Feature clustering is a feature reduction method that reduces the dimensionality of feature vectors for text classification. In this paper an incremental feature clustering approach is proposed that uses Semantic similarity to cluster the features. Pointwise Mutual Information (PMI) is widely used word similarity measure, which finds Semantic similarity between two words and is an alternative for distributional similarity. PMI computation requires simple statistics about two words for similarity measure, that is number of cooccurrences or correlations between two concepts of fixed size are computed. Once the words from preprocessed documents are fed, clusters are formed and one feature (head word) is identified for each cluster which are used for indexing the document. PMI assumes that a word have single sense, but clustering can be optimized further if polysemies of words are considered. Hence PMI may be combined with PMImax, which estimates correlation between the closest senses of two words also, thereby better feature reduction and execution time compared with other approaches.
展开▼
