Unsupervised Learning of the Morpho-Semantic Relationship in MEDLINE

Morphological analysis as applied to English has generally involved the study of rules for inflections and derivations. Recent work has attempted to derive such rules from automatic analysis of corpora. Here we study similar issues, but in the context of the biological literature. We introduce a new approach which allows us to assign probabilities of the semantic relatedness of pairs of tokens that occur in text in consequence of their relatedness as character strings. Our analysis is based on over 84 million sentences that compose the MEDLINE database and over 2.3 million token types that occur in MEDLINE and enables us to identify over 36 million token type pairs which have assigned probabilities of semantic relatedness of at least 0.7 based on their similarity as strings.