With the popularity of microblog and other social networks, a steady stream of new words emerge, Chinese word segmentation systems often cut the new words into Chinese characters. The new word discovery has become a hot topic in the field of Chinese natural language processing. Existing new word recognition methods rely on the statistical data of large-scale corpus, the ability of new low-frequency word recognition is poor. This paper presents an extension of skip-gram model and word vector projection method, after the combination of the this two methods can ease the data sparseness problem effectively in natural language processing, to identify new low-frequency words, and to improve the precision and recall rate of Chinese word segmentation system.