A NEW FEATURE SELECTION METHOD BASED ON CONCEPT EXTRACTION IN AUTOMATIC CHINESE TEXT CLASSIFICATION

Abstract

The feature selection is an important part in automatic text classification. In this paper, we use a Chinese semantic dictionary — Hownet to extract the concepts from the word as the feature set, because it can better reflect the meaning of the text. However, as the concept definition in the dictionary sometimes cannot express the word properly, we define the expression power for every sememe and every definition of the word in further process, and define the relation degree between the sememe and the definition. A threshold is set in the sememe tree, the sememe of the little information is filtered, and the words of weak definition are reserved in expression power. By this method, we construct a combined feature set that consists of both sememes and the Chinese words. The values of sememes are given according to their expression power and relation to the word. By comparing seven feature weighing methods in text classification, we propose a CHI-MCOR weighing method according to the weighing theories and classification precision. Experimental result shows that if the words are extracted properly, not only the feature dimension is smaller but also the classification precision is higher. Our method makes a good balance between the features which occur frequently in the corpus and those which only occur in one category, the difference of the classification precision among different categories is small.