A Study of Style in Machine Translation: Controlling the Formality of Machine Translation Output

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2814–2819 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics A Study of Style in Machine Translation: Controlling the Formality of Machine Translation Output Xing Niu Dept. of Computer Science University of Maryland College Park xingniu@cs.umd.edu Marianna Martindale iSchool University of Maryland College Park mmartind@umd.edu Marine Carpuat Dept. of Computer Science University of Maryland College Park marine@cs.umd.edu Abstract Stylistic variations of language, such as formality, carry speakers’ intention be- yond literal meaning and should be con- veyed adequately in translation. We pro- pose to use lexical formality models to control the formality level of machine translation output. We demonstrate the effectiveness of our approach in empiri- cal evaluations, as measured by automatic metrics and human assessments. 1 Introduction Automatically analyzing and generating natural language requires capturing not only what is said, but also how to say it. Consider the sentences “anybody hurt?” and “is someone wounded?”. The first one is less formal than the second one, and carries information beyond its literal meaning, such as the situation in which it might be used. Such differences in formality have been identi- fied as an important dimension of style (Trudgill, 1992) or tone (Halliday, 1978) variation. In this paper, we build on prior computational work that has focused on analyzing formality of texts (Lahiri and Lu, 2011; Brooke and Hirst, 2013; Pavlick and Nenkova, 2015; Pavlick and Tetreault, 2016) with a different aim: modeling formality for the purpose of controlling style in applications that generate language, with a focus on machine translation. Human translators trans- late a document for a specific audience (Nida and Taber Charles, 1969), and often ask what is the expected tone of the content when taking a new translation job. We design a machine translation system that operates under similar conditions and explicitly takes an expected level of formality as input. While ultimately we would like systems to preserve the formality of the source, this is a challenging task that requires not only automati- cally inferring the formality of the source, but also understanding how formality differs across lan- guages and cultures. As a first step, we therefore limit our study to the scenario where the expected output formality is given to the MT system as an additional input. We first select a formality model providing the most accurate scores on intrinsic formality datasets. We compare existing lexical formality models and novel variants based on inducing for- mality dimensions or subspaces in vector space models. We then turn to machine translation and show that a lexical formality model can have a positive impact when used to control the formal- ity of machine translation output. When the ex- pected formality matches the reference, we obtain improvement of translation quality evaluated by automatic metrics (BLEU). A human assessment also verified the effectiveness of our proposed sys- tem in generating translations at diverse levels of formality. 2 Formality-Sensitive MT Our goal is to provide systems with the …

Content truncated to 500 words