Viewpoint-Based Measurement of Semantic Similarity between Words

A method of measuring semantic similarity between words using a knowledge-base constructed automatically from machine-readable dictionaries is proposed. The method takes into consideration the fact that similarity changes depending on situation or context, which we calìview-point'. Evaluation shows the proposed method, although based on a simply structured knowledge-base, is superior to other currently available methods. 41.1 Introduction Measuring semantic similarity between words is important for natural language processing when searching text, performing analogical and cases-based reasoning, designing exible human interfaces to databases, and other tasks. We are exploring methods for measuring the similarity between large numbers of daily-use words with an aim toward general applications. In measuring similarity, however, we consider that similarity changes depending on situation or context, which we calìviewpoint'. For example, `horse' is more similar tòpig' than`car' from the viewpoint ofànimal'. On the other hand, `horse' is more similar tòcar' from the viewpoint of`vehicle'. There have been many studies on measuring the semantic similarity of words (for example Tversky77, Suzuki92]), but methods for acquiring knowledge of words were not considered in these studies. Babaguchi et al., proposed nding causes of similarity, which is deened as thèviewpoint' of similarity, from similar cases and executing similar case retrieval by using the obtained viewpoint of similarity Babaguchi94]. However, they assumed that the attributes of the cases were already known, since they focused on applications to an existing database. It is diicult to use their method for measuring similarity between daily-use words, for which the deening attributes may be unclear. In this paper, we propose a method that considers viewpoint when measuring the similarity between words. The method uses a knowledge base that is constructed automatically from machine-readable dictionaries. When computing the similarity between two words, conditioned on view, aspects of the words' deenitions that are important in the view are emphasized when calculating similarity. We constructed an experimental knowledge base and evaluated our method by its pre-1996 Springer-Verlag.