Using Shallow Syntax Information to Improve Word Alignment and Reordering for SMT

Proceedings of the Third Workshop on Statistical Machine Translation, pages 53–61, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Using Shallow Syntax Information to Improve Word Alignment and Reordering for SMT Josep M. Crego TALP Research Center Universitat Polit`ecnica de Catalunya 08034 Barcelona, Spain jmcrego@gps.tsc.upc.edu Nizar Habash Center for Computational Learning Systems Columbia University New York, NY 10115, USA habash@ccls.columbia.edu Abstract We describe two methods to improve SMT accuracy using shallow syntax information. First, we use chunks to refine the set of word alignments typically used as a starting point in SMT systems. Second, we extend an N-gram- based SMT system with chunk tags to better account for long-distance reorderings. Exper- iments are reported on an Arabic-English task showing significant improvements. A human error analysis indicates that long-distance re- orderings are captured effectively. 1 Introduction Much research has been done on using syntactic in- formation in statistical machine translation (SMT). In this paper we use chunks (shallow syntax infor- mation) to improve an N-gram-based SMT system. We tackle both the alignment and reordering prob- lems of a language pair with important differences in word order (Arabic-English). These differences lead to noisy word alignments, which lower the ac- curacy of the derived translation table. Addition- ally, word order differences, especially those span- ning long distances and/or including multiple levels of reordering, are a challenge for SMT decoding. Two improvements are presented here. First, we reduce the number of noisy alignments by using the idea that chunks, like raw words, have a transla- tion correspondence in the source and target sen- tences. Hence, word links are constrained (i.e., noisy links are pruned) using chunk information. Second, we introduce rewrite rules which can han- dle both short/medium and long distance reorder- ings as well as different degrees of recursive applica- tion. We build our rules with two different linguistic annotations, (local) POS tags and (long-spanning) chunk tags. Despite employing an N-gram-based SMT system, the methods described here can also be applied to any phrase-based SMT system. Align- ment and reordering are similarly used in both ap- proaches. In Section 2 we discuss previous related work. In Section 3, we discuss Arabic linguistic issues and motivate some of our decisions. In Section 4, we describe the N-gram based SMT system which we extend in this paper. Sections 5 and 6 detail the main contributions of this work. In Section 7, we carry out evaluation experiments reporting on the accuracy re- sults and give details of a human evaluation error analysis. 2 Related Work In the SMT community, it is widely accepted that there is a need for structural information to account for differences in word order between different lan- guage pairs. Structural information offers a greater potential to learn generalizations about relationships between languages than flat-structure models. The need for these ‘mappings’ is specially relevant when handling language pairs with very different word or- der, such as Arabic-English or Chinese-English. Many alternatives have been proposed on using syntactic information in SMT systems. …

Content truncated to 500 words