Emotion-aware meta-learning based cross-event fake news detection method

Abstract

The application relates to a cross-event fake news detection method based on emotion perception meta-learning, which comprises the following steps: constructing emotion perception meta-tasks and event adaptive meta-tasks; performing weighted meta-learning on the emotion perception meta-tasks; inputting the event adaptive meta-tasks and the emotion perception meta-tasks subjected to the weighted meta-learning into a basic detection model to train a preliminary detection model; fine-tuning the preliminary model by using target event data to obtain an optimal model; and inputting a post to be detected into the optimal model to output a predicted label of the post to be detected. The application deeply integrates emotion into a meta-learning process, learns fine-grained meta-knowledge from existing events, constructs emotion perception tasks, obtains initialization parameters capable of rapidly reaching a minimum value of losses of multiple existing events through double-layer optimization targets, and makes model parameters sensitive to multiple events, so that the model can rapidly adapt to a target event under the condition of only a small amount of labeled data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information detection, and in particular to a cross-event fake news detection method based on emotion perception meta-learning. BACKGROUND

[0002] The development of modern social media has greatly changed the way people access information. However, the widespread dissemination of various false information has brought serious negative effects. Therefore, many deep learning-based methods have been proposed for detecting false information, which have achieved exciting results. However, these methods are not suitable for new events because the labeled data is very limited and its data distribution is inconsistent with that of existing events. Existing scholars have proposed domain adaptation methods to alleviate these problems. However, since their goal is to align the domain information of existing events, and it is difficult to capture the fine-grained differences between real and fake posts using only semantic information, their performance is suboptimal.

[0003] In recent years, statistical-based methods and deep learning techniques have been widely used in fake news detection and have achieved good results. However, this approach has two limitations: (1) Data Greed: Training deep learning networks usually requires a large amount of labeled data; (2) Distribution Inconsistency: Existing deep learning methods require test and training data to have the same data distribution. However, when a new event occurs, its post data distribution is different from that of existing events, and there is only a small amount or no labeled data at all, which makes it difficult to use models trained by other events, and it is difficult to directly train models according to new events. Therefore, early detection of false information in new events is challenging.

[0004] Scholars have proposed some methods to use auxiliary information to help early detection, such as social networks and propagation paths. Although the introduction of such additional data can help detection, it also causes a serious data burden and cannot work properly when the required information is lacking. Other works use labeled data of existing events with domain adaptation to learn domain-invariant representations. Although domain adaptation models can learn domain-invariant representations to some extent, the learning goal is to align domain information between existing events. Therefore, model parameters are not sensitive to new events, making it difficult to quickly adapt to new events. In addition, they only focus on the semantic information of posts, ignoring the fine-grained differences between real and fake posts, so their performance is suboptimal.

[0005] Existing fake news detection methods can be roughly divided into three categories according to the type of data used:

[0006] (1) Content-based models. They mainly use text or visual features extracted from social posts for binary classification (true or false). For example, Ma et al. first applied deep learning techniques to fake news detection, inputting each sentence of the post into a recurrent neural network, using the hidden layer vector of the recurrent neural network to represent the post information, and inputting the hidden layer information into the classifier to get the result. Cheng et al. used a variational autoencoder to auto-encode text information to obtain an embedding representation of the post text, and then used the obtained vector for multi-task learning to improve the effectiveness of the model. This method usually requires a large amount of labeled data for training and cannot work effectively in the case of scarce labeled data. Therefore, it is challenging to apply them in the new event detection scenario.

[0007] (2) Social context-based models. They generally rely on rich user interaction features such as comments, retweets, and follows. For example, Jiang et al. modeled the post propagation network and user social network as a heterogeneous graph, modeled the node information in the graph through a heterogeneous graph neural network, and concatenated the post information and user information for detection. This method causes additional data burden and faces the problem of cold start for new users.

[0008] (3) External knowledge-based models. They usually rely on external knowledge for detection. Knowledge graphs and external knowledge bases contain a wealth of knowledge and rich semantic information, which can help us better understand the content of the post. At the same time, it also contains a lot of objective facts that can be compared with the content of the post to identify the falsity of the fake information. For example, Li et al. used a pre-trained fact-checking model to re-find factual evidence in an external knowledge corpus, constructed a star graph with the factual evidence and the post content, and used GCN to fuse the post content and the factual evidence to detect fake posts. However, new entity vocabularies usually appear when new events occur, and there is a lack of corresponding factual evidence; this method often does not work properly. SUMMARY

[0009] In view of the above problems existing in the prior art, the technical problem to be solved by the present application is how to detect fake information with a small amount of samples under new events.

[0010] To solve the above technical problems, the present application adopts the following technical solution: a cross-event fake news detection method based on emotion perception meta-learning, comprising the following steps:

[0011] S1: Constructing emotion perception meta-task and event adaptive meta-task:

[0012] S11: Constructing emotion perception meta-task:

[0013] Let K represent D Sevent set, M e {Real, Fake} is the label set, is D S event set, M e {Real, Fake} is the label set, is the set of all samples with event k e K and label -m e M, and a single post is treated as the i-th meta-task T i query set T i support set is composed of and x i q samples with the most similar emotion sequence and and x i q samples with the most opposite emotion sequence.

