On Exposure Bias, Hallucination and Domain Shift in Neural Machine Translation

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3544–3552 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 3544 On Exposure Bias, Hallucination and Domain Shift in Neural Machine Translation Chaojun Wang1 Rico Sennrich2,1 1School of Informatics, University of Edinburgh 2Department of Computational Linguistics, University of Zurich zippo wang@foxmail.com, sennrich@cl.uzh.ch Abstract The standard training algorithm in neural ma- chine translation (NMT) suffers from exposure bias, and alternative algorithms have been pro- posed to mitigate this. However, the practical impact of exposure bias is under debate. In this paper, we link exposure bias to another well-known problem in NMT, namely the ten- dency to generate hallucinations under domain shift. In experiments on three datasets with multiple test domains, we show that exposure bias is partially to blame for hallucinations, and that training with Minimum Risk Train- ing, which avoids exposure bias, can mitigate this. Our analysis explains why exposure bias is more problematic under domain shift, and also links exposure bias to the beam search problem, i.e. performance deterioration with increasing beam size. Our results provide a new justification for methods that reduce ex- posure bias: even if they do not increase per- formance on in-domain test sets, they can in- crease model robustness to domain shift. 1 Introduction Neural Machine Translation (NMT) has advanced the state of the art in MT (Sutskever et al., 2014; Bahdanau et al., 2015; Vaswani et al., 2017), but is susceptible to domain shift. Koehn and Knowles (2017) consider out-of-domain translation one of the key challenges in NMT. Such translations may be fluent, but completely unrelated to the input (hallucinations), and their misleading nature makes them particularly problematic. We hypothesise that exposure bias (Ranzato et al., 2016), a discrepancy between training and inference, makes this problem worse. Specifi- cally, training with teacher forcing only exposes the model to gold history, while previous predic- tions during inference may be erroneous. Thus, the model trained with teacher forcing may over-rely on previously predicted words, which would exac- erbate error propagation. Previous work has sought to reduce exposure bias in training (Bengio et al., 2015; Ranzato et al., 2016; Shen et al., 2016; Wise- man and Rush, 2016; Zhang et al., 2019). However, the relevance of error propagation is under debate: Wu et al. (2018) argue that its role is overstated in literature, and that linguistic features explain some of the accuracy drop at higher time steps. Previous work has established a link between domain shift and hallucination in NMT (Koehn and Knowles, 2017; M¨uller et al., 2019). In this paper, we will aim to also establish an empirical link between hallucination and exposure bias. Such a link will deepen our understanding of the hallu- cination problem, but also has practical relevance, e.g. to help predicting in which settings the use of sequence-level objectives is likely to be helpful. We further empirically confirm the link between exposure bias and the ‘beam search problem’, i.e. the fact that translation quality does not increase …

Content truncated to 500 words