Online review is the real evaluation of the purchased on the internet, and play a leading role in the purchase of future customers. At present, the analysis of online reviews based on emotion. The methods of analysis have word-level, sentence-level, document-level. The using of the most widely is word2vec, sen2vec, text2vec, doc2vec based on neural network, deep learning. In the traditional method, that effect of context contact and semantic understanding is not good, and the way of the neural network and deep learning is relatively high in hardware, and the time is relatively long. This article mainly uses the traditional methods and the combination of the deep learning method, first to deal with comments according to the word, word frequency, part of speech analysis to find the candidate key, using the index and cluster analysis to merge candidate key, find out the keywords. According to the feature words and index, calculate the feature weights, obtain the value of weight at the standard weight and the Norm weight. According to compare the method of weight, we think the standard weight is relevant short text. We are using feature word labeled and classifying. This paper starts with the review (documents), analysis word, from word to sentences, to review (text), and to the hotel, finally, aim to the hotel classification. This paper will lay a foundation for the intelligent recommendation of hotels.