As an unsupervised machine learning method, clustering can preliminarily group text without artificial labeling, which effectively accelerates the organization, abstraction and navigation on large news set. The length of news is long, and the text contains many homonymy and polysemy, that is one of the reason that traditional text clustering methods perform weaker on grouping news text. This paper presents a novel text representation method based on topical document embedding (TDE) to capture the semantic features of different topics. In TDE representation, document embedding of news texts is obtained by adding up word vector from Skip-Gram model weighted by TF-IDF score of all the key words in the text. While the topical document embedding is learned by joining the topic vectors obtained from LDA model and the document vectors in document embedding. By using topical document embedding to perform clustering, we implement a novel text clustering method (TDE-TC). The experimental results show that the effect of news clustering based on TDE representation is better than that of bag of words model and LDA model.