We study the influence of lexical subsumption on the semantic proximity of terms. By studying the semantic relation between lexically subsumed terms in an existing controlled vocabulary and those in a corpus, we formulate hypotheses on the underlying specific relations. These hypotheses enable us to propose an ordering of term variants found in the corpus, by probability of decreasing semantic proximity. Results from our experiment show that surface indices like the presence of proper names, compounds and the number of added words in a term variant are criteria for ordering the variants of the same term. The ranking obtained will enable us to form more semantically-cohesive clusters. Applications targeted are text mining, question-answering and science and technology watch.