Maximum entropy and conditional random field or other algorithms used for collocation extraction in the traditional assessment of Chinese language rely on manual selection of characteristics and have a high demand for semantics marking precision at the preliminary stage. In this paper，an alternative approach is suggested which substitutes term vector for the traditional semantic characteristics in collocation extracting. Specifically，the term vectors are acquired by an in-depth model completing unsupervised learning from a large corpus. In testing，the term vectors and the semantic characteristics are separately entered as inputs into three machine learning models. The results indicate that better outcomes are produced when term vectors are used in the neural network model in the sense that both the precision and recall rate are higher by nearly 3% than the best outcomes that are achievable with semantic characteristics. We also note that as the size of the corpus used for unsupervised learning training increases the resulting term vectors become more and more pragmatic.