Adverse Drug Reaction Detection in Twitter Using RoBERTa and Rules


Sedigh Khademi, Pari Delirhaghighi, Frada Burstein


This paper describes the method we developed for the Task 2 English variation of the Social Media Mining for Health Applications (SMM4H) 2020 shared task. The task was to classify tweets containing adverse effects (AE) after medication intake. Our approach combined transfer learning using a RoBERTa Large Transformer model with a rule-based post-prediction correction to improve model precision. The model’s F1-Score of 0.56 on the test dataset was 10% better than the mean of the F1-Score of the best submissions in the task.

Table of Contents


  1. Introduction
  2. Data description and preparation
  3. Model
  4. Method
    • Best model vs an ensemble
    • Rule-based correction
  5. Generated texts
  6. Results
  7. Analysis
  8. Future work


Link to full publication