Fake news

Data will be collected from social media posts containing multimodal information such as text, image, and Comment relations, etc. Two fake news datasets in English and Chinese were sourced to train and test our proposed model. They are:

- \textbf{Fakeddit}\footnote{https://github.com/entitize/Fakeddit}: an English multimodal dataset containing text, comments, and images from Reddit (n = 3,127). The samples within this dataset are labeled using a 2-way (real, fake) and a 3-way classification (real, fake with true text, fake with false text) scheme. Modelling was performed using both these classifications.

- \textbf{Multimodal Fake News Detection dataset}\footnote{https://data.beijing.gov.cn/kjzy2020/index.html}(MFND): a multimodal Chinese dataset containing text, comments, and images from Weibo (n = 2,953). The dataset is categorized into three distinct classes: uncertain, fake news, and real news.

The number of instances for each classified labels in the datasets are given in Table \ref{tab:dataset_overview}, indicating the datasets to be balanced.
\begin{table}[ht]
\centering
\caption{Fake news datasets overview}
\begin{tabular}{c|c|c|c|c}
\toprule
Dataset & 2-way & 3-way & n & Total \\
\midrule
& - & real & 1,000 & \\
MFND& - & fake & 953 & 2,953\\
& & uncertain & 1,000 & \\
\midrule
& real & real & 1,048 & \\
Fakeddit& fake & fake with true text & 1,060 & 3,127\\
& & fake with false text & 1,019 & \\
\bottomrule
\end{tabular}
\label{tab:dataset_overview}
\end{table}

Observing the two datasets, it becomes evident that the distribution of categories is relatively balanced, with no significant gaps between individual labels. This balanced distribution provides us with a more reliable and comprehensive foundation for accurate analysis and modeling. In the fake news dataset, each piece of news is composed of different modalities, namely posts, images, and comments. The following Figure \ref{fig:2} illustrates the overall distribution of these modalities.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{fig/fig2.png}
\caption{Distribution of modalities}
\label{fig:2}
\end{figure}

Taking a sample from the dataset, as illustrated in Figure \ref{fig:3}, we observe a user posting
a narrative along with an image, while numerous other users engage in comments. The
semantic gap between these comments and the post is substantial, mostly comprising
emotional opinions. Some users even verify the authenticity of the information, as evident
in this example where a commenter claims to have seen related information on YouTube.
Thus, we affirm the dataset's multimodal and semantic diversity.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{fig/4.png}
\caption{Example of dataset}
\label{fig:3}
\end{figure}