Teaching WordNet to Sing like an Angel and Cry like a Baby: Learning Affective Stereotypical Behaviors from the Web

Just as words have the potential to mean different things in different contexts, so too can their affective intent vary from one context to another. Thus, in some contexts one might feel complimented to be described as cunning, but feel aggrieved and insulted to be so described in another. As concepts become more complex and multifaceted, and accrete more layers of stereotypical associations, their ability to assume different affective profiles in different contexts also increases. Complimentary uses of the stereotype baby, for instance, will emphasize the positive behaviors of babies, while insulting uses have many negative behaviors to draw upon and accentuate. In this paper we propose a two-level organization for the affective lexicon, one that can work well with different WordNets. At the first level, stereotype-denoting terms are associated with a rich and nuanced description of their potential behaviors; at the second level, these behaviors are mapped to their affect profiles. This current paper focuses primarily on the first level.