A method of text mining is disclosed for automatically building text knowledge base. First, the text mining is applied to the web pages downloaded from internet/intranet or other text documents to extract phrases information. Then, the phrases are classified using automatic classification method or using existed classification information. In addition, the weights between the phrases are trained by using the text information in the web pages or the documents. A knowledge base system is built using the text mining results. The knowledge base is used to directly provide knowledge for a search. Also, the knowledge base helps search engine refine search results.