Identifying Conspiracy Theories News based on Event Relation Graph

We need to identify events as well as extract the four types of event relations to create an event relation graph for each individual article. We will describe the training process and the graph construction process separately.

For training, we use the annotations from the general-domain MAVEN-ERE dataset Wang et al. (2022). First, we train the event identifier, where the Longformer Beltagy et al. (2020) language model is used to encode the entire article and a classification head is built on top of the word embeddings to predict whether the word triggers an event or not. Next, we train four event relation extractors. Following the previous work Zhang et al. (2021); Yao et al. (2020), we form the training event pairs in the natural textual order, where the former event in the pair is the precedent event mentioned in text. Regarding temporal relations ²²2Time expressions such as date or time are also annotated in temporal relations. However, our event relation graph focuses on events, so we exclude temporal annotations related to time expressions and solely retain annotations between events., we process the before annotation as follows: keep the label as before if the annotated event pair aligns with the natural textual order, or assign after label to the reverse pair if not. The simultaneous, overlap, begins-on, ends-on, contains annotations are grouped into the overlap category in our event relation graph. Regarding causal relations, we assign the causes label if the natural textual order is followed, and assign caused by label otherwise. Similarly for subevent relations, we assign the contains label if the natural textual order is followed, and assign contained by label otherwise. Since the events and four event relations interact with each other to form a cohesive narrative representation, we adopt the joint learning framework from Wang et al. (2022) to train these components collaboratively.

During the event relation graph construction process, the initial step is event identification. Given a candidate news article that consists of $N$ words, the trained event identifier generates the predicted probability for each word:

$$P_{i}^{event}=(p_{i}^{non-event},p_{i}^{event})$$

(1)

where $i=1,...,N$, and $p_{i}^{event}$ is the probability of $i$-th word being an event. With the events identified, the next step is to extract and classify the four types of relations. Given a pair of predicted events $(event_{i},event_{j})$, the trained coreference identifier predicts the probability for coreference relation as:

$$P_{i,j}^{corefer}=(p_{i,j}^{non-corefer},p_{i,j}^{corefer})$$

(2)

where $p_{i,j}^{corefer}$ is the probability of $i$-th and $j$-th events corefer with each other. We also utilize the trained temporal classifier to generate the predicted probability for temporal relation:

$$P_{i,j}^{temp}=(p_{i,j}^{non-temp},p_{i,j}^{before},p_{i,j}^{after},p_{i,j}^{overlap})$$

(3)

where $p_{i,j}^{non-temp}$ denotes the probability of no temporal relation between the $i$-th and $j$-th events, and $p_{i,j}^{before},p_{i,j}^{after},p_{i,j}^{overlap}$ represents the probability of before, after, overlap relations respectively. Similarly, the trained causal classifier and subevent classifier generate the predicted probability as:

$$P_{i,j}^{causal}=(p_{i,j}^{non-causal},p_{i,j}^{cause},p_{i,j}^{caused-by})$$

(4)

$$P_{i,j}^{subevent}=(p_{i,j}^{non-subevent},p_{i,j}^{contain},p_{i,j}^{contained-by})$$

(5)

Finally, the predicted probabilities in eq(1)-(5) are incorporated as the soft labels for events and event relations in the constructed event relation graph. Accordingly, the hard label for each element is obtained by applying the $argmax$ function to the predicted probabilities.

To incorporate the constructed event relation graph into conspiracy theories identification, we first leverage the soft labels to train an event-aware language model, with the knowledge of events and event relations augmented. The soft labels provide fuzzy probabilistic features and enable the basic language model to be aware that nodes represent events and links represent event relations.

Considering the news articles are typically long, we utilize Longformer as the basic language model Beltagy et al. (2020). In order to capture the contextual information, we add an extra layer of Bi-LSTM on top Huang et al. (2015). Given a candidate news article, the embedding of each word is represented as $(w_{1},w_{2},...,w_{N})$.

