Word-Based Alignment, Phrase-Based Translation: What’s the Link?

Word-Based Alignment, Phrase-Based Translation: What’s the Link? Adam Lopez Institute for Advanced Computer Studies Department of Computer Science University of Maryland College Park, MD 20742 alopez@cs.umd.edu Philip Resnik Department of Linguistics Institute for Advanced Computer Studies University of Maryland College Park, MD 20742 resnik@umd.edu Abstract State-of-the-art statistical machine trans- lation is based on alignments between phrases – sequences of words in the source and target sentences. The learn- ing step in these systems often relies on alignments between words. It is often as- sumed that the quality of this word align- ment is critical for translation. However, recent results suggest that the relation- ship between alignment quality and trans- lation quality is weaker than previously thought. We investigate this question directly, comparing the impact of high- quality alignments with a carefully con- structed set of degraded alignments. In or- der to tease apart various interactions, we report experiments investigating the im- pact of alignments on different aspects of the system. Our results confirm a weak correlation, but they also illustrate that more data and better feature engineering may be more beneficial than better align- ment. 1 Introduction Machine translation and alignment are closely re- lated problems. In the translation problem, we are given a source sentence, and the task is to output a target sentence which conveys the same meaning. In the process, the system produces a mapping be- tween words of the source sentence and target sen- tence. In the alignment problem, we are given both sentences, and the task is simply to find this map- ping. Alignment can therefore be thought of as a more constrained version of the translation problem, in which we need find only one of the outputs in- stead of both. As in other machine learning problems, statistical machine translation systems learn to make decisions by looking at previous examples of those decisions. Because the decisions made by translation systems produce alignments between source and target sen- tences, the training data must contain examples of these alignment decisions. However, the data avail- able for training translation systems contains only pairs of sentences. There are two solutions to this problem. The first is to treat the unseen alignments as a hidden in- put and apply unsupervised learning methods. The second is to first solve the easier problem of align- ment, and then use supervised methods for learning. Current state-of-the-art models in machine transla- tion are based on alignments between phrases – se- quences of words within each sentence (Och et al., 1999; Koehn et al., 2003; Chiang, 2005; Simard et al., 2005). Unfortunately, unsupervised learning of phrase-based models is intractable. It requires nu- merous approximations and tradeoffs, and often pro- duces poor results (Marcu and Wong, 2002; DeN- ero et al., 2006; Birch et al., 2006). Therefore, su- pervised methods are usually employed (Och et al., 1999; Koehn et al., 2003; Chiang, 2005). This re- quires a method for phrase alignment. However, there are only a few examples of research on the phrase alignment problem …

Content truncated to 500 words