In this paper, we address the issue of explainability in a transformer-based subjectivity regressor trained on native English speakers’ judgements. The main goal of this work is to test how the regressor’s predictions, and therefore native speakers’ intuitions, relate to theoretical accounts of subjectivity. We approach this goal using two methods: a top-down manual selection of theoretically defined subjectivity features and a bottom-up extraction of top subjective and objective features using the LIME explanation method. The explainability of the subjectivity regressor is evaluated on a British news dataset containing sentences taken from social media news posts and from articles on the websites of the same news outlets. Both methods provide converging evidence that theoretically defined subjectivity features, such as emoji, evaluative adjectives, exclamations, questions, intensifiers, and first person pronouns, are prominent predictors of subjectivity scores. Thus, our findings show that the predictions of the regressor, and therefore native speakers’ perceptions of subjectivity, align with subjectivity theory. However, an additional comparison of the effects of different subjectivity features in author text and the text of cited sources reveals that the distinction between author and source subjectivity might not be as salient for naïve speakers as it is in the theory.

The problem of subjectivity detection is often approached as a preparatory binary task for sentiment analysis, despite the fact that theoretically subjectivity is often defined as a matter of degree. In this work, we approach subjectivity analysis as a regression task and test the efficiency of a transformer RoBERTa model in annotating subjectivity of online news, including news from social media, based on a small subset of human-labeled training data. The results of experiments comparing our model to an existing rule-based subjectivity regressor and a state-of-the-art binary classifier reveal that: 1) our model highly correlates with the human subjectivity ratings and outperforms the widely used rule-based “pattern” subjectivity regressor (De Smedt and Daelemans, 2012); 2) our model performs well as a binary classifier and generalizes to the benchmark subjectivity dataset (Pang and Lee, 2004); 3) in contrast, state-of-the-art classifiers trained on the benchmark dataset show catastrophic performance on our human-labeled data. The results bring to light the issues of the gold standard subjectivity dataset, and the models trained on it, which seem to distinguish between the origin/style of the texts rather than subjectivity as perceived by human English speakers.