This paper presents a robust approach for sentiment detection on Twitter messages, leveraging biased and noisy data from sentiment detection websites. The authors propose a two-step classification method: first, distinguishing subjective from objective tweets, and second, classifying subjective tweets as positive or negative. They use meta-information about words and specific tweet characteristics as features, which capture a more abstract representation of tweets. The approach is evaluated through experiments, showing improved performance over previous methods, especially in handling noisy and biased data. The paper also discusses the impact of different strategies for combining data sources and the effectiveness of their proposed features. The results demonstrate that their method is effective even with limited training data and robust to noisy and biased labels.This paper presents a robust approach for sentiment detection on Twitter messages, leveraging biased and noisy data from sentiment detection websites. The authors propose a two-step classification method: first, distinguishing subjective from objective tweets, and second, classifying subjective tweets as positive or negative. They use meta-information about words and specific tweet characteristics as features, which capture a more abstract representation of tweets. The approach is evaluated through experiments, showing improved performance over previous methods, especially in handling noisy and biased data. The paper also discusses the impact of different strategies for combining data sources and the effectiveness of their proposed features. The results demonstrate that their method is effective even with limited training data and robust to noisy and biased labels.