The new trend in sentiment classification is to use semantic features for representation of documents. We propose a semantic space based on WordNet senses for a supervised document-level sentiment classifier. Not only does this show a better performance for sentiment classification, it also opens opportunities for building a robust sentiment classifier. We examine the possibility of using similarity metrics defined on WordNet to address the problem of not finding a sense in the training corpus. Using three popular similarity metrics, we replace unknown synsets in the test set with a similar synset from the training set. An improvement of 6.2% is seen with respect to baseline using this approach.
展开▼
