Graph-Based Clustering for Semantic Classification of Onomatopoetic Words

This paper presents a method for semantic classication of onomatopoetic words like “ひゅーひゅー (hum)” and “からん ころん (clip clop)” which exist in every language, especially Japanese being rich in onomatopoetic words. We used a graph-based clustering algorithm called Newman clustering. The algorithm calculates a simple quality function to test whether a particular division is meaningful. The quality function is calculated based on the weights of edges between nodes. We combined two different similarity measures, distributional similarity, and orthographic similarity to calculate weights. The results obtained by using the Web data showed a 9.0% improvement over the baseline single distributional similarity measure.