A corpus analysis of online news comments using the Appraisal framework

Abstract

We present detailed analyses of the distribution of Appraisal categories (Martin & White, 2005) in a corpus of online news comments. The corpus consists of just over one thousand comments posted in response to a variety of opinion pieces on the website of the Canadian English-language newspaper The Globe and Mail. We annotated all the comments with labels corresponding to different categories of the Appraisal framework. Analyses of the annotations show that comments are overwhelmingly negative and that they favour two of the subtypes of Attitude, Judgement and Appreciation. The paper contributes a methodology for annotating Appraisal, examining the interaction of Appraisal with negation, the constructive nature of comments, and the level of toxicity found in them. The results show that highly opinionated language is expressed as opinion (Judgement and Appreciation) rather than as an emotional reaction (Affect). This finding, together with the interplay of evaluative language with constructiveness and toxicity in the comments, can be applied to the automatic moderation of online comments.