Abstract:
In this paper, we describe a weakly supervised bootstraping algorithm that reads Web texts and learns taxonomy terms. The bootstrapping algorithm starts with two seed words (a seed hypernym (Root concept) and a seed hyponym) that are inserted into a doubly anchored hyponym pattern. In alternating rounds, the algorithm learns new hyponym terms and new hypernym terms that are subordinate to the Root concept. We conducted an extensive evaluation with human annotators to evaluate the learned hyponym and hypernym terms for two categories: animals and people.