A New Subtree-Transfer Approach to Syntax-Based Reordering for Statistical Machine Translation

Proceedings of the 13th Annual Conference of the EAMT, pages 197–204, Barcelona, May 2009 A New Subtree-Transfer Approach to Syntax-Based Reordering for Statistical Machine Translation Maxim Khalilov and José A.R. Fonollosa Universitat Politècnica de Catalunya Campus Nord UPC, 08034, Barcelona, Spain {khalilov,adrian}@gps.tsc.upc.edu Mark Dras Macquarie University North Ryde NSW 2109, Sydney, Australia madras@ics.mq.edu.au Abstract In this paper we address the problem of translating between languages with word order disparity. The idea of augmenting statistical machine translation (SMT) by using a syntax-based reordering step prior to translation, proposed in recent years, has been quite successful in improving trans- lation quality. We present a new tech- nique for extracting syntax-based reorder- ing rules, which are derived through a syn- tactically augmented alignment of source and target texts. The parallel corpus with reordered source side is then passed to an N-gram-based machine translation system and the obtained results are contrasted with a monotone system performance. In exper- iments, we show significant improvement for the Chinese-to-English translation task. 1 Introduction One of the most challenging problems facing ma- chine translation (MT) is how to place the trans- lated words in the natural order of the target lan- guage. A monotone SMT system suffers from weakness in the distortion model, even if it is able to generate correct word-by-word translation. In this study we propose a reordering model that in- volves both source- and target-side syntax infor- mation in the word reordering process. Our work is inspired by the approach proposed in Imamura et al. (2005), where a complete syntax- driven SMT system based on a two-side subtree transfer is described. In their approach they con- struct a probabilistic non-isomorphic tree mapping model based on a context-free breakdown of the source and target parse trees; extract alignment templates that incorporate the constraints of the parse trees; and apply syntax-based decoding. We c⃝ 2009 European Association for Machine Translation. propose to use a similar non-isomorphic subtree mapping to extract reordering rules, but instead of involving them directly in the translation process, we use them to monotonize the source portion of the bilingual corpus. In the next step, the rules are applied to the source part of the same training corpus chang- ing the source sentence structure such that it more closely matches the word order of the target lan- guage. It leads to a simplification of the translation task due to a shorter average length of bilingual units which it is more likely to see when translat- ing an unseen set. Local and long-range word reorderings are driven by automatically extracted permutation pat- terns operating with source language constituents and underlaid by non-isomorphic subtree transfer. The target-side parse tree utilization is optional, but it greatly affects system performance: it is con- sidered as a filter constraining the reordering rules to the set of patterns covered by both the source- and target-side subtrees. Apart from the reorder- ing rules representing the order of child nodes, a set of additional rewrite rules based on a deep top- down …

Content truncated to 500 words