The University of Illinois submission to the WMT 2015 Shared Translation Task

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 192–198, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. The University of Illinois submission to the WMT 2015 Shared Translation Task Lane Schwartz, Bill Bryce, Chase Geigle, Sean Massung, Yisi Liu, Haoruo Peng, Vignesh Raja, Subhro Roy and Shyam Upadhyay University of Illinois at Urbana-Champaign Urbana, IL 61801, USA lanes@illinois.edu Abstract In this year’s WMT translation task, Finnish-English was introduced as a lan- guage pair of competition for the first time. We present experiments examining sev- eral variations on a morphologically-aware statistical phrase-based machine transla- tion system for translating Finnish into English. Our system variations attempt to mitigate the issue of rich agglutina- tive morphology when translating from Finnish into English. Our WMT sub- mission for Finnish-English preprocesses Finnish data with omorfi (Pirinen, 2015), a Finnish morphological analyzer. We also present results for two other language pairs with morphologically interesting source languages, namely German-English and Czech-English. 1 Introduction Students enrolled in the Spring 2015 graduate- level course in statistical machine translation (MT) at the University of Illinois were invited to develop MT systems within the context of the 2015 Work- shop on Statistical Machine Translation (WMT) shared translation task. Each group of 2-3 stu- dents chose one language pair, developed a base- line MT system for that language pair using Moses (Koehn et al., 2007), and chose one specific lin- guistic dimension along which to experiment. In this work, we present the results of four groups of experiments — two Finnish-English (§3.1 and §3.2), and one each for Czech-English (§4) and German-English (§5). The first author was the instructor, and the subsequent authors were students in the work described here. 2 Methodology We use the current stable release (v3) of Moses, a state-of-the-art statistical phrase-based machine translation system. We trained translation models using the Eu- roparl corpus (Koehn, 2005), using the latest avail- able versions (v7 for German-English and Czech- English, and v8 for Finnish-English), as well as the Common Crawl corpus and News Com- mentary (v10) corpus for German-English and Czech-English, and the Wiki Headlines corpus for Finnish-English. We trained a back-off language model (LM) with modified Kneser-Ney smoothing (Katz, 1987; Kneser and Ney, 1995; Chen and Goodman, 1998) on the English Gigaword v5 corpus (Parker et al., 2011) using lmplz from KenLM (Heafield et al., 2013). 3 Finnish-English We tried various morphological tokenization schemes on the source language (Finnish) in or- der to mitigate its strong agglutination. The target language (English) was tokenized with the default Moses tokenizer script. 3.1 Finnish tokenization using Morfessor and word-lattices We begin by adapting the lattice technique of Dyer et al. (2009) to Finnish. We train a standard phrase- based machine translation model on a new corpus: on the source side we concatenate the original data with its one-best segmentation according to a Mor- fessor (Creutz and Lagus, 2007) model trained on the original data, and on the target side we simply concatenate it with itself. The result is a …

Content truncated to 500 words