MultimodalTweetAnalysis

This notebook represents our approach for Task 1B of the CheckThat! Lab at CLEF 2022. We ranked #1 on the task leaderboard on CodaLab.

Task 1B

Verifiable factual claims detection: Given a tweet, predict whether it contains a verifiable factual claim. This is a binary task with two labels: Yes and No. This is a classification task.

Methodology

1. Data Augmentation

We translated the Dutch and Bulgarian datasets for the task into English to increase the amount of training data we have.

2. Feature Extraction through Twitter API

We used the twitter API to extract the following data about a tweet:

1. Numerical Features:

• number of 'followers', 'following', 'posts' of the author of the tweet
• number of 'likes', 'retweets' of the tweet itself

3. Categorical Features: -'verified' as an attribute of the author of the tweet -'url' indicating presence of a URL in the tweet

3. Multimodal Model

We used the Multimodal Toolkit for text and tabular data with HuggingFace transformers as building block for text data.