Bridging Theory and Machine: A Framework for Theory-Grounded Deep Learning in Political Text Analysis

McCarthy & Dore (2023) highlight a persistent gap: most computational text analyses in political science are either descriptively rich but atheoretical, or theoretically motivated but computationally simplistic. This research aims to create a framework where political theory (e.g., agenda-setting, framing, polarization) is embedded as an explicit constraint or regularization in transformer-based models. For instance, theory-derived codes or constructs could guide attention layers or be used as auxiliary prediction tasks (see Yeo et al., 2024). This would generate outputs that are both powerful and interpretable, helping resolve the "black box" problem of deep learning in political text analysis. The significance lies in advancing both the methodological rigor and substantive insight of computational political science, moving beyond mere pattern recognition to theory-driven discovery.

References:

1. Theory-Grounded Computational Text Analysis. Arya D. McCarthy, G. Dore (2023). Annual Meeting of the Association for Computational Linguistics.
2. Beyond Text: Leveraging Multi-Task Learning and Cognitive Appraisal Theory for Post-Purchase Intention Analysis. G. Yeo, Shaz Furniturewala, Kokil Jaidka (2024). Annual Meeting of the Association for Computational Linguistics.