Milling the Discriminative Word Sets for Bag-of-Words Model Based on Distributional Similarity Graph

摘要

Most of the previous distributional clustering methods are fundamentally unsupervised, and the discriminative property of words is not well modeled in the clustering procedure. In this paper, we propose a supervised model which involves the class conditional probability in measuring the word similarity, and transform the word-set extraction to a supervised graph-partition optimization model. A greedy algorithm is proposed to solve this model, which combines the word selecting method and the word grouping method in the unified framework. By grouping the related words, this method essentially transforms the exact match between word bins to fuzzy match between groups of related-word bins, which to some extent avoid the synonymous problems in BoW model. Experiments on data sets demonstrate that the proposed method is applicable for both text sets and image sets, and has advantages in producing better retrieval precision and meanwhile reducing the lexicon size.