Minimum Edit Distance Training for Conditional Language Generation Models

TL;DR — We propose a loss function between a pair of sequences of different lengths using edit distance. Through this, exposure bias and calibration error of the conditional language model are alleviated and generalization performance is improved.

Abstract

The utilization of attention-based encoder-decoder (AED) structures, including transformers, has further advanced the capabilities of conditional language generation (CLG) models. However, the conventional AED model training approach which aims to maximize the likelihood conditioned on the prefix of reference label sequence, introduces exposure bias and possesses limitations in that it uses different evaluation metrics in the training and inference stages. In this study, we introduce a novel AED model training technique focused on minimizing the Levenshtein distance between the reference and inferred label sequences. The proposed method effectively mitigates exposure bias and improves the generalization performance of neural machine translation and automatic speech recognition models. Furthermore, we demonstrate that a post-hoc calibration function trained with the proposed objective function significantly reduces the calibration error of the ASR model, resulting in notable performance improvements.