Preserving Pre-trained Features Helps Calibrate Fine-tuned Language Models

TL;DR

This paper investigates how preserving pre-trained features during fine-tuning improves the calibration of language models, especially under domain shifts, by reducing overconfidence and maintaining robust predictive confidence.

Contribution

The study shows that preserving pre-trained features enhances calibration of fine-tuned models and introduces a new method that encourages generative representations for better calibration.

Findings

Preserving pre-trained features improves calibration accuracy.

The proposed method achieves lower expected calibration error.

Method performs well both in-domain and out-of-domain.

Abstract

Large pre-trained language models (PLMs) have demonstrated strong performance on natural language understanding (NLU) tasks through fine-tuning. However, fine-tuned models still suffer from overconfident predictions, especially in out-of-domain settings. In this paper, we tackle the problem of calibrating fine-tuned language models. We demonstrate that the PLMs are well-calibrated on the masked language modeling task with robust predictive confidence under domain shift, yet the fine-tuned models fail to retain such property due to catastrophic forgetting, which impacts the calibration on the downstream classification task. In light of these observations, we evaluate the calibration of several methods that preserve pre-trained features and show that preserving pre-trained features can improve the calibration of fine-tuned language models. Among these methods, our proposed method that encourages the fine-tuned model to learn generative representations with auxiliary language modeling objective achieves competitive accuracy and the lowest expected calibration error compared to several strong baselines under both in-domain and out-of-domain settings on three downstream NLU tasks.