LIMSI: Learning Semantic Similarity by Selecting Random Word Subsets