[0014] Let have N samples, i.e., there are N emotion perception meta-tasks, the i-th meta-task T i is represented as

[0015]

[0016] Let e i represent the emotion sequence of x i , and the following formula is used to measure the emotion similarity between x i and x j :

[0017]

[0018] where j e {1, 2, …, |D k |} and i ≠ j;

[0019] S12: Construct event adaptive meta-tasks:

[0020] The MMD distance is defined by a specific representation of the Gaussian kernel function φ(·) acting on source event data and target event data The empirical approximation of MMD is:

[0021]

[0022] where, represents the input of the last MLP of the base detection model, i.e., the embedding of the post;

[0023] S2: Weighted meta-learning for emotion perception meta-tasks

[0024] S21: Calculate the corresponding class prototype for each The calculation formula is: ​

[0025]

[0026] where ε(x i ) is the text embedding of x i , the semantic weight is calculated as:

[0027]

[0028] where the cos function is the cosine similarity function, η is a hyperparameter, and 1(·) is an indicator function, represents a sample of event k, and K represents the number of source domain events, represents the remaining event prototypes that are not the current sample;

[0029] S22: Construct an emotion prototype d m for real information and false information in the source data, and obtain the prototype representation by averaging the emotion sequences of all corresponding samples:

[0030]

[0031] where D m is the set of all samples in D s with label m, and the emotion weight is calculated as:

[0032]

[0033] where d m represents the emotion prototype with the same label as the post, and d -m represents the emotion prototype with the opposite label as the post;

[0034] The weight w i of T i is:

[0035]

[0036] where λ is a balance factor;

[0037] S3: Obtain a preliminary detection model by training the basic detection model

[0038] Input the event adaptive meta-task and the emotion perception meta-task after weighting meta-learning in S2 into the basic detection model for training, and constantly update the parameters of the basic detection model according to the loss until the loss no longer changes to obtain the preliminary detection model;

[0039] In each iteration of training the basic detection model, the emotion perception meta-task is selected with probability ρ∈(0,1), and the event adaptation meta-task is selected with probability 1-ρ, where ρ is a hyperparameter.

[0040] S4: Using the target event data obtained from S3 Fine-tuning is performed to obtain the optimal model.

[0041] For the new event data D T According to D T Create a test task that supports random selection of sets and query sets for a test task T. i′ Using T i′ Support sets Perform n fine-tuning steps:

[0042] θ′ j =U n (θ * ;γ)

[0043] Where, θ' j This indicates the result after fine-tuning using the target event data. The parameters, U n Indicates to loss Gradient descent n times, θ * express The parameter γ represents the learning rate;

[0044] For T i′ querysets used The process will be validated. If the prediction accuracy reaches a preset threshold, the next step will be executed; otherwise, [the process will proceed]. Return to S3 as the basic detection model and continue training;

[0045] S5: Detection of posts to be detected in the target domain:

[0046] For a post to be detected, input the post into S4 to obtain the optimal model. Output the predicted tags for the post to be detected.

[0047] Preferably, the step S4 is performed on the product obtained from S3. When fine-tuning, if the number of samples in the target domain is zero, first assign weak labels based on the sentiment sequence of the posts, then assign weights to each post, and finally use these weakly labeled data to... Make fine adjustments;

[0048] Weak labels are generated for posts based on the sentiment sequence of the new event: if the sentiment sequence of a post is closer to the sentiment prototype of a real post, then the weak label of the post is real, and vice versa. The formula is as follows:

[0049]

[0050] where d Real represents the emotion prototype of real posts, d Fake represents the emotion prototype of fake posts;

[0051] If the emotion sequence is closer to its weak label prototype, it can be considered that the confidence of the weak label is stronger, the weight should be larger, and the weight of the weak sample is

[0052]

[0053] where τ is the emotion threshold, and the use represents the learning rate γ of the weak sample and the weak sample weight Gradient descent n times;

[0054]

[0055] As a preferred, the basic detection model in S3 is composed of two parts of emotion feature extractor and semantic feature extractor,

[0056] S31: Emotion feature extractor

[0057] According to prior knowledge, an emotion sequence is constructed for each post, and the post emotion sequence e is captured as a 24-dimensional vector through emotion sequence features, and then emotion features f are obtained through an MLP layer e = MLP(e);

[0058] S32: Semantic feature extractor

[0059] BERT is used as a semantic feature extractor, and the embedding of "[CLS]" in the last layer of BERT is used to represent semantic embedding f t ;

[0060] S33: The final embedding of the post is the combination of emotion embedding and semantic embedding:

[0061] f = f e + f t

[0062] Then, a linear classification detector with softmax is used to detect fake information:

[0063]

[0064] As a preferred, when the emotion perception meta-task is input into the basic detection model for training, the loss in S3 adopts a cross-entropy loss function L C :

[0065]

[0066] where y represents the true label of a post, represents the predicted label of a post;

[0067] When the event adaptive meta-task training is used to train the base detection model, the loss in S3 uses the MMD of two events as the loss function L e Calculation:

[0068]

[0069] As a preference, the S3 trains the base detection model to obtain In the process of the event adaptive meta-task, only the parameters in the emotion feature extractor and the semantic feature extractor in the base detection model are updated, and the emotion perception meta-task is used to update the parameters in the emotion feature extractor, the semantic feature extractor and the linear classification detector.

