Example-based Machine Translation Pursuing Fully Structural NLP

Example-based Machine Translation Pursuing Fully Structural NLP Sadao Kurohashi, Toshiaki Nakazawa, Kauffmann Alexis, Daisuke Kawahara Graduate School of Information Science and Technology University of Tokyo {kuro, nakazawa, alexis, kawahara}@kc.t.u-tokyo.ac.jp Abstract We are conducting Example-Based Machine Translation re- search aiming at the improvement both of structural NLP and machine translation. This paper describes UTokyo system challenged IWSLT05 Japanese-English translation tasks. 1. Introduction We are conducting research on Example-Based Machine Translation, or EBMT [1] aiming at the improvement both of structural NLP and machine translation. Machine translation has been actively studied recently, and the major approach is Statistical Machine Translation, or SMT. EBMT and SMT have something in common and something different. The important common feature is to use bilingual corpus, or translation examples, for the transla- tion of new inputs. Both methods exploit translation knowl- edge implicitly embedded in translation examples, and make MT system maintenance and improvement much easier com- pared with Rule-Based Machine Translation. The difference is that SMT supposes bilingual corpus is the only available resource (but not a bilingual lexicon and parsers); EBMT does not consider such a constraint. SMT basically combines words or phrases (relatively small pieces) with high probability [2]; EBMT tries to use larger translation examples. When EBMT tries to use larger exam- ples, it had better handle examples which are discontinuous as a word-string, but continuous structurally. Accordingly, though it is not inevitable, EBMT naturally seeks syntactic information. The difference in the problem setting is important. SMT is a natural approach when linguistic resources such as parsers and a bilingual lexicon are not available. On the other hand, in case of such linguistic resources are available, it is also natural to see how accurate MT can be achieved using all the available resources. We chose the latter problem setting and conducting EBMT research, and here we would like to mention two rea- sons we chose this setting. One reason is that we are aiming at the improve- ment of structural NLP. We have been conducting re- search on Japanese morphological analyzer, parser, and anaphora/omission analyses. MT is considered as an appli- cation of these fundamental technologies. Amelioration of �� � � � �� �� � � ����� �� ��� � the car came at me from the side at the intersection the car came at me from the side at the intersection (cross) (point) (suddenly) (that) (car) (rush out) Figure 1: An example of parallel sentence alignment. (The root of a tree is placed at the extreme left and phrases are placed from top to bottom. Correspondences of underlined words were detected by a bilingual dictionary.) fundamental NLP technologies naturally improves applica- tions, and applications give some feedback to fundamental NLP, pointing the shortcomings. Needless to say, MT is not the only NLP application, and monolingual NLP applications such as man-machine interface and information retrieval can benefit from the improvement of fundamental NLP. The second point is that, in practice, we often encounter cases to which EBMT problem setting is suitable. That is, …

Content truncated to 500 words