To integrate events knowledge into the basic language model, we construct a two-layer neural network on top of the word embeddings to learn the probability of each word triggering an event:

$$\begin{split}Q_{i}^{event}&=(q_{i}^{non-event},q_{i}^{event})\\ &=softmax(W_{2}(W_{1}w_{i}+b_{1})+b_{2})\end{split}$$

(6)

where $i=1,...,N$, and $Q_{i}^{event}$ is the learned probability of the basic language model. The soft label $P_{i}^{event}$ generated by the event relation graph is referenced as the target learning materials. By minimizing the cross entropy loss between the learned probability $Q_{i}^{event}$ and the target probability $P_{i}^{event}$, the events knowledge from the event relation graph can be infused into the basic language model:

$$Loss_{event}=-\sum_{i=1}^{N}P_{i}^{event}\log(Q_{i}^{event})$$

(7)

To integrate the event relations knowledge into the basic language model, we build four different neural networks on top of the event pair embedding, to learn the predicted probability of the four event relations respectively:

$$\begin{split}Q_{i,j}^{corefer}&=(q_{i,j}^{non-corefer},q_{i,j}^{corefer})\\ &=softmax(W_{4}(W_{3}(e_{i}\oplus e_{j})+b_{3})+b_{4})\end{split}$$

(8)

$$\begin{split}Q_{i,j}^{temp}&=(q_{i,j}^{non-temp},q_{i,j}^{before},q_{i,j}^{after},q_{i,j}^{overlap})\\ &=softmax(W_{6}(W_{5}(e_{i}\oplus e_{j})+b_{5})+b_{6})\end{split}$$

(9)

$$\begin{split}Q_{i,j}^{causal}&=(q_{i,j}^{non-causal},q_{i,j}^{cause},q_{i,j}^{caused-by})\\ &=softmax(W_{8}(W_{7}(e_{i}\oplus e_{j})+b_{7})+b_{8})\end{split}$$

(10)

$$\begin{split}Q_{i,j}^{subevent}&=(q_{i,j}^{non-subevent},q_{i,j}^{contain},q_{i,j}^{contained-by})\\ &=softmax(W_{10}(W_{9}(e_{i}\oplus e_{j})+b_{9})+b_{10})\end{split}$$

(11)

where $(e_{i}\oplus e_{j})$ is the embedding for event pair $(event_{i},event_{j})$ by concatenating the two events embeddings together. $Q_{i,j}^{corefer}$, $Q_{i,j}^{temp}$, $Q_{i,j}^{causal}$, and $Q_{i,j}^{subevent}$ are the learned probability of the basic language model for the four event relations. The soft labels generated by the event relation graph $P_{i,j}^{corefer}$, $P_{i,j}^{temp}$, $P_{i,j}^{causal}$, $P_{i,j}^{subevent}$ contain rich event relations information, and thus are referenced as the target learning materials. By minimizing the cross entropy loss between the learned probability and the target probability, the event relations knowledge within the event relation graph can be augmented into the basic language model:

$$Loss_{r}=-\sum_{i,j}P_{i,j}^{r}\log(Q_{i,j}^{r})$$

(12)

where $r\in\{corefer,temp,causal,subevent\}$ represents the four event relations.

The overall loss for training the event-aware language model based on soft labels is computed as the sum of the losses for learning each component:

$$\begin{split}Loss_{soft}=&Loss_{event}+Loss_{corefer}+Loss_{temp}\\ &+Loss_{causal}+Loss_{subevent}\end{split}$$

(13)

We further design a heterogeneous graph neural network to encode the event relation graph using hard labels. The encoder updates events embeddings with their neighbor events embeddings through interconnected relations, and produces a final graph embedding to represent each news article. This final article embedding is encoded with both events and event relations features, and will be later utilized for conspiracy theories identification.

To capture the global information of each document and explicitly indicate the connection between each document and its reported events, we introduce an extra document node and connect it to the associated events, as shown in Figure 2. The document node is initialized as the embedding of the article start token <s>, and the event node is initialized as the event word embedding.

The resulting event relation graph comprises nine fine-grained heterogeneous relations: coreference, before, after, overlap, causes, caused by, contains, contained by relations constructed from hard labels, as well as the event-doc relation. The eight event-event relations inherently carry semantic meaning, whereas the event-doc relation is a standard link without semantics. In order to incorporate the semantic meaning of event relations into graph propagation, we introduce the relation-aware graph attention network. In terms of the event-doc relation, we utilize the standard graph attention network Veličković et al. (2018).

