Improving Mid-Range Re-Ordering Using Templates of Factors

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 372–379, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Improving Mid-Range Reordering using Templates of Factors Hieu Hoang School of Informatics University of Edinburgh h.hoang@sms.ed.ac.uk Philipp Koehn School of Informatics University of Edinburgh pkoehn@inf.ed.ac.uk Abstract We extend the factored translation model (Koehn and Hoang, 2007) to allow trans- lations of longer phrases composed of fac- tors such as POS and morphological tags to act as templates for the selection and re- ordering of surface phrase translation. We also reintroduce the use of alignment in- formation within the decoder, which forms an integral part of decoding in the Align- ment Template System (Och, 2002), into phrase-based decoding. Results show an increase in transla- tion performance of up to 1.0% BLEU for out-of-domain French–English transla- tion. We also show how this method com- pares and relates to lexicalized reordering. 1 Introduction One of the major issues in statistical machine translation is reordering due to systematic word- ordering differences between languages. Often re- ordering is best explained by linguistic categories, such as part-of-speech tags. In fact, prior work has examined the use of part-of-speech tags in pre-reordering schemes, Tomas and Casacuberta (2003). Re-ordering can also be viewed as composing of a number of related problems which can be ex- plained or solved by a variety of linguistic phe- nomena. Firstly, differences between phrase or- dering account for much of the long-range re- ordering. Syntax-based and hierarchical models such as (Chiang, 2005) attempts to address this problem. Shorter range re-ordering, such as intra- phrasal word re-ordering, can often be predicted from the underlying property of the words and its context, the most obvious property being POS tags. In this paper, we tackle the issue of shorter- range re-ordering in phrase-based decoding by presenting an extension of the factored transla- tion which directly models the translation of non- surface factors such as POS tags. We shall call this extension the factored template model. We use the fact that factors such as POS-tags are less sparse than surface words to obtain longer phrase trans- lations. These translations are used to inform the re-ordering of surface phrases. Despite the ability of phrase-based systems to use multi-word phrases, the majority of phrases used during decoding are one word phrases, which we will show in later sections. Using word trans- lations negates the implicit capability of phrases to re-order words. We show that the proposed extension increases the number of multi-word phrases used during decoding, capturing the im- plicit ordering with the phrase translation, lead- ing to overall better sentence translation. In our tests, we obtained 1.0% increase in absolute for French-English translation, and 0.8% increase for German-English translation, trained on News Commentary corpora 1. We will begin by recounting the phrase-based and factored model in Section 2 and describe the language model and lexicalized re-ordering model and the advantages and disadvantages of using these models to influence re-ordering. The pro- …

Content truncated to 500 words