Elastic weight consolidation for better bias inoculation
This addresses bias mitigation in NLP models for tasks like fact verification and NLI, but it is incremental as it applies an existing method (EWC) to a known issue.
The paper tackled the problem of catastrophic forgetting when fine-tuning models to mitigate biases in sentence pair classification tasks, showing that elastic weight consolidation (EWC) reduces forgetting on the original biased dataset while maintaining accuracy gains on the unbiased fine-tuning dataset.
The biases present in training datasets have been shown to affect models for sentence pair classification tasks such as natural language inference (NLI) and fact verification. While fine-tuning models on additional data has been used to mitigate them, a common issue is that of catastrophic forgetting of the original training dataset. In this paper, we show that elastic weight consolidation (EWC) allows fine-tuning of models to mitigate biases while being less susceptible to catastrophic forgetting. In our evaluation on fact verification and NLI stress tests, we show that fine-tuning with EWC dominates standard fine-tuning, yielding models with lower levels of forgetting on the original (biased) dataset for equivalent gains in accuracy on the fine-tuning (unbiased) dataset.