Even Unassociated Features Can Improve Lexical Distributional Similarity
Kazuhide Yamamoto, Takeshi Asakura · 2010
This paper presents a new computation of lexical distributional similarity, which is a corpus-based method for computing similarity of any two words. Although the conventional method focuses on emphasizing features with which a given word is associated, we propose that even unassociated features of two input words can further improve the performance in total. We also report in addition that more than 90 % of the features has no contribution and thus could be reduced in future.