Learning Concept Hierarchies from Text with a Guided Hierarchical Clustering Algorithm

We present an approach for the automatic induction of concept hierarchies from text collections. We propose a novel guided agglomerative hierarchical clustering algorithm exploiting a hypernym oracle to drive the clustering process. By inherently integrating the hypernym oracle into the clustering algorithm, we overcome two main problems of unsupervised clustering approaches relying on the distributional similarity of terms to induce concept hierarchies. First, by only clustering two terms if they have a hypernym in common we make sure that the cluster produced in this way is actually reasonable. Second, by labeling the clusters with the corresponding hypernym we overcome the labeling problem shared by all unsupervised approaches. We present results of a comparison of our approach with Caraballo’s method, assessing the quality of the automatically learned ontologies by comparing them to a handcrafted taxonomy for the tourism domain using the similarity measures of Maedche et al. Further, we also present a human evaluation of the concept hierarchy produced by our guided algorithm.