Morphological Segmentation and OPUS for Finnish-English Machine Translation

Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 177–183, Lisboa, Portugal, 17-18 September 2015. c⃝2015 Association for Computational Linguistics. Morphological Segmentation and OPUS for Finnish-English Machine Translation J¨org Tiedemann1, Filip Ginter2, and Jenna Kanerva2,3 1 Department of Linguistics and Philology, Uppsala University, Sweden 2 Department of IT, University of Turku, Finland 3 University of Turku Graduate School (UTUGS), Finland jorg.tiedemann@lingfil.uu.se, figint@utu.fi, jmnybl@utu.fi Abstract This paper describes baseline systems for Finnish-English and English-Finnish ma- chine translation using standard phrase- based and factored models including mor- phological features. We experiment with compound splitting and morphological seg- mentation and study the effect of adding noisy out-of-domain data to the parallel and the monolingual training data. Our re- sults stress the importance of training data and demonstrate the effectiveness of mor- phological pre-processing of Finnish. 1 Introduction The basic goal of our submissions is to establish some straightforward baselines for the translation between Finnish and English using standard tech- nology such as phrase-based and factored statisti- cal machine translation, in preparation for a more focused future effort in combination with the state- of-the-art techniques in SMT for morphologically complex languages (see e.g. (Fraser et al., 2012)). The translation between Finnish and English (in both directions) is a new task in this year’s work- shop adding a new exciting challenge to the es- tablished setup. The main difficulty in this task is to manage the rich morphology of Finnish which has several implications on training and expected results with standard SMT models (see the illus- tration in Figure 1). Moreover, the monolingual and parallel training data is substantially smaller which makes the task even tougher compared with other languages pairs in the competition. In our contribution, we focus on Finnish-English empha- sizing the need of additional training data and the necessity of morphological pre-processing. In par- ticular, we explore the use of factored models with multiple translation paths and the use of morpho- logical segmentation based on proper morpholog- ical annotation and simple rule-based heuristics. Syksyll¨a taidemuseossa avataan uudet n¨ayttelyt Autumn+ADE art museum+INE open+PASS new+PL exhibi- tion+PL In autumn in art museum will be opened new exhibitions New exhibitions will be opened in the art museum in autumn Figure 1: A sentence illustrating the inflective and compounding nature of Finnish in contrast to En- glish. (ADE, INE: adessive, inessive cases, PASS: passive, PL: plural) We also add noisy out-of-domain data for better coverage and show the impact of that kind of data on translation performance. We also add a system for English-Finnish but without special treatment of Finnish morphology. In this translation direc- tion we only consider the increase of training data which results in significant improvements without any language-specific optimization. In the following, we will first present our sys- tems and the results achieved with our models be- fore discussing the translation produced in more detail. The latter analyses pinpoint issues and prob- lems that provide valuable insights for future devel- opment. 2 Basic Setup and Data Sets All our translation systems …

Content truncated to 500 words