Pathologies of Pre-trained Language Models in Few-shot Fine-tuning

TL;DR

This paper investigates the underlying reasons for performance improvements in few-shot fine-tuning of pre-trained language models, revealing that models often rely on non-task-related features, leading to potential pathologies.

Contribution

It introduces a post-hoc explanation method to analyze adaptation behavior, uncovering biases and shallow pattern reliance in few-shot fine-tuning of models like BERT and RoBERTa.

Findings

Pre-trained models exhibit strong prediction bias without fine-tuning.

Few-shot fine-tuning reduces bias but models rely on non-task features.

Models may exploit shallow data patterns, risking pathological behavior.

Abstract

Although adapting pre-trained language models with few examples has shown promising performance on text classification, there is a lack of understanding of where the performance gain comes from. In this work, we propose to answer this question by interpreting the adaptation behavior using post-hoc explanations from model predictions. By modeling feature statistics of explanations, we discover that (1) without fine-tuning, pre-trained models (e.g. BERT and RoBERTa) show strong prediction bias across labels; (2) although few-shot fine-tuning can mitigate the prediction bias and demonstrate promising prediction performance, our analysis shows models gain performance improvement by capturing non-task-related features (e.g. stop words) or shallow data patterns (e.g. lexical overlaps). These observations alert that pursuing model performance with fewer examples may incur pathological prediction behavior, which requires further sanity check on model predictions and careful design in model evaluations in few-shot fine-tuning.