Integration of POStag-based Source Reordering into SMT Decoding by an Extended Search Graph

Integration of POStag-based source reordering into SMT decoding by an extended search graph Josep M. Crego TALP Research Center (UPC) Barcelona, 08034 jmcrego@gps.tsc.upc.edu Jos´e B. Mari˜no TALP Research Center (UPC) Barcelona, 08034 canton@gps.tsc.upc.edu Abstract This paper presents a reordering frame- work for statistical machine translation (SMT) where source-side reorderings are integrated into SMT decoding, allowing for a highly constrained reordered search graph. The monotone search is extended by means of a set of reordering patterns (linguistically motivated rewrite patterns). Patterns are automatically learnt in train- ing from word-to-word alignments and source-side Part-Of-Speech (POS) tags. Traversing the extended search graph, the decoder evaluates every hypothesis making use of a group of widely used SMT models and helped by an addi- tional Ngram language model of source- side POS tags. Experiments are reported on the Euparl task (Spanish-to-English and English-to- Spanish). Results are presented regard- ing translation accuracy (using human and automatic evaluations) and computational efficiency, showing significant improve- ments in translation quality for both trans- lation directions at a very low computa- tional cost. 1 Introduction In statistical machine translation, the use of reorder- ing strategies allows for an important improvement in translation accuracy, specially when translating between language pairs with high disparity in word order. On the other hand, when arbitrary word re- orderings are permitted, the search problem is clas- sified NP-complete (Knight, 1999), while polyno- mial time search algorithms can be obtained under monotone conditions. The first SMT systems introducing reordering ca- pabilities were founded on the brute force of com- puters, aiming at finding the best hypothesis through traversing a fully reordered graph (the whole per- mutations of source-side words are allowed in the search). This approach is computationally very ex- pensive, even for very short input sentences. There- fore, different distance-based reordering constraints must be used to make the search feasible: ITG (Wu, 1996), IBM (Berger et al., 1996), Local (Kanthak et al., 2005), MaxJumps (Crego et al., 2005), etc. The use of these constraints implies a necessary balance between translation accuracy and efficiency. Typically, a distance-based reordering model is used in the search to penalize longer reorderings, only allowed when well supported by the rest of models. Obviously, this model does not follow any property of language. Lexicalized reordering mod- els, which use distance of words seen in train to score reorderings in search, (Koehn et al., 2005) , (Kumar and Byrne, 2005) have also been introduced. A main criticism to this brute force approach is that it does not make use of any linguistic informa- tion, while in linguistic theory, reorderings between linguistic phrases in different language pairs are well described. Lately, some SMT systems have introduced lin- guistic information in order to tackle the reordering 29 Proceedings of the 7th Conference of the Association for Machine Translation in the Americas, pages 29-36, Cambridge, August 2006. ©2006 The Association for Machine Translation in the Americas problem: • Reordering encoded within translation units in form of hierarchical units (Chiang, 2005), or phrases with gaps (Simard …

Content truncated to 500 words