Automatic Discovery of Definition Patterns Based on the MDL Principle

A number of approaches to construct a knowledge base for machines have been proposed. The Cyc project[1] manually constructed a knowledge base by using an artificial knowledge representation language. Such an approach, however, has not succeeded because of the heavy cost of construction and the difficulty in maintaining its consistency. On the other hand, there is a growing interest in approaches which generate a knowledge base from a large amount of text corpora automatically. Among several types of corpora, dictionaries are a promising resource for a knowledge base[2, 3]. A dictionary consists of a set of definitional sentences about words or concepts, usually using sentential patterns. This paper proposes a method of discovering these definition patterns. Once definition patterns are discovered, they can be useful for automatically generating of knowledge base from a dictionary.