Learning to Detect Few-Shot-Few-Clue Misinformation

Misinformation is created to mislead the public and can lead to severe consequences if widely propagated. Therefore, early detection is crucial. Convolutional neural networks (CNN) are used to extract interaction features between temporally sequential posts. Nguyen et al. (2017) propose another CNN-based model to learn latent representations of tweets and predict event veracity by aggregating predictions of all related statements.  Liu et al. (2018) find that only a part of comments on social platforms are helpful to classify whether a statement is fake. Then they design an attention-based detection model to evaluate relative importance of comments according to their attention values. They also conclude that the attention mechanism contributes to early misinformation detection.  Liu and Wu (2018) use recurrent and convolutional networks to extract propagation features while  Shu et al. (2020) leverage weak social supervision from multiple sources for early detection.

Misinformation detection models may not generalize well across topics and cultures (Horne et al., 2020), therefore transferable patterns have attracted researchers’ interests.  Potthast et al. (2018) have analyzed the similarity of the writing style of hyperpartisan news and its connection to misinformation.  Castelo et al. (2019) consider the changing discourse of emerging events, and propose a topic-agnostic approach that uses linguistic and web-markup features to identify misinformation.  Huang et al. (2020) point out that the word/topic distribution of articles from different media sources may be different, so they develop a synthetic network that focuses on function words and syntactic structures for a more generalized representation.  Wang et al. (2018) propose an event adversarial neural network where an event discriminator component aims to remove event-specific features. All these methods aim for features, while our work aims for model parameters that allow for strong generalization.

Fast learning is a hallmark of human intelligence. Due to high efficiency of recognizing shared patterns, meta-learning has been drawing increasing research attention (Lemke et al., 2015). There are different approaches to meta-learning: metric learning-based, optimization-based and probabilistic. The first approach to meta-learning is based on metric learning in the embedding space.  Vinyals et al. (2016)propose to match the embeddings of query examples with those of the support ones.  Zhou et al. (2018) propose to learn a concept space where the concept matching of categories is conducted. One optimization-based work in (Finn et al., 2017) is model-agnostic meta-learning (MAML) algorithms that learns a proper model initialization. Through a few steps of gradient descent-based fine-tuning, the initialization can be adapted to different tasks. From the Bayesian point of view, probabilistic meta-learning is studied in (Ravi and Beatson, 2018),where the initialization is treated as the prior of model parameter distributions. The posterior distributions are derived from the dataset of each learning task. We customize meta-learning by using topic similarity to adaptively control how the meta parameter will initialize the topic-specific parameter. One concurrent work (Wang et al., [n.d.]) combines MAML and neural processes to deal with misinformation. We differ from this work from two aspects: (1) the proposed task-aware Bayesian meta-learning algorithm can provide uncertainty estimation that is infeasible for optimization-based meta-learning including MAML, and is less prone to be under-fitting than neural processes; and (2) we use topic similarity when adapting global knowledge to tasks, which addresses the concern of dissimilar task distributions in (Wang et al., [n.d.]).

Annotating statement veracity is expensive and time-consuming. This leads to the challenge of how to improve efficiency of annotated data. Section 3 discusses the few-shot classification task for a few labeled examples per category, the key of which is to properly recognize transferable patterns among different tasks, e.g., linguistic features and user stances (Potthast et al., 2018; Wang et al., 2018). Thus, the challenge of data efficiency can be tackled by few-shot learning. From this viewpoint, we formulate misinformation detection as a few-shot learning problem. As usual, the concept “few” indicates a specific number such as 5. For a statement $\boldsymbol{s}$ and clues $\boldsymbol{c}$, the veracity $y$ is predicted by a model parameterized by $\boldsymbol{\phi}$. The model parameter $\boldsymbol{\phi}$ for different tasks shares the same prior $\boldsymbol{\theta}$. To improve data efficiency and convenient adaptation to unseen events/topics, misinformation detection models need to extract salient patterns, i.e., $\boldsymbol{\theta}$, that can be transferred across topics and events.

Besides the limited annotations, early misinformation detection faces a new challenge that has not been considered in previous studies. At the early stage of misinformation emerging on social media, there is limited evidential content. Therefore, it is important to study the shared structure of how to utilize limited evidence.

