Abstract. Dependency Grammars prove to be effective in improving sentiment analysis, because they can directly capture syntactic relations between words. However, most dependency-based systems suffer from a major drawback: they only use 1-best dependency trees for feature extraction, which adversely affects the performance due to parsing errors. Therefore, we propose an approach that applies dependency forest to sentiment analysis. A dependency forest compactly represents multiple dependency trees. We develop new algorithms for extracting features from dependency forest. Experiments show that our forest-based system obtains 5.4 point absolute improvement in accuracy over a bag-of-words system, and 1.3 point improvement over a tree-based system on a widely used sentiment dataset. Our forest-based system also achieves state-ofthe-art performance on the sentiment dataset.