Combination Features for Semantic Similarity Measure

Computing the semantic similarity between words is one of the key tasks in many language-based applications. Recent work has focused on using contextual clues for semantic similarity computation. In this paper, we propose a method to the measure semantic similarity between words using plain text contents. It takes into account information attributes (local) and topic information (global) of words to disclose their semantic similarity scores. The method models the representation of a word as a high dimensional vector of word attributes and latent topics. Thus, the semantic similarity between two words is measured by the semantic distance between their respective vectors. We have tested the proposed method on WordSimilarity-353 dataset. The empirical results have shown the combination features contribute to improve the semantic similarity results the dataset in comparison with previous work on the same task using plain text contents.