Inspired by few-shot learning, we formulate the early detection of online misinformation as a new few-shot-few-clue learning framework. As the input of each datum is made up of a statement and its evidential clues, the concept of few-shot-few-clue describes scenarios of limited size of data in two levels: (1) the number of annotated data (consisting of labels, statements and their accompanied clues) and (2) the number of clues per statement. Like few-shot misinformation detection, the concept “few” of few-clue indicates a specific number like 5-clue misinformation detection. Few-shot-few-clue learning differs from conventional few-shot learning because it considers fewness at two levels: shot and clue inside each shot.

Figure 2 is a graphical representation of variables of interest. Similar to few-shot misinformation detection, the statement veracity $y$ is decided by a detection model with: (1) the parameter $\boldsymbol{\phi}$ and (2) the input of the statement content $\boldsymbol{s}$ and the corresponding clues $\boldsymbol{c}$. What makes few-shot-few-clue learning different is the decomposition of the clue variable. The variable $\boldsymbol{c}_{m}^{\tau}$ is decomposed into a set of clues $\boldsymbol{c}_{m,n}^{\tau}$ with the set size being $N_{m}^{\tau}$. This is in line with our interest in cases when there are limited available clues per statement during the early stages, i.e.,the effect of $N_{m}^{\tau}$. Formally, we give definitions of relevant terms with reference to Figure 2.

We make three further declarations: (1) misinformation about different events/topics is mapped to different tasks; (2) a statement is in the form of text; (3) a clue is an user engagement. Based on the above definitions, the research problem of few-shot-few-clue applies when there are few statements (small $M^{\tau}$) and few clues per statement (small $N_{m}^{\tau}$) for the early detection of misinformation regarding.

The meta-parameter $\boldsymbol{\theta}$ can be interpreted as the common patterns shared by detection models across different topics; while the parameter $\boldsymbol{\phi}$ can be interpreted as features specific to certain topics. For early detection when the size of $\mathcal{D}^{\mathcal{T}}$ is limited, machine learning models usually perform poorly on $\tilde{\mathcal{D}}^{\mathcal{T}+1}$; the meta-knowledge contained in $\boldsymbol{\theta}$ can be helpful.

To efficiently learn the model parameters, we design a task-aware Bayesian meta-learning algorithm. One concern of meta-learning’s application to misinformation detection is that tasks might be dissimilar and meta-parameters are not transferable. To address this concern, we introduce topic similarity $Sim(\cdot,\cdot)$ to compute semantic similarity between topics, which determines how much we rely on the meta-parameter $\boldsymbol{\theta}$ to learn the task-specific parameter $\boldsymbol{\phi}^{\tau}$. We expect $\boldsymbol{\phi}^{\tau}$ depends heavily on $\boldsymbol{\theta}$ when a topic $\tau$ is similar to topics in the meta-training set.

Specifically, we compute a topic embedding $\boldsymbol{E}^{\tau}$ using a word embedding $\boldsymbol{e}_{i}$ (based on Word2Vec (Mikolov et al., 2013)) and its weight $w_{i}$ (the normalized frequency of a word in the topic), where $i$ indicates the $i$-th word in the vocabulary $V$:

For all topics in the meta-training set, we compute an embedding:

where $|\mathcal{D}^{\tau}|$ and $|\mathcal{D}_{\rm{meta-train}}|$ are the sizes of the dataset $\mathcal{D}^{\tau}$ and the meta-training set respectively. To avoid overfitting of $\boldsymbol{\phi}^{\tau}$ for a topic $\tau$, we introduce an adaptive parameter $z^{\tau}$ based on topic similarity:

where $\rm{cos}(\cdot,\cdot)$ is the cosine function. Then we use $z^{\tau}$ to determine the initialization of the task-specified model parameter $\boldsymbol{\phi}^{\tau}$:

where $\boldsymbol{\theta}_{\rm{rand}}$ is randomly sampled from a normal distribution. When $\tau$ is more similar to meta-training set, $\boldsymbol{\phi}^{\tau}_{\rm{init}}$ depends more on $\boldsymbol{\theta}$ learnt from the meta-training set, otherwise there will be more randomness brought by $\boldsymbol{\theta}_{\rm{rand}}$.

The task-aware Bayesian meta-learning algorithm can provide two benefits: (1) uncertainty estimation that is infeasible for alternative optimization-based meta-learning including MAML, and is less prone to be under-fitting than neural processes; and (2) topic similarity when adapting global knowledge to tasks, which addresses the concern of dissimilar task distributions in (Wang et al., [n.d.]).