[0070] Compared with the prior art, the present application has at least the following advantages:

[0071] The present application designs an emotion perception meta-learning method for cross-event fake information detection, and extends it to a zero-shot scenario using weakly supervised learning, deeply integrates emotions into the meta-learning process, learns fine-grained meta-knowledge from existing events, constructs an emotion perception task, obtains initialization parameters that can make the loss of multiple existing events quickly reach a minimum value through a double-layer optimization objective, and makes the model parameters sensitive to multiple events, so that the model can quickly adapt to the target event in the case of only a small amount of labeled data. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 is a flowchart of the method of the present application.

[0073] Figure 2 is a base detection model framework.

[0074] Figure 3 is the contribution of different components in EML, where Figure 3 (a)- Figure 3 (e) respectively represent the ablation experiment results on five different data sets.

[0075] Figure 4 is the contribution of weak samples.

[0076] Figure 5 is the model performance under different settings of two different hyperparameters, where Figure 5 (a) represents the emotion threshold, Figure 5 (b) hyperparameter p.[Explanation Figure 5 (a) andFigure 5 (b) have two lines coincide, so the figure can only see 3 curves DETAILED DESCRIPTION

[0077] The present application is further described below.

[0078] False information is defined as deliberately fabricated posts that can be verified as true or false, and each post contains its text information and corresponding labels. Our goal is to detect false information with a small number of samples under new events. Specifically, we define detection as a few-shot problem.

[0079] Each meta-task will be divided into two disjoint subsets: support set and query set. In meta-training, we fine-tune the model on the source data D S Construct meta-tasks, and then learn appropriate initialization parameters through meta-learning algorithms (double-layer optimization objectives). In meta-testing, we construct a randomly selected meta-task. Fine-tune the model parameters through a small support set, and evaluate the performance of the model by averaging the results on all query sets of the test meta-task. T Construct a randomly selected meta-task. Fine-tune the model parameters through a small support set, and evaluate the performance of the model by averaging the results on all query sets of the test meta-task.

[0080] The motivation for deeply considering emotions in the learning process is that the emotional features of posts provide additional auxiliary information from a new perspective. Emotional features are universal across different events, while the data distribution of different events differs greatly due to the differences in entity words. To illustrate its universality, we show the similarity density between the emotional sequence of the new event and the emotional prototype of the source event in Figure 1 Without loss of generality, we choose the Charlie Hebdo dataset as the new event and the other datasets as the source events. As can be seen from the figure, the emotional features of real posts in the new event are more similar to the real post prototypes in the source events, and vice versa. Similar results can also be obtained on other new events, indicating that the emotional sequence has a positive impact on cross-event false information detection. Therefore, we deeply integrate emotions into the meta-learning process in EML.

[0081] By integrating emotions into the meta-learning process, we construct emotion-aware meta-tasks based on the emotional features of the source event data, and then use meta-training algorithms to find initialization parameters that can quickly adapt to new event data. Compared with traditional meta-learning methods, our emotion-aware meta-tasks can make the model pay attention to the fine-grained differences between true and false posts.

[0082] In meta-learning based on semantic and sentiment features, we aim to notice more valuable data and reduce the impact of noise. Existing meta-learning methods treat all meta-tasks equally, making it difficult to notice more valuable tasks. During meta-training, each meta-task should have different weights, with higher-weighted meta-tasks being prioritized for model learning. Therefore, we design a novel task weighting method based on the semantic and sentiment features of posts, enabling the model to focus more on valuable meta-tasks during meta-training.

[0083] In zero-shot scenarios (the extreme case of new events where no labeled data exists), we propose a weak-labeling method based on sentiment features. We label posts in new events and then fine-tune the model using this weakly labeled data. Since weak labels inevitably contain noise, we assign weights based on the sentiment order of each post. During model fine-tuning, posts with higher weights significantly influence the parameters, thus reducing the impact of noise.

[0084] A cross-event fake news detection method based on emotion perception meta-learning includes the following steps:

[0085] S1: Constructing the emotion-aware meta-task and the event-adaptive meta-task:

[0086] Meta-tasks are the fundamental units of meta-learning, and emotion plays a crucial role in detecting fake information in new events, providing a new perspective for the creation of meta-tasks. Therefore, we construct an emotion-aware meta-task based on the emotion sequence of posts and select samples whose emotions are closest to and most opposite to the target sample as the support set. This allows the model to better capture emotional features and adapt more effectively to new event data. Compared to traditional meta-tasks that randomly select several samples from the same event, our method can capture the differences between real and fake posts with finer granularity because fake posts often possess some common emotional features.

