Abstract
We propose a new method for organizing the numerous collocates into semantic thesaurus categories. The approach introduces a thesaurus-based semantic classification model automatically learning semantic relations for classifying adjective-noun (A-N) and verb-noun (V-N) collocations into different categories. Our model uses a random walk over weighted graph derived from WordNet semantic relation. We compute a semantic label stationary distribution via an iterative graphical algorithm. The performance for semantic cluster similarity and the conformity of semantic labels are both evaluated. The resulting semantic classification establishes as close consistency as human judgments. Moreover, our experimental results indicate that the thesaurus structure is successfully imposed to facilitate grasping concepts of collocations. It might improve the performance of the state-of-art collocation reference tools.