Word Sense Identification Improves the Measurement of Short-Text Similarity

While a number of short-text similarity measures have recently been proposed, these methods have rarely focused on word sense identification. This paper present a variation, computationally efficient word sense identi-fication method that operates by comparing WordNet glosses of a target word with a con-text vector comprising the remaining words in the text fragment surrounding the target word. Empirical results show that the method performs favorably against baselines on a two-way sense-assigned dataset, and that its incorporation as a pre-processing step in a sentence similarity measure leads to superior performance on a well-known paraphrase detection dataset.