Improving machine translation via triangulation and transliteration

Improving Machine Translation via Triangulation and Transliteration Nadir Durrani School of Informatics University of Edinburgh dnadir@inf.ed.ac.uk Philipp Koehn School of Informatics University of Edinburgh pkoehn@inf.ed.ac.uk Abstract In this paper we improve Urdu→Hindi⇄English machine trans- lation through triangulation and translit- eration. First we built an Urdu→Hindi SMT system by inducing triangulated and transliterated phrase-tables from Urdu–English and Hindi–English phrase translation models. We then use it to trans- late the Urdu part of the Urdu-English parallel data into Hindi, thus creating an artificial Hindi-English parallel data. Our phrase-translation strategies give an improvement of up to +3.35 BLEU points over a baseline Urdu→Hindi system. The synthesized data improve Hindi→English system by +0.35 and English→Hindi system by +1.0 BLEU points. 1 Introduction Phrase-based machine translation models are capa- ble of producing translations of reasonable quality but only with large quantities of parallel data. Un- fortunately, bilingual data is abundantly available for only a handful of language pairs. The problem of reliably estimating statistical models for trans- lation becomes more of a challenge under sparse data conditions. Previous research has tried to ad- dress this bottleneck in two ways i) by making the best use of the existing small corpus by using gen- eralized representations such as morpho-syntactic analysis and suffix separation (Niessen and Ney, 2004; Popovi´c and Ney, 2004; Haque et al., 2012), ii) by generating additional data/inducing models c⃝ 2014 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. to overcome sparsity (Utiyama and Isahara, 2003; Resnik and Smith, 2003). Techniques like triangu- lation (Cohn and Lapata, 2007; Wu and Wang, 2007) and paraphrasing (Callison-Burch et al., 2006) have also been used to address the problem of data sparsity. Transliteration is shown to be use- ful when the languages in question are closely re- lated (Durrani et al., 2010; Nakov and Tiedemann, 2012). Our work falls in this second category of generating additional/data and models. Hindi and Urdu are widely spoken yet low re- sourced languages. Hindi descends from Sanskrit and is written in Devanagri script, where as Urdu inherits its vocabulary and language phenomenon from several languages (Arabic, Farsi and Turk- ish and Sanskrit) and is written in Arabic script. They are a closely related language pair that share grammatical structure and have a high vocabulary overlap.1 This provides a motivation to build an MT system to create Hindi and Urdu resources by translating one into another. In this paper, we exploit the relatedness of the two languages and bring together the ideas of trian- gulation and transliteration to effectively improve Urdu-to-Hindi machine translation. We make use of a tiny Hindi-Urdu parallel corpus, to build a Urdu-to-Hindi translation system. We then im- prove this system by synthesizing phrase-tables through triangulation and transliteration. We cre- ate a triangulated phrase-table using English as a pivot language following the well-known con- volution model (Utiyama and Isahara, 2007; Wu and Wang, 2007). The new phrase-table is syn- thesized using Hindi-English and Urdu-English phrase-tables. We then use the …

Content truncated to 500 words