[0087] S11: Constructing the Meta-Task of Emotion Perception

[0088] Let K represent D S The event set M∈{Real,Fake} is the tag set. It is D S The set of all samples in which event k∈K and label m∈M. It is the set of all samples where the event is k∈K and the label is -m∈M, and a single post Treat it as an i-th meta-task T i query set T i Support set Depend on Neutralize x iThe q closest samples of the emotion sequence of x and Neutralize x i The q most opposite samples of the emotion sequence of x form a set;

[0089] Let There are N samples, that is, there are N emotion perception meta-tasks, and the i-th meta-task T i is expressed as

[0090]

[0091] The emotion sequence of x is expressed as e i , and the following formula is used to measure the emotion similarity between x i and x i : j

[0092]

[0093] where j∈{1,2,…,|D k |} and i≠j; The essence of the emotion perception meta-task is false information detection, so the loss function is cross-entropy loss.

[0094] S12: Construct an event adaptive meta-task:

[0095] In addition to making the model extract emotions for detection, we also want the model to learn event-invariant features so that the model can adapt to new events. Therefore, we designed an event adaptive meta-task, which directly aligns the data distribution of different events to adapt to event differences.

[0096] In order to align the data distribution of different events, we use the Maximum Mean variance (MMD) to measure the distance between the distributions of two events. The MMD distance is calculated by a specific representation of the Gaussian kernel function φ(·), which acts on the source event data and the target event data The empirical approximation of MMD is:

[0097]

[0098] where, represents the input of the last MLP of the basic detection model, that is, the embedding of the post;

[0099] We first select two different events k 1 and k 2 from the source data, select p samples from each event to form the support set of the event adaptive task, and then select two different events k 3 and k 4 ​The query set is constructed in the same way. Note that the two events inside the support set and the query set need to be different, but the events in the support set can also be in the query set.

[0100] S2: Weighted meta-learning for emotion perception meta-tasks

[0101] S21: For each , calculate the corresponding class prototype The calculation formula is: this is obtained by averaging the embeddings of all samples in .

[0102]

[0103] where ε(x i ) is the text embedding of x i , which is obtained by taking the “[CLS]” embedding of the last layer of the pre-trained BERT. Ideally, an important training sample is semantically close to its class prototype and not too far from the prototypes of other events of the same class. Therefore, the semantic weight is calculated as:

[0104]

[0105] where the cos function is the cosine similarity function, η is a hyperparameter (0 < η < 1), and 1(·) is an indicator function that returns 1 if the Boolean input function is true, represents an event that is a sample of k, K represents the number of source domain events, represents the prototype of the remaining events that are not the current sample, for example, the current sample is about A, then

[0106] is the prototype of B or C.

[0107] S22: Construct an emotion prototype d m (irrelevant to events) for real information and false information in the source data, and obtain the prototype representation by averaging the emotion sequences of all corresponding samples:

[0108]

[0109] where D m is the set of all samples in D S with label m, similarly, samples should be closer to the emotion prototype with the same label. If the emotion of a real post is closer to the emotion prototype of a false post, the task weight corresponding to the post should be lower. Therefore, the calculation method of the emotion weight is:

[0110]

[0111] where d m represents the same emotion prototype as the post tag, d -m represents the opposite emotion prototype as the post tag.

[0112] T i is the weight of the emotion perception meta-task, and w i is:

[0113]

[0114] where λ is a balance factor. In the meta-training process, we will calculate the gradient according to the weight of each meta-task, and note that this weighted score is only used for the emotion perception meta-task.

[0115] S3: obtaining a preliminary detection model by training the basic detection model

[0116] The event adaptive meta-task and the emotion perception meta-task after weighted meta-learning in S2 are input into the basic detection model for training. The parameters of the basic detection model are updated iteratively according to the loss until the loss no longer changes, and the preliminary detection model is obtained.

[0117] When training the basic detection model M θ , in each iteration, the emotion perception meta-task is selected with a probability of ρ∈(0, 1), and the event adaptive meta-task is selected with a probability of 1-ρ. ρ is a hyperparameter, and we found that ρ=0.9 works better. Through these two types of meta-tasks, the model learns the features of event invariance and emotion perception, thereby more effectively adapting to new events.

[0118] Specifically: given a model M θ with initial parameters θ and a meta-task set First, T i is selected, and then the support set of T i is used to update the parameters of M θ internally; that is, the new parameters θ′ i are obtained by gradient descent n times:

[0119] θi i = U n (θ; α)

[0120] where U n represents gradient descent n times on the loss of , and the learning rate is α. For example, when a single gradient descent is applied, the formula is

[0121]

[0122] Then we update the model parameters θ′ internally. i Calculate the loss of the query set By minimizing relative to M θ The loss is used to further update the meta-parameters θ; this step is also known as external update. In meta-training, we need to optimize a large number of meta-tasks. At this point, our goal is...

