BOUN-TABI@SMM4H’22: Text-to-Text Adverse Drug Event Extraction with Data Balancing and Prompting

Abstract

This paper describes models developed for the Social Media Mining for Health 2022 Shared Task. We participated in two subtasks: classification of English tweets reporting adverse drug events (ADE) (Task 1a) and extraction of ADE spans in such tweets (Task 1b). We developed two separate systems based on the T5 model, viewing these tasks as sequence-to-sequence problems. To address the class imbalance, we made use of data balancing via over- and undersampling on both tasks. For the ADE extraction task, we explored prompting to further benefit from the T5 model and its formulation. Additionally, we built an ensemble model, utilizing both balanced and prompted models. The proposed models outperformed the current state-of-the-art, with an F1 score of 0.655 on ADE classification and a Partial F1 score of 0.527 on ADE extraction.

Tasks

Reproductions