6 . 863 Final Project Writeup

The goal of this system is to provide a measure of semantic similarity between any two verbs. The VerbOcean dataset contains various semantic relationships for a set of 3477 unique verbs. The pairwise relationships between verbs contained in VerbOcean were used to create a graph of semantic relations between many verbs. A tree traversal algorithm was implemented so that, given a starting word and a target word, various semantic paths from the starting word to the target word could be found. The VerbOcean dataset was generated by searching for the number of web pages which contained pairs of verbs in various simple syntaxes. The relationships incorporated into VerbOcean are similarity, strength, antonymy, enablement, and precedence. Similarity between two verbs was found by searching for instances of ‘to X and Y’ or ‘Xed and Yed’, while precedence was found by searching for ‘to X and then Y’ or ‘Xed and then Yed’. The number of results found (using Google) corresponded to the strength of a particular relationship between two verbs. The score of the shortest semantic path between two verbs intuitively gives a measure of how closely related those two verbs should be.