Are Unaligned Words Important for Machine Translation?

Proceedings of the 13th Annual Conference of the EAMT, pages 226–233, Barcelona, May 2009 Are Unaligned Words Important for Machine Translation ? Yuqi Zhang Evgeny Matusov Hermann Ney Human Language Technology and Pattern Recognition Lehrstuhl f¨ur Informatik 6 – Computer Science Department RWTH Aachen University, D-52056 Aachen, Germany {yzhang,matusov,ney}@cs.rwth-aachen.de Abstract In this paper, we deal with the problem of a large number of unaligned words in auto- matically learned word alignments for ma- chine translation (MT). These unaligned words are the reason for ambiguous phrase pairs extracted by a statistical phrase-based MT system. In translation, this phrase am- biguity causes deletion and insertion er- rors. We present hard and optional dele- tion approaches to remove the unaligned words in the source language sentences. Improvements in translation quality are achieved both on large and small vocabu- lary tasks with the presented methods. 1 Introduction Word alignment is a key part in the training of a statistical MT system because it provides map- pings of words between each source sentence and its target language translation. Because of the dif- ference in the structure of the involved languages, not all words in the source language have a corre- sponding word in the target language. So in the alignments, no matter manually created or auto- matically learned, some words are aligned, some are not. Current state-of-the-art statistical machine translation is based on phrases. First the word alignments for the training corpus are generated. Then phrase alignments are inferred heuristically from the word alignments. This approach was presented by (Och et al., 1999) and implemented by e. g. (Koehn et al., 2003). Since this widely used phrase extraction method depends on word alignments, it is often assumed that the quality c⃝ 2009 European Association for Machine Translation. of word alignment is critical to the success of translation. However, some research have shown that the large gains in alignment accuracy often lead to, at best, minor gains in translation performance (Lopez and Resnik, 2006). They concluded that it could be more useful to directly investigate ways to reduce the noise in phrase extraction than improving word alignment. The work by (Ma et al., 2007) shows that a good phrase segmentation is important for translation result. Encouraged by the work, this paper explores the influence of the unaligned words on the phrase extraction and machine translation results. We show that the presence of unaligned words causes extraction of “noisy” phrases which can lead to insertion and deletion errors in the translation output. Furthermore, we propose approaches for “hard” and “soft” deletion of the unaligned words on the source language side. We then show that better way to deal with unaligned words can substantially improve translation quality, on both small and large vocabulary tasks. In section 2, we briefly review the word align- ment concept and point out that there is a large number of unaligned words in both manual and automatic alignments used for common translation tasks. In section 3, we explain how the unaligned words affect …

Content truncated to 500 words