Classification of Semantic Relations in Different Syntactic Structures in Medical Text using the MeSH Hierarchy

Two different classification algorithms are evaluated in recognizing semantic relationships of different syntatic compounds. The compounds, which include nounnoun, adjective-noun, noun-adjective, noun-verb, and verb-noun, were extracted from a set of doctors' notes using a part of speech tagger and a parser. Each compound was labeled with a semantic relationship, and each word in the compound was mapped to its corresponding entry in the MeSH hierarchy. MeSH includes only medical terminology so it was extended to include everyday, non-medical terms. The two classification algorithms, neural networks and a classification tree, were trained and tested on the data set for each type of syntactic compound. Models representing different levels of MeSH were generated and fed into the neural networks. Both algorithms performed better than random guessing, and the classification tree performed better than the neural networks in predicting the semantic relationship between phrases from their syntactic structure. Thesis Supervisor: Peter Szolovits Title: Director, Clinical Decision-Making Group