Abstract: Most word embedding models are based on the theory of distribution hypothesis, which take a word as a basic unit and infer word representation from its external contexts. However, in some languages similar to Chinese, a word is built from several characters and these characters contains rich internal information. The semantic of a word is closely related to the semantic of its composing characters. Therefore, this paper take Chinese for example and present two model to collaborative learn word and character representation. In order to solve the phenomenon of homonymy and polysemy, multiple-prototype character embeddings and an word selection method are proposed. We evaluate the proposed models on similarity tasks and analogy tasks. The results demonstrates the proposed models outperform other baseline models.