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

Chaojun Wang, Rico Sennrich

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

The standard training algorithm in neural machine translation (NMT) suffers from exposure bias, and alternative algorithms have been proposed 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 tendency 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 Training, 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 exposure bias: even if they do not increase performance on in-domain test sets, they can increase model robustness to domain shift.
Original languageEnglish
Title of host publicationProceedings of the 58th Annual Meeting of the Association for Computational Linguistics
PublisherAssociation for Computational Linguistics (ACL)
Pages3544–3552
Number of pages9
ISBN (Electronic)978-1-952148-25-5
DOIs
Publication statusPublished - 10 Jul 2020
Event2020 Annual Conference of the Association for Computational Linguistics - Hyatt Regency Seattle, Virtual conference, United States
Duration: 5 Jul 202010 Jul 2020
Conference number: 58
https://acl2020.org/

Conference

Conference2020 Annual Conference of the Association for Computational Linguistics
Abbreviated titleACL 2020
CountryUnited States
CityVirtual conference
Period5/07/2010/07/20
Internet address

Fingerprint Dive into the research topics of 'On Exposure Bias, Hallucination and Domain Shift in Neural Machine Translation'. Together they form a unique fingerprint.

Cite this