Enriching Morphologically Poor Languages for Statistical Machine Translation

Proceedings of ACL-08: HLT, pages 763–770, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Enriching Morphologically Poor Languages for Statistical Machine Translation Eleftherios Avramidis e.avramidis@sms.ed.ac.uk Philipp Koehn pkoehn@inf.ed.ac.uk School of Informatics University of Edinburgh 2 Baccleuch Place Edinburgh, EH8 9LW, UK Abstract We address the problem of translating from morphologically poor to morphologically rich languages by adding per-word linguistic in- formation to the source language. We use the syntax of the source sentence to extract information for noun cases and verb persons and annotate the corresponding words accord- ingly. In experiments, we show improved performance for translating from English into Greek and Czech. For English–Greek, we re- duce the error on the verb conjugation from 19% to 5.4% and noun case agreement from 9% to 6%. 1 Introduction Traditional statistical machine translation methods are based on mapping on the lexical level, which takes place in a local window of a few words. Hence, they fail to produce adequate output in many cases where more complex linguistic phenomena play a role. Take the example of morphology. Predicting the correct morphological variant for a target word may not depend solely on the source words, but re- quire additional information about its role in the sen- tence. Recent research on handling rich morphology has largely focused on translating from rich morphology languages, such as Arabic, into English (Habash and Sadat, 2006). There has been less work on the op- posite case, translating from English into morpho- logically richer languages. In a study of translation quality for languages in the Europarl corpus, Koehn (2005) reports that translating into morphologically richer languages is more difficult than translating from them. There are intuitive reasons why generating richer morphology from morphologically poor languages is harder. Take the example of translating noun phrases from English to Greek (or German, Czech, etc.). In English, a noun phrase is rendered the same if it is the subject or the object. However, Greek words in noun phrases are inflected based on their role in the sentence. A purely lexical mapping of English noun phrases to Greek noun phrases suffers from the lack of information about its role in the sen- tence, making it hard to choose the right inflected forms. Our method is based on factored phrase-based statistical machine translation models. We focused on preprocessing the source data to acquire the needed information and then use it within the mod- els. We mainly carried out experiments on English to Greek translation, a language pair that exemplifies the problems of translating from a morphologically poor to a morphologically rich language. 1.1 Morphology in Phrase-based SMT When examining parallel sentences of such lan- guage pairs, it is apparent that for many English words and phrases which appear usually in the same form, the corresponding terms of the richer target language appear inflected in many different ways. On a single word-based probabilistic level, it is then obvious that for one specific English word e the probability p(f|e) of it being translated into …

Content truncated to 500 words