Evaluation of Index Term Discovery in Medical Reference Text

Currently, and in the foreseeable future, the amount of text in electronic form is multiplying. While retrieval of this text is facilitated by increased indexing, there is a danger of overwhelming the user with too much information. As well, the current rise in available computing power means that we can consider computationally intense solutions. One such solution is dynamic taxonomy. Dynamic taxonomy uses an ontology based index, giving it the ontological advantages of structure, and increased number of interconnections. Subsequent complexity is managed by dynamic pruning. Dynamic taxonomy needs a text base that has been deeply indexed with concepts from a taxonomy. In this paper, we evaluate various algorithms for generating such an index. For the text, we use a subset of an electronic medical reference database, and the indexing terms are extracted from an established medical ontology, the Unified Medical Language System (UMLS). This paper compares an index created by various algorithms against one devised by medical domain experts. The algorithms extract candidate phrases from the text using techniques such as NLP based phrase extraction, combinations of N consecutive words, and combinations of preexisting book