Domain adaptation for neural machine translation

The development of deep learning techniques has allowed Neural Machine Translation (NMT) models to become extremely powerful, given sufficient training data and training time. However, such translation models struggle when translating text of a specific domain. A domain may consist of text on a well-defined topic, or text of unknown provenance with an identifiable vocabulary distribution, or language with some other stylometric feature. While NMT models can achieve good translation performance on domain-specific data via simple tuning on a representative training corpus, such data-centric approaches have negative side-effects. These include over-fitting, brittleness, and `catastrophic forgetting' of previous training examples. In this thesis we instead explore more robust approaches to domain adaptation for NMT. We consider the case where a system is adapted to a specified domain of interest, but may also need to accommodate new language, or domain-mismatched sentences. We explore techniques relating to data selection and curriculum, model parameter adaptation procedure, and inference procedure. We show that iterative fine-tuning can achieve strong performance over multiple related domains, and that Elastic Weight Consolidation can be used to mitigate catastrophic forgetting in NMT domain adaptation across multiple sequential domains. We develop a robust variant of Minimum Risk Training which allows more beneficial use of small, highly domain-specific tuning sets than simple cross-entropy fine-tuning, and can mitigate exposure bias resulting from domain over-fitting. We extend Bayesian Interpolation inference schemes to Neural Machine Translation, allowing adaptive weighting of NMT ensembles to translate text from an unknown domain. Finally we demonstrate the benefit of multi-domain adaptation approaches for other lines of NMT research. We show that NMT systems using multiple forms of data representation can benefit from multi-domain inference approaches. We also demonstrate a series of domain adaptation approaches to mitigating the effects of gender bias in machine translation.

Name dc:type.qualificationname

Doctor of Philosophy (PhD)

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2020

Author dc:creator

• Saunders, Danielle

Advisor dc:contributor.advisor

• Byrne, William

• Neural machine translation
• Natural language processing
• Domain adaptation

Licence

• https://apollo8-f-pro.lib.cam.ac.uk/bitstreams/d83224bb-fb83-48c9-87ea-4152b36a85d3/download
• https://www.rioxx.net/licenses/all-rights-reserved/

Language dc:language

eng

Author Identifier

0000-0002-7943-3102

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/319335