Sentiment analysis of social media text is a very interesting and important study, which has a wide range of applications both in academia and industry. However, most researchers focus only on the study of semantic features without taking the effect of sentiment information into consideration. Motivated by the needs to sentiment information for sentiment analysis, we propose an Emotion Feature Highlighting Method (EFHM), which is able to utilize both the semantic information and emotion information. In the first step, emotion punctuation and word are combined as a new word in our method to enrich the emotion information of the text in the preprocessing stage. In the second step, the emotional words are extended by calculating the emotional relevance on the basis of the emotion dictionary. In the last step, we extract the feature representation of a text using an improved Continuous Bag-of-Words (CBOW) model, in which relation-specific vector offset is updated according to the emotional weight. Our experiments on three social media datasets show the superior performance for sentiment analysis tasks both in Chinese and English text.