[0123]

[0124] The learning rate for external updates is β, and the formula for performing a single external update is:

[0125]

[0126] Where g i It is a meta-task T i The meta-gradient on can be extended to

[0127]

[0128] As can be seen from the above equation, we have introduced a higher-order gradient into the calculation of the meta-gradient. To reduce computational cost, we employ a first-order approximation algorithm.

[0129] S4: Using the target event data obtained from S3 Fine-tuning is performed to obtain the optimal model.

[0130] For the new event data D T According to D T Create a test task that supports random selection of sets and query sets for a test task T. i′ Using T i′ Support sets Perform n fine-tuning steps:

[0131] θ′ j =U n (θ * ;γ)

[0132] Where, θ′ j This indicates the result after fine-tuning using the target event data. The parameters, U n Indicates to loss Gradient descent n times, θ * express The parameter γ represents the learning rate.

[0133] For T i′ querysets used Verification is carried out, and when the prediction accuracy reaches the preset threshold, the next step is performed, otherwise the Returning to S3 as the basic detection model continues training;

[0134] S5: Detection of target field posts to be detected: for a post to be detected, input the optimal model obtained by S4 Output the predicted label of the post to be detected.

[0135] Specifically, in S4, the S3-derived When fine-tuning, if the number of target field samples is zero, the false

[0136] Information

[0137] Zero-sample detection of information is an extreme case of few-sample detection, that is, there is no labeled data for the target event. This detection is of great significance because we need to quickly detect false information when a new event occurs. First, weak labels are assigned according to the emotion sequence of the post, then weights are assigned to each post to reduce the influence of noise, and finally the weak label data is fine-tuned using these weak label data.

[0138] Generate weak labels for posts according to the emotion sequence of the new event: if the emotion sequence of the post is closer to the emotion prototype of the real post, the weak label of the post is real, and vice versa, the formula is

[0139]

[0140] Where d Real represents the emotion prototype of the real post, and d Fake represents the emotion prototype of the false post;

[0141] During fine-tuning, different weights should be assigned to different weak samples. Specifically, if the emotion sequence is closer to its weak label prototype, it can be considered that the confidence of the weak label is stronger, and the weight should be larger. The weight of the weak sample is

[0142]

[0143] Where τ is the emotion threshold, and represents the learning rate γ of the weak sample and the weak sample weight Gradient descent n times; for example, for weak sample x i The formula for applying single gradient descent is:

[0144]

[0145] Meta-training remains unchanged in the zero-sample scenario, and meta-testing is:

[0146] ​

[0147] Specifically, the basic detection model in S3 is composed of an emotion feature extractor and a semantic feature extractor. The model first maps the post text to an emotion-semantic space, and then performs fake information detection.

[0148] S31: Emotion feature extractor

[0149] The emotional features of the post are essential for early detection. We constructed an emotion sequence for each post according to prior knowledge. The content of the emotion sequence is shown in Table 1. The sequence vector is calculated by the length-weighted term frequency of the post.

[0150] According to prior knowledge, an emotion sequence is constructed for each post. The post emotion sequence e is captured as a 24-dimensional vector through the emotion sequence feature, and then the emotion feature f is obtained through an MLP layer. e = MLP(e);

[0151] S32: Semantic feature extractor

[0152] BERT is used as a semantic feature extractor, and the embedding of “[CLS]” in the last layer of BERT is used to represent the semantic embedding f t ;

[0153] S33: The final embedding of the post is the combination of the emotion embedding and the semantic embedding:

[0154] f = f e + f t

[0155] Then, a linear classification detector with softmax is used to detect fake information:

[0156]

[0157] Specifically, when the emotion perception meta-task is trained by inputting the basic detection model, the loss in S3 uses the cross-entropy loss function L C :

[0158]

[0159] where y represents the true label of the post, and f represents the predicted label of the post.

[0160] When the event adaptation meta-task training is trained by inputting the basic detection model, the loss in S3 uses the MMD of two events as the loss function L e :

[0161]

[0162] Specifically, the S3 trains the basic detection model to obtain In the process, the event adaptive meta-task only updates the parameters in the emotion feature extractor and the semantic feature extractor in the basic detection model, and the emotion perception meta-task is used to update the parameters in the emotion feature extractor, the semantic feature extractor and the linear classification detector.

[0163] Experimental verification

[0164] 1. Dataset

[0165] Our experiments are based on five real Twitter event datasets PHEME, which contain five specific events: Ferguson unrest, Ottawa shooting, Sydney siege, Charlie Hebdo shooting, and Germanwings plane crash. The authors collected tweets related to newsworthy events that may trigger the initiation and spread of misinformation from the Twitter streaming API. Detailed statistics of the dataset are shown in Table 1.

