A Context-based Model of Semantic Similarity

Lexical co-occurrence counts from large corpora have been used to construct highdimensional vector-space models of language.. Distances between word vectors extracted from these models are generally considered to reflect semantic similarity. Implicit in this assumption is that ‘semantic distance’ measurements correspond to human intuitions. This paper investigates the validity of one such measure, contextual similarity, calculated from the spoken part of the British National Corpus. In Experiment 1, a moderate correlation is found between human judgements of the semantic similarity between pairs of nouns and the model’s measure of contextual similarity. The correlation between the two measures is confirmed in two additional experiments, using a new set of elicited ratings. The semantic similarity of same-category word pairs (Experiment 2A) and the similarity between words differing in syntactic category (Experiment 2B) is found to be predictable from contextual similarity. The results from the three experiments provide support for the role of lexical co-occurrence information in modelling semantic similarity.