To optimize the task-aware Bayesian meta-learning algorithm with variational inference, we first need to derive the Evidence Lower BOund (ELBO) of the data likelihood:

Derivation of Eq. 10 can be found in Appendix A. Here, $\psi$ and $\{\boldsymbol{\lambda}^{\tau}|\tau=1,\ldots,\mathcal{T}\}$ are the variational parameters (such as a mean and standard deviation of each weight) of the approximate posteriors over the global latent variables $\boldsymbol{\theta}$ and the local latent variables $\boldsymbol{\phi}^{\tau}$. If we maintain distinct variational parameters $\boldsymbol{\lambda}^{\tau}$, each of which indexes a distribution over task-specific model parameters $q(\boldsymbol{\phi}^{\tau};\boldsymbol{\lambda}^{\tau})$, the number of variational parameters is in the scale of $O(1+\mathcal{T})$.

To scale the optimization process, we adopt the Amortized Variational Inference (AVI) for meta-learning proposed by Ravi and Beatson (2018) to infer the posterior distributions of $\boldsymbol{\phi}^{\tau}$ and $\boldsymbol{\theta}$. As shown by Model-Agnostic Meta-Learning (Finn et al., 2017), a proper initialization can produce the decent task-specific parameters. Hence, from a global initialization $\boldsymbol{\theta}$, we produce the variational parameter $\boldsymbol{\lambda}^{\tau}$ by conducting several steps of gradient descent.

First, we need to specify the loss function on $\mathcal{D}^{\tau}=\left\{{S}^{\tau},{C}^{\tau},{Y}^{\tau}\right\}$,

Then, we modify the procedure of stochastic gradient descent, $SGD_{K^{\tau}}(\mathcal{D}^{\tau},\boldsymbol{\theta})$, based on the adaptive parameter $z^{\tau}$, to produce $\boldsymbol{\lambda}^{\tau}$ from the the global initialization $\boldsymbol{\theta}$:

$\alpha$ is the learning rate and $K^{\tau}$ is the number of steps of gradient descents that is determined by the adaptive parameter $z^{\tau}$. $K_{\text{min}}$ and $K_{\text{max}}$ denote the lower and upper bounds of $K^{\tau}$, and $\lceil\cdot\rceil$ is a ceil function. Hereby $q(\boldsymbol{\phi}^{\tau};\boldsymbol{\lambda}^{\tau})=q(\boldsymbol{\phi}^{\tau};SGD_{K^{\tau}}(\mathcal{D}_{t}^{\tau},\boldsymbol{\theta}))=q(\boldsymbol{\phi}^{\tau}|\mathcal{D}^{\tau},\boldsymbol{\theta})$. With this form of the variational distribution, $\boldsymbol{\theta}$ serves as both the global initialization of local variational parameters $\boldsymbol{\lambda}$ and the parameters of the prior $p(\boldsymbol{\phi}|\boldsymbol{\theta})$.

Finally, we estimate $\boldsymbol{\theta}$. We let $q(\boldsymbol{\theta};\psi)$ be a dirac delta function $q(\boldsymbol{\theta})=\mathbbm{l}\{\boldsymbol{\theta}=\boldsymbol{\theta}^{\star}\}$, where $\boldsymbol{\theta}^{\star}$ is the solution to the optimization problem. This conduction removes the need for global variational parameters $\psi$ and simplifies the optimization problem of Eq. 10 to

Algorithm 1 provides the pseudo code for the task-aware Bayesian meta-training algorithm.

We compare the TABML algorithm and the following state-of-the-art models. RFC (Kwon et al., 2017) is a random forest with features from the source tweets and engaged user profiles. CRNN (Liu and Wu, 2018) combines convolutional and recurrent neural networks to extract features from engaged users and retweet texts. CSI (Ruchansky et al., 2017) incorporates relevant articles and analyses group behaviour of engaged users. dEFEND (Shu et al., 2019) uses a co-attention mechanism to study source claims and user features. GCAN (Lu and Li, 2020) utilises the co-attention graph networks to encapsulate the propagation structure of heterogeneous data. EANN (Wang et al., 2018) contains a event discriminator that removes event-specific features and discovers the common pattern among events. MetaFEND (Wang et al., [n.d.]) is a meta neural process that combines meta-learning and neural processes to detect few-shot misinformation.

