Eliciting Subjectivity and Polarity Judgements on Word Senses

There has been extensive work on eliciting human judgements on the sentiment of words and the resulting annotated word lists have frequently been used for opinion mining applications in Natural Language Processing (NLP). However, this word-based approach does not take different senses of a word into account, which might differ in whether and what kind of sentiment they evoke. In this paper, we therefore introduce a human annotation scheme for judging both the subjectivity and polarity of word senses. We show that the scheme is overall reliable, making this a well-defined task for automatic processing. We also discuss three issues that surfaced during annotation: the role of annotation bias, hierarchical annotation (or underspecification) and bias in the sense inventory used.