Using Context in Neural Machine Translation Training Objectives

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7764–7770 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 7764 Using Context in Neural Machine Translation Training Objectives Danielle Saunders and Felix Stahlberg∗and Bill Byrne Department of Engineering, University of Cambridge, UK ds636@cam.ac.uk fstahlberg@google.com wjb31@cam.ac.uk Abstract We present Neural Machine Translation (NMT) training using document-level met- rics with batch-level documents. Previous sequence-objective approaches to NMT train- ing focus exclusively on sentence-level met- rics like sentence BLEU which do not corre- spond to the desired evaluation metric, typ- ically document BLEU. Meanwhile research into document-level NMT training focuses on data or model architecture rather than training procedure. We find that each of these lines of research has a clear space in it for the other, and propose merging them with a scheme that allows a document-level evaluation metric to be used in the NMT training objective. We first sample pseudo-documents from sen- tence samples. We then approximate the ex- pected document BLEU gradient with Monte Carlo sampling for use as a cost function in Minimum Risk Training (MRT). This two- level sampling procedure gives NMT per- formance gains over sequence MRT and maximum-likelihood training. We demon- strate that training is more robust for document-level metrics than with sequence metrics. We further demonstrate improve- ments on NMT with TER and Grammatical Er- ror Correction (GEC) using GLEU, both met- rics used at the document level for evaluations. 1 Introduction Neural Machine Translation (NMT) research has explored token-level likelihood functions (Sutskever et al., 2014; Bahdanau et al., 2015) and sequence-level objectives inspired by reinforce- ment learning (Ranzato et al., 2016; Bahdanau et al., 2016) or expected Minimum Risk Training (MRT) (Shen et al., 2016). A typical sequence ob- jective in these cases is based on sentence-level BLEU (sBLEU) (Edunov et al., 2018). However ∗Now at Google sBLEU, even if aggregated over sentences, is only an approximation of the desired metric, document- level BLEU. Beyond translation, many metrics for natural language tasks do not have robust sentence- level approximations. A logical progression is the extension of sequence-level NMT training objec- tives to include context from outside the sentence. Document-based NMT, by contrast, aims to use out-of-sentence context to improve translation. Re- cent research explores lexical consistency by pro- viding additional sentences during training (Maruf et al., 2019; Voita et al., 2018, 2019) or inference (Voita et al., 2019; Stahlberg et al., 2019), poten- tially with adjustments to model architecture. How- ever, to the best of our knowledge, no attempt has been made to extend sequence-level neural training objectives to include document-level reward func- tions. This is despite document-level BLEU being arguably the most common NMT metric, and being the function originally optimised by Minimum Er- ror Rate Training (MERT) for Statistical Machine Translation (SMT) (Och, 2003). We propose merging lines of research on train- ing objectives and document-level translation. We achieve this by presenting a document-level ap- proach to sequence-level objectives which brings the training objective closer to the …

Content truncated to 500 words