Distributionally Robust Classifiers in Sentiment Analysis

TL;DR

This paper introduces a distributionally robust sentiment classifier based on BERT with additional layers, designed to maintain performance under dataset shifts, and demonstrates its effectiveness on IMDb and Rotten Tomatoes datasets.

Contribution

The paper presents a novel BERT-based sentiment classifier with integrated DRO and extra layers to enhance robustness against distributional shifts.

Findings

DRO model improves performance under dataset shift

Additional layers contribute to robustness

Effective on IMDb to Rotten Tomatoes shift

Abstract

In this paper, we propose sentiment classification models based on BERT integrated with DRO (Distributionally Robust Classifiers) to improve model performance on datasets with distributional shifts. We added 2-Layer Bi-LSTM, projection layer (onto simplex or Lp ball), and linear layer on top of BERT to achieve distributionally robustness. We considered one form of distributional shift (from IMDb dataset to Rotten Tomatoes dataset). We have confirmed through experiments that our DRO model does improve performance on our test set with distributional shift from the training set.