Dartmouth CS at WNUT-2020 Task 2: Informative COVID-19 Tweet Classification Using BERT

TL;DR

This paper presents a system that enhances COVID-19 tweet classification by fine-tuning BERT with tweet-specific features and training an SVM, achieving top performance in the WNUT-2020 shared task.

Contribution

The study introduces a novel BERT-based approach combined with tweet features and SVM for improved COVID-19 tweet informativeness classification.

Findings

BERT+ achieved an F1-score of 0.8713.

The combined BERT and SVM approach outperformed other models.

Tweet-specific features improved classification accuracy.

Abstract

We describe the systems developed for the WNUT-2020 shared task 2, identification of informative COVID-19 English Tweets. BERT is a highly performant model for Natural Language Processing tasks. We increased BERT's performance in this classification task by fine-tuning BERT and concatenating its embeddings with Tweet-specific features and training a Support Vector Machine (SVM) for classification (henceforth called BERT+). We compared its performance to a suite of machine learning models. We used a Twitter specific data cleaning pipeline and word-level TF-IDF to extract features for the non-BERT models. BERT+ was the top performing model with an F1-score of 0.8713.