A Syntactic Skeleton for Statistical Machine Translation

A Syntactic Skeleton for Statistical Machine Translation Bart Mellebeek, Karolina Owczarzak, Declan Groves, Josef Van Genabith and Andy Way Dublin City University, National Centre for Language Technology, Dublin 9, Ireland {mellebeek | owczarzak | dgroves | josef | away}@computing.dcu.ie Abstract We present a method for improving statistical machine translation perfor- mance by using linguistically motivated syntactic information. Our algo- rithm recursively decomposes source language sentences into syntactically simpler and shorter chunks, and recomposes their translation to form target language sentences. This improves both the word order and lexical selection of the translation. We report statistically significant relative improvements of 3.3% BLEU score in an experiment (English→Spanish) carried out on an 800-sentence test set extracted from the Europarl corpus. 1 Introduction Almost all research in MT being carried out today is corpus-based, with by far the most dominant paradigm being phrase-based Sta- tistical Machine Translation (SMT). Phrase- based models have recently achieved con- siderable improvements in translation qual- ity; however, they still face difficulty when it comes to modeling long-distance depen- dencies or differences in word order between source and target languages. An obvious way to help overcome these obstacles is to try to add a syntactic level to the mod- els. While a number of attempts have been made to incorporate syntactic knowledge into phrase-based SMT, this has led to little improvement in translation and the loss of language-independence for the systems. Our novel approach uses TransBooster, a wrapper technology designed to improve the output of wide-coverage MT systems (Melle- beek, Khasin, Van Genabith, & Way, 2005) by exploiting the fact that both rule-based and statistical MT systems tend to per- form better at translating shorter sentences than longer ones. TransBooster decom- poses source language sentences into syntac- tically simpler and shorter chunks, sends the chunks to a baseline MT system and recom- poses the translated output into target lan- guage sentences. It has already proved suc- cessful in experiments with rule-based MT systems (Mellebeek, Khasin, Owczarzak, Van Genabith, & Way, 2005). In this paper we apply the TransBooster wrapper technol- ogy to a state-of-the-art phrase-based En- glish → Spanish SMT model constructed with Pharaoh (Koehn, 2004) and we re- port a statistically significant improvement in BLEU and NIST score. The paper is organised as follows. In sec- tion 2, we give a short overview of the most relevant methods that incorporate syntactic knowledge in SMT models. We explain our approach in section 3 and demonstrate it with a worked example. Sections 4 and 5 contain the description, results and analy- sis of our experiments. We summarize our findings in section 6. 2 Related Research One of the major difficulties SMT faces is its inability to model long-distance depen- dencies and correct word order for many language pairs. In this respect, phrase- based SMT systems fare much better than word-based systems, but are still far from perfect. As reported by (Koehn, Och, & Marcu, 2003), even increasing phrase length above three words does not lead to signifi- cant improvements due …

Content truncated to 500 words