Why Exposure Bias Matters: An Imitation Learning Perspective of Error Accumulation in Language Generation

TL;DR

This paper investigates how exposure bias in language models causes error accumulation, leading to issues like repetition and hallucinations, by analyzing it through an imitation learning lens and providing empirical evidence.

Contribution

It offers a novel analysis of exposure bias from an imitation learning perspective and demonstrates its impact on error accumulation in language generation.

Findings

Exposure bias causes error accumulation in language models.

Perplexity does not effectively measure error accumulation.

Error accumulation leads to poor generation quality.

Abstract

Current language generation models suffer from issues such as repetition, incoherence, and hallucinations. An often-repeated hypothesis is that this brittleness of generation models is caused by the training and the generation procedure mismatch, also referred to as exposure bias. In this paper, we verify this hypothesis by analyzing exposure bias from an imitation learning perspective. We show that exposure bias leads to an accumulation of errors, analyze why perplexity fails to capture this accumulation, and empirically show that this accumulation results in poor generation quality. Source code to reproduce these experiments is available at https://github.com/kushalarora/quantifying_exposure_bias