The relation-aware graph attention network is designed to handle event-event relations, with the relations semantics integrated into event nodes embeddings. Given a pair of events $(event_{i},event_{j})$, their relation $r_{ij}$ is initialized as the embedding of the corresponding relation word. At the $l$-th layer, the input for $i$-th event node are output features produced by the previous layer denoted as $h_{i}^{(l-1)}$. During the propagation process, the relation embedding between $i$-th and $j$-th event is updated as:

$$r_{ij}=W^{r}[h_{i}^{(l-1)}\oplus r_{ij}\oplus h_{j}^{(l-1)}]$$

(14)

where $\oplus$ represents feature concatenation and $W^{r}$ is a trainable matrix. Then the attention weights $\alpha_{ij}$ across neighbor event nodes are computed as:

$$\alpha_{ij}=softmax_{j}\big{(}(W^{Q}h_{i}^{(l-1)})(W^{K}r_{ij})^{T}\big{)}$$

(15)

The output features $h_{i,r}^{(l)}$ are formulated as:

$$h_{i,r}^{(l)}=\sum_{j\in\mathcal{N}_{i,r}}\alpha_{ij}W^{V}r_{ij}$$

(16)

where $W^{Q}$, $W^{K}$, $W^{V}$ are trainable matrices, and $\mathcal{N}_{i,r}$ denotes the neighbor event nodes connecting with $i$-th event via the relation type $r$. After collecting $h_{i,r}^{(l)}$ for all relation types $r\in R=$ {coreference, before, after, overlap, causes, caused by, contains, contained by}, we aggregate the final output feature for $i$-th event at $l$-th layer as:

$$h_{i}^{(l)}=\sum_{r\in R}h_{i,r}^{(l)}/|R|$$

(17)

The standard graph attention network is employed to process event-doc relation, and update the document node embedding from connected events using the standard attention mechanism. The document node embedding at $l$-th layer is denoted as $d^{(l)}$. During propagation, the attention weights $\alpha_{i}$ across the connected events are computed as:

$$e_{i}=LeakyRelU\big{(}a^{T}\big{[}Wd^{(l-1)}\oplus Wh_{i}^{(l-1)}\big{]}\big{)}$$

(18)

$$\alpha_{i}=softmax_{i}(e_{i})=\frac{\exp(e_{i})}{\sum_{i}\exp(e_{i})}$$

(19)

where $i\in$ the events set, $a$ and $W$ are trainable parameters. The final output feature for the document node at $l$-th layer is calculated as:

$$d^{(l)}=\sum_{i}\alpha_{i}Wh_{i}^{(l-1)}$$

(20)

This document node embedding is the final embedding to represent the whole article, which contains both the information of graph structure and article context. We further build a two-layer classification head on top to predict conspiracy theories, and use cross-entropy loss for training.

Most prior work studied conspiracy theories within social media short text. LOCO Miani et al. (2021) is the only publicly available dataset that provides conspiracy theories labels for long news documents. LOCO consists of a large number of documents collected from 58 conspiracy theories media sources and 92 mainstream media sources. LOCO labels articles from conspiracy theories websites with the conspiracy class, and articles from mainstream media sources with the benign class. To prevent the conspiracy theory classifier from relying on stylistic features specific to a media source for identifying conspiracy stories, we will create media source aware train / dev / test data splits for our experiments as well, in addition to standard random data splitting that disregards data sources.

We design two different settings: 1) identifying conspiracy theories from unseen new media sources, or 2) from random media sources:

• •

  Media Source Splitting: We split the dataset based on media sources, and articles from the same media source will not appear in the same set: training, development, or testing. This ensures that the model is evaluated on articles whose sources were not seen during training. This setting will remove the inflation in performance that is due to the conspiracy theory classifier relying on the shortcuts of media sources features to make prediction. Table 2 and 3 presents the statistics of media sources and articles within the train, dev, and test sets.

• •

  Random Splitting: We adopt the standard method to split train / dev / test set randomly. For fair comparison, we ensure the equal size of training and development set, and also evaluate on the same testing set used in the media source splitting. In this setting, articles from the same media source can appear in both training and evaluation sets.

The event relation graph is trained on MAVEN-ERE dataset Wang et al. (2022) which contains massive general-domain news articles. The current state-of-art model framework Wang et al. (2022) is adopted to learn different components collaboratively. The performance of event identification is presented in Table 5. Table 4 shows the results of event coreference relation identification. Following the previous work Cai and Strube (2010), MUC Vilain et al. (1995), $B^{3}$ Bagga and Baldwin (1998), $CEAF_{e}$ Luo (2005), and BLANC Recasens and Hovy (2011) are used as evaluation metrics. Table 6 shows the performance of other components in event relation graph, including temporal, causal, and subevent relation tasks. The standard macro-average precision, recall, and F1 score are used for evaluation.

