Twitter Rumour Classification with BERTweet

1 minute read


In this blog post, I would like to present my submission to the COMP90042 Natural Language Processing Project at the University of Melbourne (Semester 1 2021).

Transformers Logo

The task of the project was to

  • Develop a system for rumour identification (task 1) and
  • Analyze the nature of rumours that are being propagated on Twitter (task 2).

The dataset for the project was published by the COMP90042 teaching team and consisted of a set of source tweets and their replies (incl. corresponding metadata) that had been extracted from the Twitter API. In total, the training data consisted of 4641 events that had been labeled as either RUMOUR or NON-RUMOUR (binary classification).

For this project, I have implemented three classification systems:

  1. A BERT-based implementation that uses the textual representation of tweets (called “PureBERT”)
  2. An extension of the PureBERT architecture that combines the textual features with tabular data (called “MultimodalBERT”)
  3. A language model that has been pre-trained on a large corpus (850 million) of English Tweets (called “BERTweet”).

Using the best-performing model BERTweet, I managed to achieve a F1 score of 86.17% (which put me on rank 12 out of 308 participants in the final CodaLab competition).

A detailed write-up of the implementation details (pre-processing routine etc.) for the models mentioned above is available here, and if you are interested in further details, please refer to the following repository: COMP90042-Rumour-Detection-on-Twitter

I have also used BERTweet to participate in the “Disaster Tweets” Kaggle challenge. The notebook is available here: Disaster Tweets - BERTweet