Semantic Similarity of Words in a Taxonomy: A Fuzzy and Multiple-Context Approach

In this paper, the semantic similarity of words organized according to a taxonomy is investigated. For this purpose, word senses are fundamental, and word senses depend on contexts. In general, in the literature, the methods for disambiguating the meaning of words are costly, have some limitations, often require human intervention, and sometimes identifying word senses is an ambiguous task also for human beings. Due to the frequent cases in which word disambiguation is inherently subjective, in this paper, we argue for the importance of considering all possible contexts and relative senses of words in these contexts. In our proposal, contexts and word senses are extracted from BabelNet, and words are fuzzy sets compared according to the Jaccard’s fuzzy similarity measure. This proposal has been applied to some of the most representative state-of-the-art taxonomy-based similarity measures and experimented with the Miller and Charles ( $M\& C$ ) golden dataset. Such a dataset is particularly relevant to test our proposal in that it stems from an in-depth analysis of the meaning of a word and the contexts in which the word is used. The experiment is very encouraging because it shows an improvement of the Pearson’s correlation values with human judgment for all the selected similarity measures.