Sheffield Systems for the Finnish-English WMT Translation Task

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 172–176, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. Sheffield Systems for the Finnish-English WMT Translation Task Karin Sim Smith Lucia Specia David Steele Department of Computer Science The University of Sheffield Sheffield, UK {kmsimsmith1, l.specia, dbsteele1}@sheffield.ac.uk Abstract This paper provides an overview of the Sheffield University submission to the WMT15 Translation Task for the Finnish- English language pair. The submitted translations were created from a system built using the CDEC decoder. Finnish is a morphologically rich language with ele- ments such as nouns and verbs carrying a large number of inflectional types. Con- sequently, our improvements are based on morphology and include preprocessing steps to handle of morphological inflec- tions inherent in the language, and which otherwise result in lexical sparsity and loss of information. 1 Introduction This paper outlines The University of Sheffield’s submission for the shared translation task, which is part of the 2015 Workshop on Machine Trans- lation. We participated in the Finnish-English lan- guage pair task which used news-test-2015 data. 23 systems from 12 organisations took part in this task. Finnish is an inflectional language containing a productive morphology. The morphological phe- nomena can lead to a great many inflectional forms. This complex productive morphology can be a barrier to machine translation, with many forms unseen at training. As such, our work was focussed on handling the morphological variation in Finnish with the aim of extracting and transfer- ring as much information as possible - in terms of nominal forms and declensions. For this paper we describe our baseline system in Section 3, followed by our improvements in Section 4 and potential gains in Section 5. We re- port our results in Section 6. 2 Related work In terms of previous work in the translation of morphologically rich languages in MT, Finnish- English has previously featured as a language pair, in the 2005 shared task (Koehn and Monz, 2005). Chahuneau et al. (2013) experimented specifi- cally with models into morphologically rich lan- guages, we opted to do from Finnish, as a mor- phologically rich language, into English. Their ap- proach is, however, more systematic, deploying a morphological grammar. Another approach, used by Ammar et al. (2013) is that of synthetic translation options, supple- menting the phrase tables to compensate for the sparseness in translating from/to highly inflected languages. Luong et al. (2010) also investigate morpheme- level extraction, but integrate this into the decod- ing process itself, instead of the pre-processing step we have. They also incorporate unsuper- vised morphological analysis and do not rely on language-specific tools, whereas we used a Finnish parser for our morphological analysis 3 Baseline system For our decoder we used CDEC (Dyer et al., 2010), which essentially is used for rule extraction and decoding. CDEC uses synchronous context- free grammars (SCFGs) as the model for natural language syntax. The initial tokenization and lower casing were performed using the ‘tokenize-anything’ and ‘low- ercase.pl’ scripts respectively. They are both in- …

Content truncated to 500 words