... experiment only using the Hownet and Rocling categories, the model had the same purity as the baseline (52.17%) and a low NMI-ratio (4.22%) when using 2 clusters. The purity improved when using more ... purity improved when using more clusters (70.43% with 4 clusters and 76.09% with 6), but it was still much lower than the purity achieved by using the semantic taxonomy and the NMI-ratio dropped ... expectation step (E-step), using equation (3), ∑∏∏===cmiimiimcfpcpcfpcpfffcp1121)|()()|()(), ,,|(~ (3) 3 In our experiments, for verbs with more than 3 senses,...