A Probability Model to Improve Word Alignment

A Probability Model to Improve Word Alignment Colin Cherry and Dekang Lin Department of Computing Science University of Alberta Edmonton, Alberta, Canada, T6G 2E8 {colinc,lindek}@cs.ualberta.ca Abstract Word alignment plays a crucial role in sta- tistical machine translation. Word-aligned corpora have been found to be an excellent source of translation-related knowledge. We present a statistical model for comput- ing the probability of an alignment given a sentence pair. This model allows easy in- tegration of context-specific features. Our experiments show that this model can be an effective tool for improving an existing word alignment. 1 Introduction Word alignments were first introduced as an in- termediate result of statistical machine translation systems (Brown et al., 1993). Since their intro- duction, many researchers have become interested in word alignments as a knowledge source. For example, alignments can be used to learn transla- tion lexicons (Melamed, 1996), transfer rules (Car- bonell et al., 2002; Menezes and Richardson, 2001), and classifiers to find safe sentence segmentation points (Berger et al., 1996). In addition to the IBM models, researchers have proposed a number of alternative alignment meth- ods. These methods often involve using a statistic such as φ2 (Gale and Church, 1991) or the log likeli- hood ratio (Dunning, 1993) to create a score to mea- sure the strength of correlation between source and target words. Such measures can then be used to guide a constrained search to produce word align- ments (Melamed, 2000). It has been shown that once a baseline alignment has been created, one can improve results by using a refined scoring metric that is based on the align- ment. For example Melamed uses competitive link- ing along with an explicit noise model in (Melamed, 2000) to produce a new scoring metric, which in turn creates better alignments. In this paper, we present a simple, flexible, sta- tistical model that is designed to capture the infor- mation present in a baseline alignment. This model allows us to compute the probability of an align- ment for a given sentence pair. It also allows for the easy incorporation of context-specific knowl- edge into alignment probabilities. A critical reader may pose the question, “Why in- vent a new statistical model for this purpose, when existing, proven models are available to train on a given word alignment?” We will demonstrate exper- imentally that, for the purposes of refinement, our model achieves better results than a comparable ex- isting alternative. We will first present this model in its most general form. Next, we describe an alignment algorithm that integrates this model with linguistic constraints in order to produce high quality word alignments. We will follow with our experimental results and dis- cussion. We will close with a look at how our work relates to other similar systems and a discussion of possible future directions. 2 Probability Model In this section we describe our probability model. To do so, we will first introduce some necessary no- tation. Let E be an English sentence e1, e2, . . . , …

Content truncated to 500 words