Information Dynamics of Language Communication

A computational framework that measures the directed flow of meaning in conversation has been introduced, offering a new way to quantify how semantic content moves between speakers. The work, posted to arXiv on 29 June 2026, uses large language models to track predictive influence in dialogue [1][2]. The framework computes two principal measures: semantic transfer entropy (STE), which captures the directed predictive influence one speaker exerts on another, and semantic partial information decomposition (SPID), which resolves how multiple sources jointly shape a target's language into redundant, unique, and synergistic components [2]. The approach treats large language models as probabilistic estimators of natural language, enabling the quantification of semantic flow without relying on hand-crafted linguistic features [1][2]. Across four experiments, the authors report that the framework detects reduced information flow in cognitively rigid dialogue, captures the dominant role of persuaders in shaping discourse, and distinguishes high- from low-quality psychotherapy by examining the directionality of therapist-client information exchange [2]. It also reveals synergistic premise contributions in argumentative essays, where multiple lines of reasoning combine to produce effects not present in any single premise alone [2]. The study addresses a gap in computational linguistics, where quantifying how meaning propagates through communicative exchanges has remained underdeveloped [2]. While interpersonal communication research has long examined how humans adjust verbal and nonverbal cues during face-to-face interaction, and how messages are produced and interpreted, the computational modeling of semantic influence has lagged behind [5]. The new framework provides a formal, information-theoretic lens on dynamics that scholars in group dynamics and communication studies have explored qualitatively for decades [4][5]. The work also operates in a domain where nonverbal channels — including tone, inflection, emphasis, and other vocal characteristics — convey significant meaning alongside verbal content [3]. The framework's reliance on text-based language models means it captures only the verbal dimension of exchange, leaving the integration of prosodic and paralinguistic signals as an open challenge [2][3]. The submission, totaling 3,597 KB, was posted by Leonardo Goodall and is hosted on arXiv under the Computation and Language category [1]. The authors suggest the framework opens avenues for studying information dynamics in digital discourse, pedagogical interactions, and clinical dialogues [2].