Transformers have achieved a state of the art performance across most natural language processing tasks. However the performance of these models degrade when being trained on skewed class distributions (class imbalance) because training tends to be biased towards head classes with most of the data points . Classical methods that have been proposed to handle this problem (re-sampling and re-weighting) often suffer from unstable performance, poor applicability and poor calibration. In this paper, we propose to use Bayesian methods and Venn-Abers predictors for well calibrated and robust training against class imbalance. Our proposed approach improves f1-score of the baseline RoBERTa (A Robustly Optimized Bidirectional Embedding from Transformers Pretraining Approach) model by about 6 points (79.0% against 72.6%) when training with class imbalanced data.

The paper describes our submissions for the Social Media Mining for Health (SMM4H) workshop 2022 shared tasks. We participated in 2 tasks: (1) classification of adverse drug events (ADE) mentions in english tweets (Task-1a) and (2) classification of self-reported intimate partner violence (IPV) on twitter (Task 7). We proposed an approach that uses RoBERTa (A Robustly Optimized BERT Pretraining Approach) fine-tuned with a label distribution-aware margin loss function and post-hoc posterior calibration for robust inference against class imbalance. We achieved a 4% and 1 % increase in performance on IPV and ADE respectively when compared with the traditional fine-tuning strategy with unweighted cross-entropy loss.