[0166] Table 1: Dataset details

[0167]

[0168] 2. Implementation details

[0169] We implemented all deep learning baselines and EML (the method of the present application is referred to as EML) using NVIDIA GeForce RTX 3090 and PyTorch 1.9.0. We selected q = 5 similar and opposite samples in each class to construct the emotion perception meta-task, and p = 16 samples in each event to construct the event adaptive meta-task. The hyperparameter η is set to 0.5, the balancing factor λ is set to 1.5, the update step n is set to 2, the number of sampled meta-tasks used for one outer update is set to 32, and the emotion threshold τ is set to 0.5. For the optimizer of the inner and outer updates, we use Adam with a learning rate of 5e-5, while for the gradient update during meta-testing, we set the learning rate to 1e-5. In meta-training, we perform one outer update using all meta-tasks; that is, the number of times the meta-model is updated is the number of meta-tasks. In meta-validation, we fine-tune until the gradient of the model parameters is less than 0.05. To ensure the effectiveness of the test results, we evaluate the test performance based on 100 test sets and report the average performance. In our experiments, we take the current event as the target event and the data of the other four events as the source data.

[0170] 3. Baseline methods

[0171] Unless otherwise stated, baseline methods are trained on source data and then adapted to target data. To demonstrate the effectiveness of EML, we choose the following competitive baselines for comparison:

[0172] (1) Traditional machine learning methods. We implement various machine learning algorithms, including decision tree (DT), Naive Bayes (NB), and support vector machine (SVM). They are based on TF-IDF, as it is a commonly used weighting technique in information retrieval and data mining. It measures the importance of a word in a text and is often used as a text feature representation.

[0173] (2) Base_Model. It directly uses the support set of the target task for fine-tuning and validates on its query set. It reflects the performance of the basic detection model on a small amount of samples without the need for additional data.

[0174] (3) Base-Pretrain. It also uses the base detection model, which uses the source data pre-trained model and then fine-tunes it on the target event.

[0175] (4) BERT-pro. The prototype network creates a prototype for each class and then classifies it by comparing the similarity of the input to the prototype. Its feature is the general classification ability, which can not only be applied to the current event data. To ensure the fairness of the experiment, we use BERT as the embedding representation of the prototype network.

[0176] (5) MLADA. A meta-learning adversarial network for few-shot text classification. It can identify important lexical features and generate high-quality sentence embeddings to new domains.

[0177] (6) DAFD. Early detection model for misinformation in new domains. It applies domain adaptation techniques to align the data distribution of the two domains during pre-training and generates adversarial examples during fine-tuning to enhance the robustness and generalization ability of the model. Since DAFD requires source domain data for an event, we use four source event data in turn for pre-training and use the best result as the final result.

[0178] 4. Results comparison

[0179] In our experiments, we have 5-shot and 10-shot settings. 5-shot means that there are only 5 samples (real and fake) for each class in the data used for fine-tuning, which is a widely used setting in few-shot learning scenarios. Table 2 shows the performance of different methods on the dataset. It can be seen that EML achieves the best results on all indicators in 5-shot and 10-shot settings.

[0180] In the 5-shot setting, traditional machine learning algorithms always get the worst results, which are much worse than deep learning algorithms. Among these algorithms, SVM almost achieves the best results, while decision tree performs poorly. We conjecture that constructing a hyperplane for classification is more suitable for the current scenario. For deep learning algorithms, the performance of Base_Model is always the worst because it is directly fine-tuned on a few samples and is prone to overfitting, thus failing to learn meaningful embedding representations. The performance of BERT-pro varies significantly between different event datasets. We conjecture that this is because the prototype network is highly data-sensitive. In the case of few samples, different data has a greater impact on the construction of prototypes, so the overall stability of the model is poor. In most cases, the effect of the base pre-training is not as good as DAFD because it is only pre-trained on the source data and then directly fine-tuned on the target data. After all, the two parts of data have different distributions, and the data of the target event is very small, so the result is suboptimal. EML achieves the best results on all indicators for all datasets.

[0181] In the 10-shot setting, we can see similar results, where traditional machine learning algorithms achieve the worst results, indicating that this method is not good at handling few-shot learning problems. EML is still the best, indicating that EML performs well in handling few-shot detection of new events.

[0182] Table 2: Performance comparison of detection models on each event dataset in 5 and 10-shot settings

[0183]

[0184] 5. Ablation experiments

[0185] We also explore the effectiveness of the EML overall framework, weighted meta-learning, and event-adaptive meta-tasks using several variants of EML to explore the effectiveness of these strategies:

[0186] EML w / o-d: A variant of EML that does not have event-adaptive meta-tasks in meta-training, only using emotion perception meta-tasks for meta-training.

[0187] EML w / o-w: A variant of EML that does not have a weighted meta-learning step, so all meta-tasks are treated equally in meta-training.

[0188] EML w / o-d / w: A variant of EML that does not have weighted meta-learning and event-adaptive meta-tasks, only using emotion perception meta-tasks in meta-training and treating them equally.

[0189] EML w / o-all (i.e., basic detection model): EML variant without all components and meta-training steps, in this case, it degenerates to the basic detection model.

[0190] We conduct ablation experiments on these variants in the 5-shot setting, and the average results of five events are shown in Figure 5 From the results, we can get the following points.

