As a fundamental problem of text categorization, text representation is widely concerned. Currently, there are three main ways of text representation: bag-of-words model, latent semantic representation and knowledge-based explicit semantic representation. The paper analyzes and compared the effects of these methods applied to text categorization. Experiments show that the knowledge-based explicit semantic representation cannot improve the text categorization performance as expected. To tackle the problem that the knowledge-based explicit semantic representation easily introduces noise in extending text, a supervised explicit semantic representation method is proposed. The dataset label information is used to identify the most relevant concepts in document and the document is represented in explicit semantic based on expanding those key concepts. The results of three datasets confirm the effectiveness of the proposed method.