POLITICAL-ADS: An annotated corpus for modeling event-level evaluativity

This paper presents a corpus targeting evaluative meaning as it pertains to descriptions of events. The corpus, POLITICAL-ADS is drawn from 141 television ads from the 2008 U.S. presidential race and contains 3945 NPs and 1549 VPs annotated for scalar sentiment from three different perspectives: the narrator, the annotator, and general society. We show that annotators can distinguish these perspectives reliably and that correlation between the annotator’s own perspective and that of a generic individual is higher than those with the narrator. Finally, as a sample application, we demonstrate that a simple compositional model built off of lexical resources outperforms a lexical baseline.