Scaling up Discourse Quality Annotation for Political Science

The empirical quantification of the quality of a contribution to a political discussion is at the heart of deliberative theory, the subdiscipline of political science which investigates decision-making in deliberative democracy. Existing annotation on deliberative quality is time-consuming and carried out by experts, typically resulting in small datasets which also suffer from strong class imbalance. Scaling up such annotations with automatic tools is desirable, but very challenging. We take up this challenge and explore different strategies to improve the prediction of deliberative quality dimensions (justification, common good, interactivity, respect) in a standard dataset. Our results show that simple data augmentation techniques successfully alleviate data imbalance. Classifiers based on linguistic features (textual complexity and sentiment/polarity) and classifiers integrating argument quality annotations (from the argument mining community in NLP) were consistently outperformed by transformer-based models, with or without data augmentation.