Supervised Contrastive Learning for Pre-trained Language Model Fine-tuning

TL;DR

This paper introduces a supervised contrastive learning objective for fine-tuning pre-trained language models, improving generalization, robustness, and performance in few-shot settings without extra data or architecture changes.

Contribution

It proposes a novel supervised contrastive loss for fine-tuning, enhancing model performance and robustness over traditional cross-entropy loss in NLP tasks.

Findings

Significant improvements on GLUE benchmark in few-shot learning.

Enhanced robustness to noise in training data.

Better generalization to related tasks with limited labels.

Abstract

State-of-the-art natural language understanding classification models follow two-stages: pre-training a large language model on an auxiliary task, and then fine-tuning the model on a task-specific labeled dataset using cross-entropy loss. However, the cross-entropy loss has several shortcomings that can lead to sub-optimal generalization and instability. Driven by the intuition that good generalization requires capturing the similarity between examples in one class and contrasting them with examples in other classes, we propose a supervised contrastive learning (SCL) objective for the fine-tuning stage. Combined with cross-entropy, our proposed SCL loss obtains significant improvements over a strong RoBERTa-Large baseline on multiple datasets of the GLUE benchmark in few-shot learning settings, without requiring specialized architecture, data augmentations, memory banks, or additional unsupervised data. Our proposed fine-tuning objective leads to models that are more robust to different levels of noise in the fine-tuning training data, and can generalize better to related tasks with limited labeled data.