Extracting contextual evaluativity
Kevin Scott Reschke, Pranav Anand · 2011
Recent work on evaluativity or sentiment in the language sciences has focused on the contributions that lexical items provide. In this paper, we discuss contextual evaluativity, stance that is inferred from lexical meaning and pragmatic environments. Focusing on assessor-grounding claims like We liked him because he so clearly disliked Margaret Thatcher, we build a corpus and construct a system employing compositional principles of evaluativity calculation to derive that we dislikes Margaret Thatcher. The resulting system has an F-score of 0.90 on our dataset, outperforming reasonable baselines, and indicating the viability of inferencing in the evaluative domain. 1 Contextual Evaluativity A central aim of contemporary research on sentiment or evaluative language is the extraction of evaluative triples: 〈evaluator, target, evaluation〉. To date, both formal (e.g., Martin and White 2005, Potts 2005) and computational approaches (e.g., Pang and Lee 2008) have focused on how such triples are lexically encoded (e.g., the negative affect of scoundrel or dislike). While lexical properties are a key source of evaluative information, word-based considerations alone can miss pragmatic inferences resulting from context. (1), for example, communicates that the referent of we bears not only positive stance