Multimodal Pipeline for Collection of Misinformation Data from Telegram

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 1480–1489 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 1480 Multimodal Pipeline for Collection of Misinformation Data from Telegram Jose Sosa, Serge Sharoff University of Leeds Woodhouse Lane, LS2 9JT, Leeds, UK {scjasm, s.sharoff}@leeds.ac.uk Abstract The paper presents the outcomes of AI-COVID19, our project aimed at better understanding of misinformation flow about COVID-19 across social media platforms. The specific focus of the study reported in this paper is on collecting data from Telegram groups which are active in promotion of COVID-related misinformation. Our corpus collected so far contains around 28 million words, from almost one million messages. Given that a substantial portion of misinformation flow in social media is spread via multimodal means, such as images and video, we have also developed a mechanism for utilising such channels via producing automatic transcripts for videos and automatic classification for images into such categories as memes, screenshots of posts and other kinds of images. The accuracy of the image classification pipeline is around 87%. Keywords: Telegram data collection, COVID-19 misinformation, multimodal classification 1. Introduction The widespread distribution of COVID-19 misinfor- mation leads to confusion, anti-health policy senti- ment, and risk-tolerant behaviour (Chou et al., 2021). AI-COVID19 is our project which aims to use AI tools to understand the dynamics of misinformation spread across several social media platforms, includ- ing Twitter, Facebook and Telegram. This paper con- centrates on the methods and outcomes of data collec- tion from Telegram, a platform that is currently less ac- tive in policing misinformation, so that we can collect a greater variety of COVID-19 misinformation exam- ples. We designed a pipeline to collect COVID-19 misin- formation from Telegram public channels. Then, we used this pipeline to build one of the first multimodal datasets of COVID-19 misinformation; i.e. in addi- tion to the prevailing text data, our dataset includes im- ages, videos, and documents. Overall the dataset com- prises almost one million messages from 2k different public channels related to spreading COVID-19 mis- leading information. In addition, it includes 38k im- ages, 15k videos, and 522 documents (mostly in the PDF and DOCX formats) from those channels. Fur- thermore, it organises the collected images into three categories: memes, posts and others by means of auto- matic image classification. Finally, it incorporates a set of transcripts for the collected videos. We summarised our contributions as follows: • Automatic pipeline for collecting misinformation from Telegram. • Joint collection of text and multimedia data. Our pipeline for collecting data allows us to get text and also any multimedia data that can be part of the messages, e.g. images, videos, documents, au- dios, and stickers. • A classifier for the collected images. We train a CNN classifier based on AlexNet (Krizhevsky et al., 2012) to identify memes and images from text- based posts. • A new telegram multimodal dataset on COVID-19 related misinformation. The complete code for collecting data, the classifier, and the first …

Content truncated to 500 words