Effects of Morphological Analysis in Translation between German and English

Proceedings of the Third Workshop on Statistical Machine Translation, pages 135–138, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Effects of Morphological Analysis in Translation between German and English Sara Stymne, Maria Holmqvist and Lars Ahrenberg Department of Computer and Information Science Link¨oping University, Sweden {sarst,marho,lah}@ida.liu.se Abstract We describe the LIU systems for German- English and English-German translation sub- mitted to the Shared Task of the Third Work- shop of Statistical Machine Translation. The main features of the systems, as compared with the baseline, is the use of morphologi- cal pre- and post-processing, and a sequence model for German using morphologically rich parts-of-speech. It is shown that these addi- tions lead to improved translations. 1 Introduction Research in statistical machine translation (SMT) increasingly makes use of linguistic analysis in order to improve performance. By including abstract cat- egories, such as lemmas and parts-of-speech (POS), in the models, it is argued that systems can become better at handling sentences for which training data at the word level is sparse. Such categories can be integrated in the statistical framework using factored models (Koehn et al., 2007). Furthermore, by pars- ing input sentences and restructuring based on the result to narrow the structural difference between source and target language, the current phrase-based models can be used more effectively (Collins et al., 2005). German differs structurally from English in sev- eral respects (see e.g. Collins et al., 2005). In this work we wanted to look at one particular aspect of restructuring, namely splitting of German com- pounds, and evaluate its effect in both translation di- rections, thus extending the initial experiments re- ported in Holmqvist et al. (2007). In addition, since German is much richer in morphology than English, we wanted to test the effects of using a sequence model for German based on morphologically sub- categorized parts-of-speech. All systems have been specified as extensions of the Moses system pro- vided for the Shared Task. 2 Part-of-speech and Morphology For both English and German we used the part-of- speech tagger TreeTagger (Schmid, 1994) to obtain POS-tags. The German POS-tags from TreeTagger were re- fined by adding morphological information from a commercial dependency parser, including case, number, gender, definiteness, and person for nouns, pronouns, verbs, adjectives and determiners in the cases where both tools agreed on the POS-tag. If they did not agree, the POS-tag from TreeTagger was chosen. This tag set seemed more suitable for SMT, with tags for proper names and foreign words which the commercial parser does not have. 3 Compound Analysis Compounding is common in many languages, in- cluding German. Since compounding is highly pro- ductive it increases vocabulary size and leads to sparse data problems. Compounds in German are formed by joining words, and in addition filler letters can be inserted or letters can be removed from the end of all but the last word of the compound (Langer, 1998). We have chosen to allow simple additions of letter(s) (-s, -n, -en, -nen, -es, -er, -ien) and simple …

Content truncated to 500 words