes a cross-domain fake news detection system based on domain and example-level transfer, including:

[0093] Module 1 is used to acquire news data from multiple source domains and one target domain as training data. Based on this training data, a generalization model is trained through meta-learning.

[0094] Module 2 is used to perform masked language training tasks based on news data in the target domain, train a masked language model, and obtain a language model for transferability measurement.

[0095] Module 3 is used to calculate the transferability of the language model on the news data of the source domain, and to weight the news data of the source domain based on the transferability;

[0096] Module 4 is used to guide the generalization model to adapt to the target domain using news data from the target domain and weighted news data from the source domain, thereby obtaining a fake news detection model for the target domain.

[0097] Module 5 is used to perform fake news detection in the target field using this fake news detection model.

[0098] The aforementioned cross-domain fake news detection system based on domain and example-level transfer, wherein module 1 is used for:

[0099] In the meta-learning approach, in addition to model parameters, each task has its own set of model parameters. During each model parameter update, a batch of data is divided into n task data. For each task data, it is divided into a support set and a query set. The model is first trained on the support set and the corresponding classification loss is calculated using cross-entropy.

[0100]

[0101] Where m s This is the current data volume of the support set, y i For the true value of the i-th data point, Given the predicted value of the i-th data point, update the model parameters θ corresponding to the current task data based on this loss. d :

[0102]

[0103] Then, the model for the current task is assigned the parameter θ. d And calculate the loss on the query set of the current task data:

[0104]

[0105] Where m q y is the data size of the current task query set. i For the true value of the i-th data point, For the predicted value of the i-th data point, the loss is calculated by summing the results of all tasks in a batch of data after iterating through all tasks. And update the model parameters based on this loss:

[0106] The aforementioned cross-domain fake news detection system based on domain and example-level transfer, wherein...

[0107] This module 2 is used for:

[0108] For each piece of data in the target domain, after word segmentation, a sequence {w1,...,w...} containing n word elements is obtained. n The sequence is replaced with a mask [MASK]. The mask language model is then used to predict the word at the [MASK] position based on the other words in the sequence. During the training of the mask language model, the language model is updated by minimizing the cross-entropy loss to obtain a target domain-adaptive language model.

[0109] This module 3 is used for:

[0110] P, each piece of data in the source domain s Word segmentation yields P s ={[CLS],w0,...,w m-1 [SEP]}, and then the masked data P is obtained by sequentially masking the tokens at each position. mask = {[CLS],...,w i-1 [MASK],w i+1 ,...,[SEP]};Use a target domain-adaptive language model to predict each word in the source domain data and calculate the probability of its correct prediction prob(w i ), to calculate the perplexity pp of the language model:

[0111] prob(w i ) = LM(P mask ,w i )

[0112]

[0113] The portability of the source domain data is obtained by the following formula: w = 1 / pp.

[0114] The aforementioned cross-domain fake news detection system based on domain and example-level transfer, wherein module 4 includes: updating the model parameters through cross-entropy loss to obtain a fake news detection model for the target domain:

[0115]

[0116]

[0117] Where y is the true value of the sample. For the sample predicted value, Let w(x) be the cross-entropy loss function, and w(x) be the sample weights.

[0118] The present invention also proposes a storage medium for storing a program that executes any of the cross-domain fake news detection methods based on domain and example-level migration.

[0119] This invention also proposes a client for any of the aforementioned cross-domain fake news detection systems based on domain and example-level migration.

Claims

1. A cross-domain fake news detection method based on domain and example-level transfer, characterized in that, include: Step 1: Obtain news data from multiple source domains and one target domain as training data. Based on this training data, train a generalization model using meta-learning. Step 2: Based on news data in the target domain, perform a masked language training task to train the masked language model and obtain a language model for transferability measurement. Step 3: Calculate the transferability of the language model on the news data of the source domain, and weight the news data of the source domain according to the transferability; Step 4: Using news data from the target domain and weighted news data from the source domain, guide the generalization model to adapt to the target domain, and obtain the fake news detection model for the target domain; Step 5: Use the fake news detection model to detect fake news in the target area; Step 2 includes: For each piece of data in the target domain, word segmentation yields a sequence containing n words. The word in the sequence is replaced with a mask [MASK]. The mask language model is used to predict the word at the [MASK] position based on the other words in the sequence. During the training of the mask language model, the language model is updated by minimizing the cross-entropy loss to obtain a target domain adaptive language model. Step 3 includes: Each piece of data from the source domain Word segmentation is performed to obtain Then, the masked data is obtained by sequentially masking the words at each position. Use a target domain-adaptive language model to predict each word in the source domain data and calculate the probability of a correct prediction. To calculate the perplexity of the language model : The transferability of source domain data is obtained using the following formula. .

