Relational Similarity Measurement Between Word-pairs using Multi-Task Lasso

Abstract—Relational similarity measurement as a popular research area in the field of natural language processing, is widely used in information retrieval, word sense disambiguation, machine translation and so on. The existing approaches are mostly based on extracting semantic features as feature matrixes from the large-scale corpus and using the corresponding method to process these feature matrixes to compute the relational similarity between word-pairs. However, the extracted semantic features are loosely distributed, which make the sparseness of feature matrixes. This paper proposes a Multi-Task Lasso based Relational similarity measure method (MTLRel), which makes snippets retrieved from a web search engine as the semantic information sources of a word-pair, then builds the feature matrix by extracting predefined patterns from snippets, compress and denoise the feature matrix into a feature vector using a multi-task lasso method, finally measures the relational similarity between two word-pairs by computing the cosine of the angle between two feature vectors. The MTLRel approach achieves an accuracy rate of 50.3 % by testing 374 SAT analogy questions with lower time consumption. Index Terms — Relational similarity, multi-task learning, Lasso I