Why Exposure Bias Matters: An Imitation Learning Perspective of Error Accumulation in Language Generation
This addresses a key issue for natural language processing researchers and practitioners by verifying a long-standing hypothesis about model brittleness, though it is incremental as it builds on existing imitation learning perspectives.
The paper tackles the problem of exposure bias in language generation models, showing that it causes error accumulation leading to poor generation quality, with empirical evidence demonstrating its negative impact.
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