Word embeddings have been extensively used for sentiment analysis tasks. However, typical existing algorithms only model the syntactic context of words but fail to capture sufficient sentiment information of text, which affects the performance of sentiment analysis. Therefore, this paper presents an word vector refinement model based on improved genetic algorithm, which uses sentiment lexicon to obtain the sentiment ranking of the semantic nearest neighbors of the target word, and uses an improved genetic algorithm to optimize the vector representations of words such that they get the sentiment information of the word. Experimental results show that the proposed model can improve conventional word embeddings for binary classification on Internet Movie Database (IMDB) and fine-grained classification on Stanford Sentiment Treebank (SST).