[0191] Without weighted meta-learning, the performance of the model on each dataset decreases, which shows that it can enhance the robustness and generalization ability of the model to new events. Without the event adaptive meta-task, a similar phenomenon can be observed, and the performance of the model on each dataset decreases. This also shows that our analysis is reasonable, and combining the event adaptive meta-task in meta-training can improve the generalization ability of the model. Without the meta-learning process, the performance of the model is the worst, and the model cannot work correctly in the few-shot scenario.

[0192] Through the component analysis of EML, we can know that: (1) both the event adaptive meta-task and the weighted meta-learning component contribute to the improvement of fake news detection performance; (2) the emotion-aware meta-learning framework is very important, as it can make good use of source data and adapt to new event data more effectively.

[0193] 6. Analysis of zero-shot detection results

[0194] We first weakly label the target event data according to the emotion prototypes of the source data. When we choose an event as the target event, the data of the remaining four events are all source data, so there is no problem of data leakage. The accuracy of the effective weak label is shown in Table 3.

[0195] Table 3: Weak label accuracy of each event

[0196]

[0197] Since the weak label has noise, we calculate the weight corresponding to each weak sample, and the effective weak label refers to the sample with a weight of 0. From Table 3, we can see that our weak label accuracy reaches 64.22 on average over multiple events, and 69.39 on the Charlie Hebdo data. This is much higher than random labeling, proving that our idea is correct; posts with the same emotion are more likely to have the same label. Without new event data, applying the emotional features of the post can better assist detection.

[0198] After our weak labeling, all models can be used for new events, and we use the following baselines to evaluate EML:

[0199] Base Model: The basic detection model is directly trained and tested on the new event data, and the performance of the model is verified without other data.

[0200] Base-Pretrain: We pretrain the basic detection model using source data, and then fine-tune it using weak samples from the target event.

[0201] BERT-pro: The prototype network is usually used in zero-shot scenarios, and we construct prototypes using source data and then directly test on new event samples.

[0202] EML: The meta-training process of EML remains unchanged, and it is fine-tuned using weak samples from the new event.

[0203] We select 32 samples in each category for training, and the average results are shown in Table 4. We can draw the following conclusions:

[0204] (1) Most models achieve more competitive results with the support of weak samples, which illustrates the necessity of weak samples in the zero-shot scenario.

[0205] (2) The proposed model EML achieves the best performance in most indicators, indicating that EML can also effectively handle zero-shot detection.

[0206] (3) Events with high annotation accuracy in weak samples can achieve better results. Without labeled data, our method achieves an accuracy of 80.94% on Charlie Hebdo.

[0207] Table 4: Performance comparison of detection models on each event dataset in the 0-shot setting

[0208]

[0209] We also conduct ablation experiments to demonstrate the necessity of weak sample generation. When EML is meta-trained, we compare the results of directly testing on target event data and fine-tuning using weak samples before testing. The results are shown in Figure 4 .

[0210] The results show that directly testing the meta-trained model on the target event does not work well. This is because the purpose of meta-learning is to learn the initialization parameters that can quickly adapt to new event data. The current parameters are not the extreme points of the loss function, but they can quickly reach the extreme points after a small amount of fine-tuning, therefore, the generation of weak samples is essential for zero-shot detection.

[0211] We use the t-SNE algorithm to visualize the emotion sequences of new events and the emotion prototypes of source data. Real posts in new events are closer to the real post prototypes of source domains, and vice versa, which indicates that the emotion features between different events are universal and the effectiveness of weakly labeled using emotion sequences.

[0212] 7. Impact of hyperparameters

[0213] We further explore the impact of two important hyperparameters in EML: the emotion threshold τ and the selection task proportion parameter ρ. The experimental results are shown in Figure 5

[0214] For the emotion threshold τ, as the value of τ increases, the effect of EML first increases and then decreases. The best effect is when τ = 0.5. This is because the weakly labeled samples have noise. When τ is small, EML considers a large amount of noise, which easily makes EML unable to learn effective knowledge. When τ is large, there are fewer available samples, and EML also cannot learn effective knowledge.

[0215] For the hyperparameter ρ, as ρ increases, the impact of EML also presents the same trend. When ρ is small, the model uses a large number of event adaptation tasks in meta-training, which may make it difficult to learn knowledge for detecting false information. When ρ is large, especially when ρ = 1, EML will degenerate into not using event adaptation meta-tasks, and the model cannot learn event-invariant representations well.

