A Refined TF-IDF Algorithm Based on Channel Distribution Information for Web News Feature Extraction

TF-IDF algorithm is widely used in text feature extraction, in which IDF value demonstrates the importance of a term. While applying to the procession of web news, the traditional IDF doesn´t work well, especially in a collection divided according to channels. In order to solve this problem, a refined IDF schema is proposed, named Channel Distribution Information (CDI) IDF, which is based on the information among the IDF values of each channel collections. According to the statistical features, the Top terms and the meaningless terms could be identified. Experiments on a manual labeled test set indicated that, related to the traditional TF-IDF, the CDI TF-IDF increases the Recall, Precise and F0.5 measure by 2.71%, 3.07% and 3.00%.