Sub-Sentential Alignment Method by Analogy

SUB-SENTENTIAL ALIGNMENT METHOD BY ANALOGY Tantely Andriamanankasina, Kenji Araki, and Koji Tochinai Graduate School of Engineering, Hokkaido University, Japan Email: {tantely,araki,tochinai}tomedia.eng.hokudai.ac.jp ABSTRACT This paper describes a method for searching word correspondences between pairs of translation sen- tences. In the Example-Based Machine Translation, translation patterns can be extracted easily if word correspondences between pair of translation sentences are defined. The popular methods for aligning bilingual corpus at a sub-sentential level are unable to produce reliable result when the size of the cor- pus is limited, because they are based on statistics. We propose a method for incrementing a word correspondence-included initial corpus automatically. It is appropriate for new languages whose huge corpus as well as machine readable dictionaries are still not available. The method was evaluated with French-Japanese spoken language texts. As the number of translation examples goes beyond 1,000, more than 80.0% of correct word correspondence rates were earned. 1. INTRODUCTION Being able to define accurately and efficiently word correspondences between parallel texts is an issue in corpus-based machine translation. Translation patterns can be extracted very easily if word corre- spondences between pairs of translation sentences are defined. For sub-sentential alignment, a simple consultation of a machine readable dictionary seems to be a very obvious method. However, the presence of unregistered words, and the difference which may be seen in the surface form of words, or after being processed by a lexical analyzer cause problems. In addition, compound words, which often appear in one-to-many or many-to-many word correspondences, are not covered by single word entry-based dictio- naries. The use, not of dictionaries, but of the parallel corpus itself, has been therefore suggested in order to search word correspondences [4, 5, 6]. Sub-sentential alignment methods which have been proposed are based on statistics. The problem with statistical methods is that they are not able to produce reliable results when the corpus size is limited. For unexplored languages whose huge parallel corpus are still not available, these approaches cannot be applied. In addition, correspondences involving multiple tokens have not yet been entirely resolved [4]. Melamed avoids these cases and considers only one-to-one correspondence [5]. On the other hand, although Kitamura' s [6] main goal is not the sub-sentential alignment, but the extraction of translation patterns, possible correspondences between functional words are not considered. It weakens the method since it decreases the number of translation patterns which could be extracted. We have proposed a French-Japanese example-based machine translation which uses word correspondence-included translation examples [1, 2]. As a relatively new field, a huge parallel corpus is not available. Manual construction of translation examples are very time consuming. An automatic method which can efficiently be applied to size-limited corpus is necessary. In this paper, we propose a method for estimating the word correspondences between new pairs of translation sentences by analogy. Similar pairs are selected from a word correspondences-included initial translation examples. Word correspondences between the new pairs are predicted according to the word correspondences between these similar pairs. By doing so, the method …

Content truncated to 500 words