Machine Translation, it’s a question of style, innit? The case of English tag questions

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2507–2512 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Machine Translation, it’s a question of style, innit? The case of English tag questions Rachel Bawden LIMSI, CNRS, Univ. Paris-Sud, Universit´e Paris-Saclay, F-91405 Orsay, France rachel.bawden@limsi.fr Abstract In this paper, we address the problem of generating English tag questions (TQs) (e.g. it is, isn’t it?) in Machine Transla- tion (MT). We propose a post-edition so- lution, formulating the problem as a multi- class classification task. We present (i) the automatic annotation of English TQs in a parallel corpus of subtitles and (ii) an ap- proach using a series of classifiers to pre- dict TQ forms, which we use to post-edit state-of-the-art MT outputs. Our method provides significant improvements in En- glish TQ translation when translating from Czech, French and German, in turn im- proving the fluidity, naturalness, grammat- ical correctness and pragmatic coherence of MT output. 1 Introduction When it comes to the machine translation (MT) of discourse, revisiting the question of what consti- tutes a high quality translation is essential; which aspects of language should be tackled and how to evaluate them. A first step is to identify the many stylistic aspects of speech that pose a prob- lem for current MT techniques and to study how they could be taken into account and evaluated. We take a step in this direction by addressing a new aspect of discourse in MT, related to speaker attitude and style: the English tag question (here- after TQ), i.e. utterances such as catchy, ain’t it? and it wasn’t him, was it?. When translating into English, TQs present two main challenges. The first is knowing when to generate one. Similar to the translation of discourse connectives, TQ use is a question of style and speaker attitude, and im- portantly, there is not often a direct, lexical corre- spondence across languages. TQs are common in English and far less so in other languages, which means that other contextual cues are necessary to determine whether a TQ should appear in the En- glish translation. The second, in particular for canonical TQs (e.g. was it?, isn’t it?), is that the overall grammaticality of the utterance is deter- mined by the correct choice of tag, which, in a similar way to anaphor translation, is grammati- cally dependent on the rest of the MT output. Our aim in this paper is to improve the gener- ation of English TQs in an MT setting with En- glish as the target language. We formulate this as a multi-class classification task, using features from both source sentences and machine translated out- puts. The prediction of the appropriate question tag to use in the English translation (if any) is then used to post-edit MT outputs. Our results, when translating from Czech (CS), French (FR) and German (DE) into English (EN), display sig- nificant improvements, as shown by automatic and manual evaluations (Sec. 5).1 2 English tag questions (TQs) TQs …

Content truncated to 500 words