Regional prejudice is prevalent in Chinese cities in which native residents and migrants lack a basic level of trust in the other group. Like Twitter, Sina Weibo is a social media platform where people actively engage in discussions on various social issues. Thus, it provides a good data source for measuring individuals’ regional prejudice on a large scale. We find that a resentful tone dominates in Weibo messages related to migrants. In this paper, we propose a novel approach, named DKV, for recognizing polarity and direction of sentiment for Weibo messages using distributed real-valued vector representation of keywords learned from neural networks. Such a representation can project rich context information (or embedding) into the vector space, and subsequently be used to infer similarity measures among words, sentences, and even documents. We provide a comprehensive performance evaluation to demonstrate that by exploiting the keyword embeddings, DKV paired with support vector machines can effectively recognize a Weibo message into the predefined sentiment and its direction. Results demonstrate that our method can achieve the best performances compared to other approaches.