Building a large-scale commonsense knowledge base by converting an existing one in a different language

This paper describes our effort to build a large-scale commonsense knowledge base in Korean by converting a pre-existing one in English, called ConceptNet. The English commonsense knowledge base is essentially a huge net consisting of concepts and relations. Triplets in the form of ConceptRelation-Concept in the net were extracted from English sentences collected from volunteers through a Web site, who were interested in entering commonsense knowledge. Our effort is an attempt to obtain its Korean version by utilizing a variety of language resources and tools. We not only employed a morphological analyzer and existing commercial machine translation software but also developed our own special-purpose translation and out-of-vocabulary handling methods. In order to handle ambiguity, we also devised a noisy concept filtering and concept generalization methods. Out of the 2.4 million assertions, i.e. triplets of concept-relation-concept, in the English ConceptNet, we generated about 200,000 Korean assertions so far. Based on our manual judgments of a 5% sample, the accuracy was 84.4%.