Induction of Japanese-English Translation Rules from Ambiguous Examples and a Large Semantic Hierarchy

概要

論文の詳細を見る
This work addresses the problem of constructing translation rules for ALT-J/E - a knowledge-based Japanese-English translation system developed at NTT. We present our work towards the automatic acquisition of the condition part of these rules from examples of Japanese sentences and their English translations. From the viewpoint of machine learning research, this learning task is challenging for two reasons : (i) It involves exploiting a huge amount of background knowledge. (ii) Training examples in this domain are ambiguous in the sense that each of the variables describing an example could have more than one value. We investigate two approaches. First, we apply Haussler's algorithm for learning internal disjunctive concepts. This algorithm is designed to directly exploit the kind of background knowledge we have. However, there is no obvious way to make this algorithm learn from ambiguous examples. Second, we apply Quinlan's ID3 algorithm. As it is, this algorithm is not able to utilize the background knowledge of our domain, nor is it capable of dealing with ambiguous training examples. We give, however, a transformation method to "compile" the relevant background knowledge along with the ambiguous training examples into a modified set of training examples on which we can directly run ID3. Our experimental results show that the rules learned directly from ambiguous examples in the latter approach are quite as accurate as those obtained by Haussler's algorithm with the ambiguity explicitly removed. Overall, the experiments show that our machine learning approach yields rules that are highly accurate compared to the manually created rules.
社団法人人工知能学会の論文
1994-09-01

Induction of Japanese-English Translation Rules from Ambiguous Examples and a Large Semantic Hierarchy

スポンサーリンク

概要

著者

関連論文

スポンサーリンク