Combining Rule-Based Methods and Latent Semantic Analysis for Ontology Structure Construction

We report on a pilot study of techniques for organizing terms derived from relevant corpora into semantic clusters. We combine rule-based techniques and an unsupervised statistical method to semantically cluster terms extracted from medical reports. The rule-based approach identifies lexicosyntactic patterns suggestive of parent-child relations and coordination structures suggestive of sibling-sibling relations. Latent semantic analysis (LSA) is applied to those relations to weed spurious pairs. Two domain experts evaluated the results. Applying LSA as a filter to the relations extracted by rules did not improve the precision for either parent-child relations (52%) or sibling-sibling relations (71%).