Standard language variety conversion for content localisation via SMT

Standard Language Variety Conversion for Content Localisation Via SMT Federico Fancellu, Andy Way Morgan O’Brien Centre for Global Intelligent Content McAfee Ireland Inc School of Computing Citygate Dublin City University Mahon Dublin, Ireland. Cork, Ireland. ffancellu@cngl.ie,away@computing.dcu.ie Morgan O’Brien@mcafee.com Abstract Translation between varieties of the same language is a widespread reality in the lo- calisation industry. However, monolin- gual statistical machine translation (SMT) is still a solution that has not yet been ade- quately explored; to the best of our knowl- edge, previous work in this area has never directly applied SMT to varieties of the same language for the precise purpose of reducing the time and cost of human trans- lation and editing of content that needs to be localised. In this paper, we start exploring the prob- lem by deploying SMT to translate Brazil- ian Portuguese into European Portuguese. Our exploration mainly takes into consid- eration the use of bilingual dictionaries to guide the decoder and modify the transla- tion output. We also consider the option of mining a bilingual dictionary from word alignments obtained after standard SMT training. On good-quality data provided by Intel, we show that the SMT baseline already consti- tutes a strong system which in a number of experiments we fail to improve upon. We conjecture that bilingual dictionaries mined from client data would help if more heterogeneous training data were to be added. 1 Introduction Localising content does not only involve trans- lating across different languages, but often also translating between varieties of the same language. c⃝ 2014 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. These varieties might differ in different repects, in- cluding spelling (e.g. British English colour vs. American English colour), lexicon (e.g. British English autumn vs. American English fall), word usage (e.g. British English I’m pissed off vs. American English I’m pissed), grammar (e.g. Irish English You’re after spilling my pint vs. British English You’ve just spilt my pint), etc. Consid- ering that often such translation tasks are carried out by humans, monolingual translation becomes costly and time-consuming, especially when one takes into account how much the two languages have in common. Deploying Statistical Machine Translation (SMT, e.g. Koehn et al. (2003)) would appear to offer a solution to the problem. Given that the two varieties are essentially the same language except for some minor differences, we expect most of the translation variants to be captured by an SMT system. Moreover, we rely on the SMT system to be able to capture those structures that are not only acceptable in a language variety but are also preferable; in a rule-based system (RBMT), these could only be handled by complex hand-written rules. The present study was run as a short-term (3∼4- week) innovation project between CNGL and In- tel. The main goal of the present paper is to as- sess to what extent automatic methods can deliver a good translation between language varieties for localised content. For this reason, in the …

Content truncated to 500 words