2. The cross-domain fake news detection method based on domain and example-level transfer as described in claim 1, characterized in that, Step 1 includes: In the meta-learning approach, in addition to model parameters, each task has its own set of model parameters. During each model parameter update, a batch of data is divided into n task data. For each task data, it is divided into a support set and a query set. The model is first trained on the support set and the corresponding classification loss is calculated using cross-entropy. in This is the current data volume of the support set. For the first The true value of each data point. For the first The predicted value of each data point is used to update the model parameters corresponding to the current task data based on this loss. : In the formula The learning rate; the model for the current task is assigned parameters. And calculate the loss on the query set of the current task data: ] in This is the data volume of the current task query set. For the first The true value of each data point. For the first The predicted values ​​for each data point are summed up after all tasks for a batch of data have been completed, resulting in a loss calculated by adding the predicted values ​​to the query set of each task's data. And update the model parameters based on this loss: In the formula This is the learning rate.

3. The cross-domain fake news detection method based on domain and example-level transfer as described in claim 1, characterized in that, Step 4 includes updating the model parameters using cross-entropy loss to obtain a fake news detection model for the target domain. ， in, For the sample true value, For the sample predicted value, Let cross-entropy be the loss function. These are the sample weights.

4. A cross-domain fake news detection system based on domain and example-level transfer, characterized in that, include: Module 1 is used to acquire news data from multiple source domains and one target domain as training data. Based on this training data, a generalization model is trained through meta-learning. Module 2 is used to perform masked language training tasks based on news data in the target domain, train a masked language model, and obtain a language model for transferability measurement. Module 3 is used to calculate the transferability of the language model on the news data of the source domain, and to weight the news data of the source domain based on the transferability; Module 4 is used to guide the generalization model to adapt to the target domain using news data from the target domain and weighted news data from the source domain, thereby obtaining a fake news detection model for the target domain. Module 5 is used to detect fake news in the target area using this fake news detection model; This module 2 is used for: For each piece of data in the target domain, word segmentation yields a sequence containing n words. The word in the sequence is replaced with a mask [MASK]. The mask language model is used to predict the word at the [MASK] position based on the other words in the sequence. During the training of the mask language model, the language model is updated by minimizing the cross-entropy loss to obtain a target domain adaptive language model. This module 3 is used for: Each piece of data from the source domain Word segmentation is performed to obtain Then, the masked data is obtained by sequentially masking the words at each position. Use a target domain-adaptive language model to predict each word in the source domain data and calculate the probability of a correct prediction. To calculate the perplexity of the language model : The transferability of source domain data is obtained using the following formula. .

5. The cross-domain fake news detection system based on domain and example-level transfer as described in claim 4, characterized in that, This module 1 is used for: In the meta-learning approach, in addition to model parameters, each task has its own set of model parameters. During each model parameter update, a batch of data is divided into n task data. For each task data, it is divided into a support set and a query set. The model is first trained on the support set and the corresponding classification loss is calculated using cross-entropy. in This is the current data volume of the support set. For the first The true value of each data point. For the first The predicted value of each data point is used to update the model parameters corresponding to the current task data based on this loss. : In the formula The learning rate; the model for the current task is assigned parameters. And calculate the loss on the query set of the current task data: in This is the data volume of the current task query set. For the first The true value of each data point. For the first The predicted values ​​for each data point are summed up after all tasks for a batch of data have been completed, resulting in a loss calculated by adding the predicted values ​​to the query set of each task's data. And update the model parameters based on this loss: In the formula This is the learning rate.

6. The cross-domain fake news detection system based on domain and example-level transfer as described in claim 4, characterized in that, Module 4 includes: updating the model parameters using cross-entropy loss to obtain a fake news detection model for this target domain. ， in, For the sample true value, For the sample predicted value, Let cross-entropy be the loss function. These are the sample weights.

7. A storage medium for storing a program that performs the cross-domain fake news detection method based on domain and instance-level migration as described in any one of claims 1 to 3.

8. A client for the cross-domain fake news detection system based on domain and example-level migration as described in any one of claims 4 to 6.

Images