A Direct Syntax-Driven Reordering Model for Phrase-Based Machine Translation

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 849–857, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics A Direct Syntax-Driven Reordering Model for Phrase-Based Machine Translation Niyu Ge IBM T.J.Watson Research Yorktown Heights, NY 10598 niyuge@us.ibm.com Abstract This paper presents a direct word reordering model with novel syntax-based features for sta- tistical machine translation. Reordering models address the problem of reordering source lan- guage into the word order of the target language. IBM Models 3 through 5 have reordering com- ponents that use surface word information but very little context information to determine the traversal order of the source sentence. Since the late 1990s, phrase-based machine translation solves much of the local reorderings by using phrasal translations. The problem of long- distance reordering has become a central re- search topic in modeling distortions. We present a syntax driven maximum entropy reordering model that directly predicts the source traversal order and is able to model arbitrarily long dis- tance word movement. We show that this model significantly improves machine translation qual- ity. 1 Introduction Machine translation reordering models model the problem of the word order when translating a source language into a target language. For exam- ple in Spanish and Arabic, adjectives often come after the nouns they modify whereas in English modifying adjectives usually precede the nouns. When translating Spanish or Arabic into English, the position of the adjectives need to be properly reordered to be placed before the nouns to make fluent English. In this paper, we present a word reordering model that models the word reordering process in transla- tion. The paper is organized as follows. §2 out- lines previous approaches to reordering. §3 details our model and its training and decoding process. §4 discusses experiments to evaluate the model and §5 presents machine translation results. §6 is discussion and conclusion. 2 Previous Work The word reordering problem has been one of the major problems in statistical machine translation (SMT). Since exploring all possible reorderings of a source sentence is an NP-complete problem (Knight 1999), SMT systems limit words to be re- ordered within a window of length k. IBM Models 3 through 5 (Brown et.al. 1993) model reorderings based on surface word information. For example, Model 4 attempts to assign target-language posi- tions to source-language words by modeling d(j | i, l, m) where j is the target-language position, i is the source-language position, l and m are respectively source and target sentence lengths. These models are not effective in modeling reorderings because they don’t have enough context and lack structural information. Phrase-based SMT systems such as (Koehn et.al. 2003) move from using words as translation units to using phrases. One of the advantages of phrase- based SMT systems is that local reorderings are inherent in the phrase translations. However, phrase-based SMT systems capture reordering in- stances and not reordering phenomena. For exam- ple, if the Arabic phrase “the car red” and its English translation “the red …

Content truncated to 500 words