Phrase-level Active Learning for Neural Machine Translation

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 1087–1099 November 10–11, 2021. ©2021 Association for Computational Linguistics 1087 Phrase-level Active Learning for Neural Machine Translation Junjie Hu∗ University of Wisconsin-Madison junjie.hu@wisc.edu Graham Neubig Carnegie Mellon University gneubig@cs.cmu.edu Abstract Neural machine translation (NMT) is sensi- tive to domain shift. In this paper, we ad- dress this problem in an active learning set- ting where we can spend a given budget on translating in-domain data, and gradually fine- tune a pre-trained out-of-domain NMT model on the newly translated data. Existing active learning methods for NMT usually select sen- tences based on uncertainty scores, but these methods require costly translation of full sen- tences even when only one or two key phrases within the sentence are informative. To ad- dress this limitation, we re-examine previous work from the phrase-based machine transla- tion (PBMT) era that selected not full sen- tences, but rather individual phrases. However, while incorporating these phrases into PBMT systems was relatively simple, it is less trivial for NMT systems, which need to be trained on full sequences to capture larger structural prop- erties of sentences unique to the new domain. To overcome these hurdles, we propose to se- lect both full sentences and individual phrases from unlabelled data in the new domain for routing to human translators. In a German- English translation task, our active learning ap- proach achieves consistent improvements over uncertainty-based sentence selection methods, improving up to 1.2 BLEU score over strong active learning baselines.1 1 Introduction Machine translation (MT) models are very sensi- tive to domain shift (Koehn and Knowles, 2017; Chu and Wang, 2018), and one typical way to ad- dress this problem is adding in-domain data to the MT training process (Luong and Manning, 2015; Chu et al., 2017). However, this data may not be available a priori, and hiring professional transla- tors with knowledge of specific domains (such as medicine or law) is usually costly. ∗Work done at Carnegie Mellon University 1Code/data is released at https://github.com/ JunjieHu/phrase-al-nmt. Out-of-Domain Parallel0Data In-Domain Source0Data 1.0Train0A Base0NMT Sentences Phrases 3.0Annotate 4.0Mixed Fine-tuning In-Domain Parallel0Data 2.0Hybrid0 Phrase/Sentence0 Selection Figure 1: Overview of the active learning process As a result, active learning approaches (Gangad- haraiah et al., 2009; Haffari et al., 2009; Blood- good and Callison-Burch, 2010) have been widely adopted to reduce the annotation cost by translat- ing a smaller representative subset of the in-domain data, with the hope that models trained on this trans- lated subset approximate those trained on a much larger labeled set. In general, active learning (AL) approaches iterate between two steps: data selec- tion/annotation, and model update. With regards to data selection for machine translation, most exist- ing works (Haffari et al., 2009; Peris and Casacu- berta, 2018; Zeng et al., 2019) focus on selecting sentences that are most useful for training either phrase-based machine translation (PBMT) or neu- ral machine translation (NMT) models. However, even the most informative sentences inevitably involve segments that the MT system can already translate well, and …

Content truncated to 500 words