A graph-based approach to the automatic generation of multilingual keyword clusters

In this paper, we report an effective graph-theoretic method for generating Japanese and English bilingual keyword clusters using the keyword lists assigned to academic papers by the authors where each of the generated clusters contains keywords with similar meanings from both languages. The advantages of the method are that (i) various domain-dependent keyword pairs useful for IR can be automatically extracted and clustered, (ii) the computation cost is reasonable, and (iii) low-frequency keywords can be properly treated and maintained for later use in IR applications. We apply the method to a set of Japanese and English keywords extracted from academic conference papers in computer science, and show that our method gives a very promising result.