Building a fine-grained subjectivity lexicon from a web corpus