A Study on Word Similarity using Context Vector Models

Abstract

There is a need to measure word similarity when processing natural languages, especially when using generalization, classification, or example-based approaches. Usually, measures of similarity between two words are defined according to the distance between their semantic classes in a semantic taxonomy. The taxonomy approaches are more or less semantic-based that do not consider syntactic similarities. However, in real applications, both semantic and syntactic similarities are required and weighted differently. Word similarity based on context vectors is a mixture of syntactic and semantic similarities. In this paper, we propose using only syntactic related co-occurrences as context vectors and adopt information theoretic models to solve the problems of data sparseness and characteristic precision. The probabilistic distribution of co-occurrence context features is derived by parsing the contextual environment of each word, and all the context features are adjusted according to their IDF (inverse document frequency) values. The agglomerative clustering algorithm is applied to group similar words according to their similarity values. It turns out that words with similar syntactic categories and semantic classes are grouped together.