[0216] Finally, it should be pointed out that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.​

Claims

1. A cross-event fake news detection method based on emotion-aware meta-learning, characterized in that: Comprising the following steps: S1: constructing emotion perception sub-tasks and event adaptive sub-tasks: S11: constructing emotion perception sub-tasks: Let K denote the set of events of D S , M e {Real, Fake} is the set of labels, is the set of all samples in D S with event k e K and label m e M, is the set of all samples in D S with event k e K and label -m e M, and let be the set of queries for the ith meta-task T i be the support set for T i be the set of q samples in D i closest to x in terms of sentiment sequence, and be the set of q samples in D i most opposite to x in terms of sentiment sequence. Set There are N samples, that is, there are N emotional perception meta-tasks in total, the i-th meta-task T i is expressed as Use e i to represent the sequence of emotions for x i , and measure the emotional similarity between x i and x j using the following formula: where j e {1,2,..., |D k |} and i≠j; S12: constructing event adaptive sub-tasks: The definition of the MMD distance is computed by a specific representation of the Gaussian kernel function, φ(·), which acts on the source event data and the target event data An empirical approximation of the MMD is: wherein, represents the input of the last MLP of the base detection model, i.e. the embedding of the post; S2: weighted meta-learning on emotion perception sub-tasks S21: For each The calculation formula of the corresponding class prototype The calculation formula is: where ε(x i ) is a text embedding of x i , and the semantic weight is computed as: where the cos function is the cosine similarity function, η is a hyperparameter, and 1(·) is an indicator function, represents a sample of event k, and K represents the number of source domain events, represents the rest of the event prototypes that are not the current sample. S22: Construct an emotion prototype d for the real information and false information in the source data m The prototype representation is obtained by averaging the emotion sequences of all corresponding samples: where D m is the set of all samples in D s with label m, the mood weight is calculated as wherein d m represents the same emotion prototype as the post tag, d -m represents the opposite emotion prototype as the post tag; T i The weight w i is: Wherein λ is a balance factor; S3: obtaining a preliminary detection model M by training the basic detection model θ* : Input the event adaptive sub-tasks and the emotion perception sub-tasks after S2 into the basic detection model for training, and update the parameters of the basic detection model according to the loss until the loss no longer changes to obtain a preliminary detection model; In each iteration of training the basic detection model, select emotion perception sub-tasks with a probability ρ∈(0,1) and event adaptive sub-tasks with a probability of 1-ρ, and ρ is a hyperparameter; S4: fine-tune the optimal model using the target event data ​ For new event data D T , create test tasks according to D T , where the support set and query set are randomly selected, for a test task T i′ , use the support set of T i′ to fine-tune n times on : θ' j = U n (θ * ; γ) where θ′ j denotes the parameters of the target event data fine-tuned model, U n denotes the parameters of the target event data loss function gradient descent n times, θ * denotes the parameters of the target event data learning rate; Query set for T i′ Validation is performed when the prediction accuracy reaches a preset threshold, otherwise the basic detection model is returned to S3 for further training.​ S5: detection of target domain posts to be detected: For a post to be detected, input the post to be detected into S4 to obtain the optimal model Output the predicted label of the post to be detected.

2. The emotion-aware meta-learning based cross-event fake news detection method of claim 1, wherein: The S4 in the above formula is obtained by fine-tuning the S3 If the number of target domain samples is zero, first assign weak labels according to the emotion sequence of the post, then assign weights to each post, and finally fine-tune the S3 using these weak label data ; According to the emotion sequence of the new event to generate weak labels for the posts: if the emotion sequence of the post is closer to the emotion prototype of the real post, then the weak label of the post is real, and vice versa, the formula is wherein d Real represents the emotion prototype of a real post, d Fake represents the emotion prototype of a fake post; If the emotion sequence is closer to its weak label prototype, it can be considered that the confidence of the weak label is stronger, and the weight should be larger, and the weight of the weak sample is where τ is the emotion threshold, using denotes the learning rate γ and the weak sample weight performing gradient descent n times; 3. The emotion-aware meta-learning based cross-event fake news detection method according to claim 1 or 2, characterized in that: The basic detection model in S3 is composed of two parts, an emotion feature extractor and a semantic feature extractor, S31: emotion feature extractor According to prior knowledge, an emotion sequence is constructed for each post, the post emotion sequence e is captured as a 24-dimensional vector through emotion sequence features, and then emotion features f are obtained through an MLP layer e = MLP(e); S32: semantic feature extractor Using BERT as semantic feature extractor, the embedding of "[CLS]" in the last layer of BERT is used to represent semantic embedding f t ; S33: the final embedding of the post is the combination of the emotion embedding and the semantic embedding: f = f e + f t Then detect the false information through a linear classification detector with softmax:

4. The emotion-aware meta-learning based cross-event fake news detection method of claim 3, wherein: When the input basic detection model is trained for the emotion perception meta-task, the loss in S3 adopts a cross-entropy loss function L C Calculation: where y represents the true label of the post, represents the predicted label of the post; When the input base detection model is trained by event adaptive meta-task training, the loss in S3 adopts the MMD of two events as the loss function L e Calculation:

5. The emotion-aware meta-learning based cross-event fake news detection method of claim 3, wherein: The S3 trains the basic detection model to obtain In the process of updating the basic detection model, the event adaptive meta-task only updates parameters in the emotion feature extractor and the semantic feature extractor in the basic detection model, and the emotion perception meta-task is used to update parameters in the emotion feature extractor, the semantic feature extractor, and the linear classification detector.

Images