Applying Latent Semantic Analysis to Optimize Second-order Co-occurrence Vectors for Semantic Relatedness Measurement

摘要

Measures of semantic relatedness are largely applicable in intelligent tasks of NLP and Bioinformatics. By taking these automated measures into account, this paper attempts to improve Second-order Co-occurrence Vector semantic relatedness measure for more effective estimation of relatedness between two given concepts. Typically, this measure, after constructing concepts definitions (Glosses) from a thesaurus, considers the cosine of the angle between the concepts' gloss vectors as the degree of relatedness. Nonetheless, these computed gloss vectors of concepts are impure and rather large in size which would hinder the expected performance of the measure. By employing latent semantic analysis (LSA), we try to conduct some level of insignificant feature elimination to generate economic gloss vectors. Applying both approaches to the biomedical domain, using MEDLINE as corpus, UMLS as thesaurus, and reference standard of biomedical concept pairs manually rated for relatedness, we show LSA implementation enforces positive impact in terms of performance and efficiency.