Exploring Content Predictability in Turn-taking Through Different Computer-Mediated Communications
Published in 31st International Conference on Computational Linguistics (COLING), 2025
Published: COLING 2025, Abu Dhabi, UAE (pp. 7949–7962).
This paper introduces an LLM-based approximation of human next-word prediction as a scalable, behavior-grounded measure of content predictability during turn-taking, replacing more invasive neuroscience-based measurements. We compare predictability profiles across multiple computer-mediated communication (CMC) modalities to quantify how the loss of multimodal cues (gaze, prosody, gesture) reshapes how speakers and listeners anticipate upcoming content.
Results reveal systematic differences in predictability across CMC modalities, with text-only and voice-only conversations showing different patterns of prediction breakdown compared to face-to-face baselines. The paper provides a methodological contribution—reusable LLM-based prediction metrics for psycholinguistic CMC research—and an empirical foundation for designing AI scaffolds that target the specific prediction failures induced by each communication channel.
Recommended citation: He, W., MacDonald, C. C., Yoo, Y., Eizayaga, M., Shim, R., Katreczko, L. D., & Fussell, S. R. (2025). "Exploring Content Predictability in Turn-taking Through Different Computer-Mediated Communications." Proceedings of the 31st International Conference on Computational Linguistics (COLING 2025), pages 7949–7962. Abu Dhabi, UAE. Association for Computational Linguistics.
Download Paper