Non-Compositional Language Model and Pattern Dictionary Development for Japanese Compound and Complex Sentences

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 353–360 Manchester, August 2008 Non-Compositional Language Model and Pattern Dictionary Development for Japanese Compound and Complex Sentences Satoru Ikehara, Masato Tokuhisa, Jin’ichi Murakami Tottori University, Koyama-Minami Tottori, 680-8552, Japan {ikehara, tokuhisa, murakami}@ike.tottori-u.ac.jp Abstract To realize high quality machine transla- tion, we proposed a Non-Compositional Language Model, and developed a sen- tence pattern dictionary of 226,800 pat- tern pairs for Japanese compound and complex sentences consisting of 2 or 3 clauses. In pattern generation from a par- allel corpus, Compositional Constituents that could be generalized were 74% of independent words, 24% of phrases and only 15% of clauses. This means that in Japanese-to-English MT, most of the translation results as shown in the parallel corpus could not be obtained by methods based on Compositional Semantics. This dictionary achieved a syntactic coverage of 98% and a semantic coverage of 78%. It will substantially improve translation qual- ity. 1 Introduction A wide variety of machine translation (MT) meth- ods are being studied(Nagao, 1996; Brown et al., 1990; Vogel et al., 2003), but to obtain high-quality translations between languages belonging to dif- ferent families that are alien each other is diffi- cult. Most practical systems still employ a trans- fer method based on compositional semantics. A problem with this method is that it produces trans- lations by separating the syntactic structure from meaning, and is thus liable to lose the meaning of the source text. c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. Better translation quality can be expected from pattern-based MT and example-based MT where the syntactic structure and semantics are handled together. However, pattern-based MT require im- mense pattern dictionaries that are difficult to de- velop(Jung et al., 1999; Uchino et al., 2001). Meanwhile, example-based MT(Nagao, 1984; Sato, 1992; Brown, 1999) obtains translation re- sults by substituting semantically similar elements in structurally matching translation examples, so a pre-prepared pattern dictionary is not needed. However, the capability to substitute a constituent in an example changes from one example to the next, and to automate this judgement is impossi- ble. This problem could be addressed by manually tagging each example beforehand to specify which constituents can be substituted, but the resulting method would be just another pattern-based trans- lation method. Attention has been focused on the use of cog- nitive grammar(Langacker, 1987) and construc- tion grammar(Fillmore, 1988) in the search to find methods that might help to resolve this problem. However, the standards for determining the struc- tural meaning units and the granularity needed for meaning analysis have not been clarified. As a method in which the syntactic structure and meaning are dealt with as an integral whole, a sen- tence pattern (SP)-dictionary called A-Japanese Lexicon has already been developed for Japanese simple sentences(Ikehara et al., 1997). This dictio- nary includes 14,800 valency patterns. The trans- lation quality of Japanese simple sentences into English was 90% and this could be improved up to …

Content truncated to 500 words