Quantifying Conversational Coherence: Empirical Evidence of Shared Informational Structure in Human Dialogue

This dataset contains the full materials and computational implementation used in the study “Quantifying Conversational Coherence: Empirical Evidence of Shared Informational Structure in Human Dialogue.” It includes: Human annotations of six natural dialogues (≈23 turns each) by five independent annotators, marking valleys (ruptures of coherence) and peaks (recoveries of coherence). The Python script, implementing the tolerant evaluation framework described in the paper (tolerant Precision–Recall–F₁, permutation baselines, Fleiss’ κ and Light’s κ on dilated labels). The reproducible results (valleys_peaks_final_results.xlsx) obtained from the final dataset. Analyses were performed using pooled and window-tolerant matching (±1, ±2, ±3 turns) and 300 permutation baselines preserving per-annotator event frequencies. All computations are fully replicable. The materials support the claim—consistent with the Theory of Informational Emergence (TIE)—that conversational coherence is not merely semantic but an emergent informational property of interactional dynamics. Future work: A forthcoming pilot study will correlate these human agreement results with automated coherence metrics generated by TIE–Dialog, an analytical tool designed to simulate informational coherence and detect structural ruptures and recoveries in real-time conversational data. Contact: Adolfo J. Céspedes Jiménez — University of Groningen 📧 [e-mail]