Measuring Taxonomic Similarity between Words Using Restrictive Context Matrices

Measuring taxonomic similarity between words plays an important role in many semantic-based applications but still remains a challenging task today. We propose a new method which utilizes restrictive context matrices for this problem. We learn a set of special lexico-syntactic patterns automatically and use them to extract taxonomic related contexts of words from raw text. These restrictive contexts are then transformed into real matrices and similarities between them are calculated to reflect the taxonomic similarities between words. The main contribution of our work is that taxonomic related context of words can be mined, evaluated, and used to measure taxonomic similarities between words. Experimental results on Miller-Charles benchmark dataset achieve a correlation coefficient of 0.856.