We first examine the effectiveness of Twitter-LDA (Diao et al., 2012) on determining the topic cluster that each tweet belongs to. We present the results of topic clustering in the form of word cloud on Twitter15 and Twitter16 in Figure 5. From this figure, it is clear that each topic cluster focuses well on a social event. For instance, the first topic of Twitter15 is about the “MH17” accident of Malaysia Airline, and the second topic of Twitter16 is more about politics. Moreover, there are sill overlap words between clusters, which indicates similarity in terms of semantics. This further verifies our assumption of similar tasks and serves as the basis of using semantic similarity to modify the Bayesian meta-learning based misinformation detection (that is, Equation 9).

The Pheme dataset has clearly split statement tweets into various events, hence there is no need to do topic clustering. These events are “Charlie Hebdo”, “Sydney siege”, “Ferguson”, “Ottawa shooting”, “Germanwings-crash”, “Putin missing”, “Prince Toronto”, “Gurlit” and “Ebola Essien”.

Subsequently, we aim to evaluate the effectiveness of the proposed task-aware Bayesian meta-learning algorithm with all available statements and clues. Table 3 summarizes the four evaluation metrics of all models on the three datasets. The reported numbers in Table 3 are the means and standard deviations of TABML model running with different weights sampled from the same distribution that is learned at the meta-training phase. Our model outperforms all the baselines on most of all the metrics across the three datasets, achieving a performance improvement of around 4% in terms of the accuracy. As for recall and precision, our model obtains the best values in most cases or the second best with a fairly minor gap in other cases. On the larger Pheme dataset where state-of-the-art GCAN has severely discounted performance, our model enjoys a more significant improvement. Compared to the baseline with transferable patterns, i.e., EANN and MetaFEND, our model outperforms them in terms of all metrics over three datasets. Therefore, our model is able to generalize better to new topics.

Annotating statement veracity is expensive and time-consuming. To understand the effect of the number of training statements, we conduct experiments using a different number of annotated statements. Specifically, we use five different percentages (i.e., 60%, 70%, 80%, 90% and 100%) of all training data to train the models. From each topic cluster in the dataset Twitter15 and Twitter16, we still randomly sample 6 true and 6 false shots to make a batch, with each shot containing all available clues. So it is still 2-way-6-shot-30-clue learning on Twitter15 and Twitter16. In the same way, it is 2-way-6-shot-25-clue learning on Pheme. Evaluation performance is reported in Figure 6. We have three observations: First, it is clear that the size of training dataset exerts a significant effect on the performance of all models. Specifically, the performance increases as there are more training data in most cases. Therefore, the most effective way to improve misinformation detection model is by increasing the number of annotated statements. Second, our TABML is able to outperform the state-of-the-art baselines with different numbers of shots. Third, our TABML gains more stable performance improvement compared with baselines when more training data are available.

Next, we study early misinformation detection when there are a limited number of clues per statement. From each topic cluster, we fix the number of shots at 6. Then we change the number of clues $x$ in each shot, where $x$ can be [5, 10, 15, 20, 25, 30] and [5, 10, 15, 20, 25] for Twitter datasets and Pheme dataset respectively. Figure 7 illustrates how the detection performance changes over different numbers of clues. Obviously, our proposed TABML consistently shows better performance when varying the clues. Another obvious finding is that, despite of the overall increasing accuracy with more clues, performance is improvement with different pace. This indicates that some clues hinder veracity prediction, and should be excluded. We leave this in the future work. Baseline models show greater variances, which means instability and lower performance, while our TABML has more stability. In fact, Figure 7 shows TABML is stable enough to deal with few-clue problems.

Finally, we study how the adaptive parameter $\boldsymbol{z}$ affects the performance of our model. Figure 8 shows the comparative result. As we can see, our TABML model with the adaptive parameter $\boldsymbol{z}$ always outperforms the counterpart without $\boldsymbol{z}$ across three datasets no matter what metric it is, leading to $10\%\sim 30\%$ improvements. The result illustrates that TABML with $\boldsymbol{z}$ can adaptively learn about the meta information among the topics and then effectively apply them into the learning of new tasks, avoiding the phenomenon of overfitting and the reduced performance in the learning of new tasks, which always contain the out of distributional training data.