A LABEL-ORIENTED APPROACH FOR TEXT CLASSIFICATION

摘要

Text classification is a well known problem in the machine learning community. A widely used approach is that based on the Term Frequency - Inverse Document Frequency (TF-IDF) feature. This feature represents very well the characteristic of a text. However, this feature could not clearly represent the relationship from a text to its assigned label. This paper presents a Label-Oriented (LO) approach for text classification problem. This approach takes account of the relationship between a text and its assigned label by introducing a new feature label-oriented score. This score represents the level of the importance of the term regarding all terms and texts assigned to the label compared to all terms and texts unassigned to the label. In the training phase, this model calculates the label-oriented score of each term to a label. In the testing phase, the sum of this score of all terms in a text will help us to determine whether the text should be assigned to the label or not. The proposed model is then evaluated in two cases: short and regular texts. The experiment results indicate that the proposed model is significantly better than baseline models on the considered datasets.