A tool for the acquisition of Japanese-English machine translation rules using inductive learning techniques

Addresses the problem of constructing translation rules for ALT-J/E-a knowledge-based Japanese-English translation system developed at NTT. We introduce the system ATRACT, which is a semi-automatic knowledge acquisition tool designed to facilitate the construction of the desired translation rules through the use of inductive machine learning techniques. Rather than building rules by hand from scratch, a user of ATRACT can obtain good candidate rules by providing the system with a collection of examples of Japanese sentences along with their English translations. This learning task is characterized by two factors: (i) it involves exploiting a huge amount of semantic information as background knowledge; (ii) training examples are “ambiguous”. Currently, two learning methods are available in ATRACT. Experiments show that these methods lead to rules that are very close to those composed manually by human experts given only a reasonable number of examples. These results suggest that ATRACT will significantly contribute to reducing the cost and improving the quality of ALT-J/E translation rules