Bootstrapping information extraction mappings by similarity-based reuse of taxonomies

Many practical information extraction systems use simple taxonomies for mapping extracted strings to client-specific concept codes. In such taxonomies, concepts are defined as groups of semantically similar words and phrases. For the mapping to be accurate, a new client-specific taxonomy – usually nothing more than a set of concept codes, each with a single description – needs to be enriched with the domain-specific terminology variations, which is a very labor-intensive task. In this paper, we describe a method to significantly reduce the required manual effort for this task. Our approach is based on combining multiple existing client-specific taxonomies into a single semantic space. On a set of gold standard taxonomies our method achieves an average precision of 91% and a recall of 55%. An additional practice test shows that the method saves at least 62% of the manual effort needed to enrich a new taxonomy.