IRIS: English-Irish Machine Translation System

IRIS: English-Irish Machine Translation System Mihael Arcan†, Caoilfhionn Lane†, Eoin ´O Droighne´ain‡, Paul Buitelaar† †Insight Centre for Data Analytics, National University of Ireland, Galway [firstname.lastname]@insight-centre.org ‡Acadamh na hOllscola´ıochta Gaeilge, National University of Ireland, Galway eoin.odroighneain@oegaillimh.ie Abstract We describe IRIS, a statistical machine translation (SMT) system for translating from English into Irish and vice versa. Since Irish is considered an under-resourced language with a limited amount of machine-readable text, building a machine translation system that produces reasonable translations is rather challenging. As translation is a difficult task, current research in SMT focuses on obtaining statistics either from a large amount of parallel, monolingual or other multilingual resources. Nevertheless, we collected available English-Irish data and developed an SMT system aimed at supporting human translators and enabling cross-lingual language technology tasks. Keywords: Statistical Machine Translation, Under-resourced languages, Irish language 1. Introduction The META-NET Language White Paper Series1 showed that English has the best language technology support amongst all European languages, followed by languages such as Dutch, French, German, Italian and Spanish with moderate support. The same study showed that transla- tions from moderately supported languages, such as French or Spanish, into English can achieve acceptable quality for many practical applications. Due to the lack of good trans- lation resources, the Irish language has been categorised as a weak or not supported language by the META-NET re- port (Judge et al., 2012). Despite the limited amount of resources, the demonstrated English-Irish SMT system IRIS2 can nevertheless support human translators and enable information access across Irish and English language and culture. Additionally, we believe that through an SMT system, the broader commu- nity can gain access to information that might have other- wise been unavailable. 2. Related Work Past research on translation systems for under-resourced languages focused on improving translation quality either by collaboratively collecting or generating parallel data needed for SMT models. Additionally, research focused on using closely related languages for translation improvement or using a pivot language to overcome the data sparseness. With the aim of language preservation, Lewis and Yang (2012) show an SMT system for English to and from White Hmong (Hmong-Mien language). They built their system from dictionary entries and translations of introductions or 1http://www.meta-net.eu/whitepapers/ key-results-and-cross-language-comparison 2http://server1.nlp.insight-centre.org/ iris/ phrases from localisation projects. To extend their parallel resources, they manually searched for Hmong phrases and its translations on the web, whereby they collected around 45,000 parallel sentences in overall. A different deploy- ment of SMT systems in an under-resourced scenario was shown in Lewis et al. (2011) and Lewis (2010) as a con- sequence of the earthquake crisis in Haiti supporting emer- gency responders to find trapped people. Differently, Babych et al. (2007) compare results be- tween a direct transfer of an SMT system (source→target language) and translations via a cognate language (source→pivot→target language). Their approach focused on Slavic languages with Russian as the pivot language. The results showed the efficiency of the usage of dictionar- ies, grammars as well as lexical and syntactic similarities of closely related languages for translation …

Content truncated to 500 words