Evaluation of clustering algorithms for word sense disambiguation

Word sense disambiguation in text is still a difficult problem as the best supervised methods require laborious and costly preparation of training data. This work focuses on evaluation of a few selected clustering algorithms in the task of word sense disambiguation. We used five datasets for two languages (English and Polish). Five clustering algorithms (k-means, k-medoids, hierarchical agglomerative clustering, hierarchical divisive clustering, graph-partitioning-based clustering) and two weighting schemes were tested. The best parameters of the algorithms were chosen using 5 × 2 cross validation. BCubed measure was employed for evaluation of clustering. We conclude that with these settings agglomerative hierarchical clustering achieves best results for all the datasets.