View Distillation with Unlabeled Data for Extracting Adverse Drug Effects from User-Generated Data

This paper presents an algorithm based on multi-layer transformers for identifying Adverse Drug Reactions (ADR) in social media data. The approach extracts two views from documents and uses classifiers trained on each view to label unlabeled documents, which then serve as initializers for classifiers in the other view. The model significantly outperforms transformer-based models pretrained on domain-specific data when evaluated on the largest publicly available ADR dataset.