Getting to Know Moses: Initial Experiments on German-English Factored Translation

Proceedings of the Second Workshop on Statistical Machine Translation, pages 181–184, Prague, June 2007. c⃝2007 Association for Computational Linguistics Getting to know Moses: Initial experiments on German–English factored translation Maria Holmqvist, Sara Stymne, and Lars Ahrenberg Department of Computer and Information Science Linköpings universitet, Sweden {marho,sarst,lah}@ida.liu.se Abstract We present results and experiences from our experiments with phrase-based statisti- cal machine translation using Moses. The paper is based on the idea of using an off- the-shelf parser to supply linguistic infor- mation to a factored translation model and compare the results of German–English translation to the shared task baseline sys- tem based on word form. We report partial results for this model and results for two simplified setups. Our best setup takes ad- vantage of the parser’s lemmatization and decompounding. A qualitative analysis of compound translation shows that decom- pounding improves translation quality. 1 Introduction One of the stated goals for the shared task of this workshop is “to offer newcomers a smooth start with hands-on experience in state-of-the-art statis- tical machine translation methods”. As our previ- ous research in machine translation has been mainly concerned with rule-based methods, we jumped at this offer. We chose to work on German-to-English trans- lation for two reasons. Our primary practical inter- est lies with translation between Swedish and Eng- lish, and of the languages offered for the shared task, German is the one closest in structure to Swedish. While there are differences in word order and morphology between Swedish and German, there are also similarities, e.g., that both languages represent nominal compounds as single ortho- graphic words. We chose the direction from Ger- man to English because our knowledge of English is better than our knowledge of German, making it easier to judge the quality of translation output. Experiments were performed on the Europarl data. With factored statistical machine translation, different levels of linguistic information can be taken into account during training of a statistical translation system and decoding. In our experi- ments we combined syntactic and morphological factors from an off-the-shelf parser with the fac- tored translation framework in Moses (Moses, 2007). We wanted to test the following hypotheses: • Translation models based on lemmas will im- prove translation quality (Popovič and Ney, 2004) • Decompounding German nominal compounds will improve translation quality (Koehn and Knight, 2003) • Re-ordering models based on word forms and parts-of-speech will improve translation qual- ity (Zens and Ney, 2006). 2 The parser The parser, Machinese Syntax, is a commercially available dependency parser from Connexor Oy 1. It provides each word with lemma, part-of-speech, morphological features and dependency relations (see Figure 1). In addition, the lemmas of com- pounds are marked by a ‘#’ separating the two parts of the compound. For the shared task we only used shallow linguistic information: lemma, part- of-speech and morphology. The compound bound- ary identification was used to split noun com- 1 Connexor Oy, http://www.connexor.com. 181 pounds to make the German input more similar to English text. 1 Mit mit pm>2 …

Content truncated to 500 words