[Paper Review] Quantifying Exposure Bias for Neural Language Generation

Tianxing He, Jingzhao Zhang+2 more|arXiv (Cornell University)|Sep 25, 2019

Topic Modeling49 references24 citations

TL;DR

This paper proposes EB-bleu and EB-C metrics to quantify exposure bias in neural language generation, evaluating its real impact on LSTM and transformer models across autoregressive LM and seq2seq translation tasks. Surprisingly, it finds that removing training-generation discrepancy via exposure bias mitigation yields only marginal performance gains, suggesting the mismatch between model and data distributions lies within a 'comfortable zone' where performance loss is minimal.

ABSTRACT

The exposure bias problem refers to the training-generation discrepancy, caused by teacher forcing, in maximum likelihood estimation (MLE) training for auto-regressive neural network language models (LM). It has been regarded as a central problem for neural language generation (NLG) model training. Although a lot of algorithms have been proposed to avoid teacher forcing and therefore` to alleviate exposure bias, there is little work showing how serious the exposure bias problem actually is. In this work, we first identify the self-recovery ability of MLE-trained LM, which casts doubt on the seriousness of exposure bias. We then propose sequence-level (EB-bleu) and word-level (EB-C) metrics as reasonable proxies to quantify the impact of exposure bias. We conduct experiments for the LSTM/transformer model, in both real and synthetic settings. In addition to the unconditional NLG task, we also include results for a seq2seq machine translation task. Surprisingly, all our measurements suggest that removing the training-generation discrepancy only brings very little performance gain. In our analysis, we hypothesise that although there exist a mismatch between the model distribution and the data distribution, the mismatch is still in the model's comfortable zone, and is not big enough to induce significant performance loss.

Motivation & Objective

• To investigate the actual severity of exposure bias in maximum likelihood-trained neural language models.
• To develop reliable, sequence-level and word-level metrics to quantify exposure bias impact.
• To evaluate whether removing training-generation discrepancy significantly improves model performance.
• To analyze whether the distributional mismatch caused by teacher forcing is large enough to cause substantial performance degradation.
• To assess the robustness of MLE-trained models to exposure bias across different architectures and tasks.

Proposed method

• Propose EB-bleu, a sequence-level metric based on BLEU scores computed under controlled exposure conditions to estimate exposure bias impact.
• Introduce EB-C, a word-level metric that measures the model's sensitivity to exposure bias by evaluating token-level prediction reliability.
• Train LSTM and transformer models using standard teacher-forcing MLE for both unconditional language modeling and seq2seq translation tasks.
• Conduct experiments in both real and synthetic data settings to isolate exposure bias effects from other training factors.
• Use self-recovery analysis to assess whether models can recover from exposure-induced errors during inference.
• Compare model performance under standard MLE training versus exposure-bias-removal techniques to measure marginal gains.

Experimental results

Research questions

• RQ1 How severe is exposure bias in practice for MLE-trained neural language models?
• RQ2 To what extent do exposure bias mitigation techniques improve model performance on language generation tasks?
• RQ3 Does the distributional mismatch caused by teacher forcing significantly degrade model generation quality?
• RQ4 Is the model's performance loss due to exposure bias large enough to justify complex mitigation strategies?
• RQ5 How do the proposed EB-bleu and EB-C metrics compare in capturing exposure bias effects across different model architectures?

Key findings

• The proposed EB-bleu and EB-C metrics effectively quantify exposure bias impact across both real and synthetic language generation tasks.
• Performance gains from removing training-generation discrepancy are minimal, suggesting exposure bias is less severe than commonly assumed.
• Models trained with MLE exhibit strong self-recovery ability, indicating the distributional mismatch lies within a 'comfortable zone' where performance degradation is limited.
• Even in synthetic settings designed to amplify exposure bias, the performance improvements from mitigation remain small.
• The results hold across both LSTM and transformer architectures, as well as for both unconditional generation and seq2seq translation tasks.
• The study challenges the prevailing assumption that exposure bias is a major performance bottleneck in neural language generation.