Our paper presents the first attempt to identify conspiracy theories in news articles. There are few established methods available for comparison. We experimented the following systems as baselines:

• •

  all-conspiracy: a naive baseline that categorizes all the documents into conspiracy class

• •

  chatgpt: where we designed an instruction prompt (A.1) that allows the large language model ChatGPT to automatically generate predicted labels for each news article within the same test set.

• •

  longformer: where we use the same language model Longformer Beltagy et al. (2020) and add an extra layer of Bi-LSTM on top. The embedding of the article start token is used as article embedding. The same two-layer classification head is built on top of the article embedding to predict conspiracy theories. This baseline model is equivalent to our developed model without event relation graph.

• •

  longformer + additional features: where we concatenate the soft labels eq (1)-(5) as additional features into the article embedding.

Table 7 reports the experimental results of conspiracy theories news identification under two different splitting settings. Precision, recall, and F1 score of the conspiracy class is shown.

In the media source splitting setting, we can see that incorporating the event relation graph substantially boosts precision by 3.59% and recall by 5.13%, compared to the longformer baseline. This suggests that the event relation graph encapsulates events and their logical interrelations, thereby has the ability to understand complex narratives of conspiracy theories. These improvements also show that our proposed method can generalize well for unseen new media sources.

In the random splitting setting, incorporating the event relation graph can also bring significant improvement to both precision and recall. The results indicate the effectiveness of the event relation graph method under either the hard and easy setting, with the F1 score increased by 4.22% to 4.47%. Unsurprising, the system performance in this setting is overall higher than in the media source splitting, probably due to the inflation caused by the classifier taking the shortcut of using media source recognition for conspiracy theory detection.

The event relation graph method outperforms the simple feature concatenation baseline. We observe that compared with the longformer baseline, incorporating probabilistic vectors as additional features (longformer + additional features) only slightly improve the performance. This demonstrate that developing more sophisticated methods to fully leverage the information embedded within the event relation graph is necessary.

The method based on event relation graph performs significantly better than the chatgpt. Comparing our full model to the chatgpt baseline, there is still a large gap in performance, especially the recall. We observe that chatgpt chooses conspiracy label mostly when false narratives are explicit. On the contrary, the event relation graph concentrates on logical reasoning, thus can better deal with implicit cases and yields higher recall.

The ablation study is also shown in Table 7. The event relation graph encoder based on hard labels contributes to enhancing recall, while the event-aware language model leveraging soft labels contributes to improving precision. It is probably because that the hard labels employ the four event relations in a relatively aggressive manner, by constructing the corresponding relation regardless of the confidence level. This enables the encoder to fully exploit the intrinsic information embedded in the event relation graph, thereby enhancing recall. On the contrary, the soft labels incorporate predicted probabilities of the event relation graph, allowing the model to learn from confidence levels and ultimately improving precision. The utilization of soft labels and hard labels are complementary with each other. Incorporating both soft labels and hard labels (the full model) exhibit the best performance.

We further study the effect of the different event relations in the event relation graph. Table 8 shows the experimental results of removing each type of event relations from the full model. The results show that removing any type of event relations leads to a performance drop, reducing both precision and recall. Therefore, each type of event relations plays an essential role in constructing a comprehensive content structure, and is crucial for identifying conspiracy theories news. Further, the experimental results demonstrate the importance of different event relations: removing temporal, causal, or subevent relations leads to larger performance decrease.

Figure 3 shows an example where our method succeeds in solving false negative. By identifying the events and extracting the relations between events, the model is aware of the author’s intention to attribute a causal relation between taking vaccination and autism. At the same time, the model is encoded with the information that the reaction event consisting of disease, pain, and inability happened after taking vaccination, which may not be adequate to reach a causal conclusion. The model not only learns the overall distribution of the four event relations, but is also aware of the specific relations described within a story, thereby can better comprehend complex narratives.

An example where our method fails to identify conspiracy theories is also shown in Figure 3. The author intends to imply a causal relation between COVID-19 pandemic and emergent 5G technology. However, the text expresses this relation in a very implicit way by using the phrase has something to do with. The event relation graph algorithm fails to recognize this implicitly stated causal relation and leads to a false negative error. In order to further improve the performance of conspiracy theory identification, it is necessary to improve event relation graph construction and better extract implicit event relations.