Controlling Politeness in Neural Machine Translation via Side Constraints

Proceedings of NAACL-HLT 2016, pages 35–40, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Controlling Politeness in Neural Machine Translation via Side Constraints Rico Sennrich and Barry Haddow and Alexandra Birch School of Informatics, University of Edinburgh {rico.sennrich,a.birch}@ed.ac.uk, bhaddow@inf.ed.ac.uk Abstract Many languages use honorifics to express po- liteness, social distance, or the relative so- cial status between the speaker and their ad- dressee(s). In machine translation from a lan- guage without honorifics such as English, it is difficult to predict the appropriate honorific, but users may want to control the level of po- liteness in the output. In this paper, we per- form a pilot study to control honorifics in neu- ral machine translation (NMT) via side con- straints, focusing on English→German. We show that by marking up the (English) source side of the training data with a feature that en- codes the use of honorifics on the (German) target side, we can control the honorifics pro- duced at test time. Experiments show that the choice of honorifics has a big impact on translation quality as measured by BLEU, and oracle experiments show that substantial im- provements are possible by constraining the translation to the desired level of politeness. 1 Introduction Many languages use honorifics to express polite- ness, social distance, or the relative social status be- tween the speaker and their addressee(s). A wide- spread instance is the grammatical T-V distinction (Brown and Gilman, 1960), distinguishing between the familiar (Latin Tu) and the polite (Latin Vos) second person pronoun. In machine translation from a language without honorifics such as English, it is difficult to predict the appropriate honorific, but users may want to control the level of politeness in the output. The research presented in this publication was conducted in cooperation with Samsung Electronics Polska sp. z o.o. - Sam- sung R&D Institute Poland. We propose a simple and effective method for in- cluding target-side T-V annotation in the training of a neural machine translation (NMT) system, which allows us to control the level of politeness at test time through what we call side constraints. It can be applied for translation between languages where the T-V distinction is missing from the source, or where the distribution differs. For instance, both Swedish and French make the T-V distinction, but reciprocal use of T pronouns is more widespread in Swedish than in French (Schüpbach et al., 2006). Hence, the Swedish form is not a reliable signal for the appro- priate form in the French translation (or vice-versa). Our basic approach of using side constraints to control target-side features that may be missing from the source, or are unreliable because of a cate- gory mismatch, is not limited to the T-V distinc- tion, but could be applied to various linguistic fea- tures. This includes grammatical features such as tense and the number/gender of discourse partici- pants, and more generally, features such as dialect and register choice. This paper has the following contributions: • we describe rules to automatically annotate …

Content truncated to 500 words