Comparing measures of semantic similarity

The aim of this paper is to compare different methods for automatic extraction of semantic similarity measures from corpora. The semantic similarity measure is proven to be very useful for many tasks in natural language processing like information retrieval, information extraction, machine translation etc. Additionally, one of the main problems in natural language processing is data sparseness since no language sample is large enough to seize all possible language combinations. In our research we experiment with four different measures of association with context and eight different measures of vector similarity. The results show that the Jensen-Shannon divergence and L1 and L2 norm outperform other measures of vector similarity regardless of the measure of association with context used. Maximum likelihood estimate and t-test show better results than other measures of association with context.