With the explosive growth of textual data from social media (forums, blogs, and social networks), exploitation of these new information sources has become crucial. Our work focuses on sentiment analysis in this social media context. In order to identify sentiments in messages (e.g., tweets, SMS), original text-mining techniques have to be proposed. This paper presents a new method that integrates semantic and lexical knowledge for sentiment analysis. The proposed approach gives an important weight to "sentiment words" for a classification task. Our study compares two corpora (i.e. 88milSMS and DEFT'2015) in order to highlight the specific characteristics of SMS data in social media context.