Natural language processing for non-violent conflict detection in institutional discourse

This paper proposes a methodological pipeline for studying non-violent institutional conflict in formal deliberative settings, where disagreement is rarely expressed through overt hostility and instead emerges through tone, interactional positioning, and procedural dynamics. The approach combines computational methods with human-in-the-loop annotation to enable scalable yet interpretable analysis of conflict in public administration. The pipeline is illustrated through a case study of the Regional Council of Aosta Valley (Italy), using official multilingual parliamentary transcripts. We construct an original dataset integrating institutional metadata, speaker characteristics, and affective evaluations. A representative three-month subset is annotated along two dimensions, i.e., intrinsic tone and contextual conflict, by human coders, large language models (LLMs), and a multilingual BERT sentiment classifier, allowing systematic comparison across annotation sources. Results show that human annotations constitute the most coherent benchmark, while LLM-based labels are noisier, especially for contextual conflict. BERT-derived sentiment provides a stable but only partially aligned signal. Descriptive analyses reveal pronounced asymmetries in participation and evaluation, including exploratory gender-disaggregated patterns that motivate further investigation. Predictive models trained on non-linguistic metadata achieve moderate accuracy when supervised with automated sentiment, but near-perfect performance when human-coded tone is included, highlighting both the limits of automation and the interpretive nature of institutional conflict. Overall, the study demonstrates that non-violent conflict is empirically tractable yet deeply context-dependent, and offers a transferable framework for conflict analysis in democratic institutions.