Translation of Multiword Expressions Using Parallel Suffix Arrays

Translation of Multiword Expressions Using Parallel Suffix Arrays Paul McNamee and James Mayfield Johns Hopkins University Applied Physics Laboratory 11100 Johns Hopkins Road Laurel, Maryland 20723-6099 USA {paul.mcnamee,james.mayfield}@jhuapl.edu Abstract Accurately translating multiword expres- sions is important to obtain good per- formance in machine translation, cross- language information retrieval, and other multilingual tasks in human language technology. Existing approaches to induc- ing translation equivalents of multiword units have focused on agglomerating in- dividual words or on aligning words in a statistical machine translation system. We present a different approach based upon information theoretic heuristics and the exact counting of frequencies of occur- rence of multiword strings in aligned par- allel corpora. We are applying a technique introduced by Yamamoto and Church that uses suffix arrays and longest common prefix arrays. Evaluation of the method in multiple language pairs was performed using bilingual lexicons of domain- specific terminology as a gold standard. We found that performance of 50-70%, as measured by mean reciprocal rank, can be obtained for terms that occur more than 10 or so times. 1 Introduction When processing human language it is difficult to operate only at the level of individual words. While for some tasks, perhaps monolingual infor- mation retrieval in particular, this might seem rea- sonable, for others such as machine translation, cross-language question answering, and translin- gual information retrieval, restriction to processing single words is a significant impediment. There has been much recent interest in computational ap- proaches to dealing with multiword expressions (MWEs) as workshops at ACL-2006, SIGIR-2005, ACL-2004 and other conferences attest. Even identifying what constitutes a good MWE is difficult, and there are a variety of interesting classes such as idiomatic expressions, complex noun phrases, phrasal verbs, and collocations. Many researchers have focused attention on short- length n-grams (notably n=2) or phrases that com- pose a syntactic unit. Longer length n-grams have been eschewed due to computational complexity and data sparseness. Here we examine arbitrary sequences of words and attempt to translate them. To effect translation we will rely on informa- tion-theoretic measures such as Dice scores or Mu- tual Information, and large aligned bilingual texts. To score translation candidates we will compute global and local frequencies of occurrence of sub- strings using the method described by Yamamoto and Church (2001) which is based on suffix arrays and longest common prefix arrays. In the rest of this paper we review earlier work in phrase translation (Section 2) and fully describe our approach (Section 3). In Section 4 details of our evaluation are presented and we conclude the paper with a discussion of our results (Section 5). 2 Related Work The problem of phrasal translation has been at- tempted in three different ways: (1) fusion of trans- lations of component words; (2) translation using word alignments and bilingual suffix trees; and (3) purported phrases induced by word alignments in a SMT system. The present work can be seen as a 100 Proceedings of the 7th Conference of the Association for Machine Translation in the …

Content truncated to 500 words