Task-Conditioned Multi-Task Learning for Actor-Centric Rhetorical Positioning in Political Discourse

Automatically modelling how political actors are evaluated in discourse requires methods that move beyond general sentiment classification. Existing NLP techniques typically capture overall polarity but struggle to attribute evaluative language to the specific social entities referenced within a text. This paper introduces actor-centric rhetorical positioning, a token-level framework that links references to social actors with Positive, Neutral, or Negative evaluative labels. To model the dependency between entity identification and evaluation, we propose a Task-Conditioned Multi-Task Learning (TC-MTL) architecture in which rhetorical positioning is conditioned on entity detection. We compare this approach against Single-Task Learning (STL) and parallel multi-task baselines using DistilBERT, RoBERTa, and DeBERTa on a corpus of U.S. political discourse spanning rally speeches, formal addresses, and online political commentary. TC-MTL improves NER performance by approximately 2–4 F1 points over STL baselines. Explicit conditioning strengthens entity extraction while maintaining positioning performance, resulting in a more reliable framework for analysing actor-centred political discourse. Our approach enables the systematic detection of how actors are constructed and positioned in political speech, supporting large-scale analysis of rhetorical framing and identity-based polarisation.