Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Shared meaning in dialogue emerges through sustained semantic reorganization
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Shahrzad bahmani1, Ivan Toni2, Amir Homayun Hallajian3; 1Donders Institute for Brain, Cognition, and Behaviour, Radboud University
Human communication is often modeled as inference over a shared space of possible meanings, where a signal reduces uncertainty within a set of referential options already available to both interlocutors. Yet many interactions do not operate inside a pre-defined space. They progressively reshape it. What becomes contrastive, salient, or worth saying next depends on the history of the interaction itself. Here, we assume that communication is a coupled dynamical process in which interlocutors sculpt a shared conceptual structure and keep it stable enough to support meaning. We ask how communicating dyads achieve semantic alignment by actively supporting this conceptual reorganization. We analyzed dialogues in 43 dyads from the CABB corpus (Eijk et al. 2022), recorded during a face-to-face communication task where participants develop referential expressions to indicate novel objects through 96 alternating director-matcher trials. To quantify dialogue-related change in semantic alignment within a dyad, each participant independently named the referents before and after the dialogue. To quantify the conceptual dynamics of each dialogue, conversational turns were embedded contextually with text-embedding-3-large, generating a weighted turn-by-turn semantic similarity graph from contextual cosine similarities, sparsified to retain only statistically informative connections (Serrano et al. 2009). The graph structure was tracked through its intrinsic dimensionality (ID), the effective number of independent semantic directions occupied by the unfolding dialogue, across its trajectory (Jazayeri and Ostojic 2021). Real dyads were compared against Erdős–Rényi (ER) and degree-preserving Configuration Model (CM) nulls. Real dyads compressed their semantic space far below both nulls (p<.001, d>4.7) with a fraction of the turn-by-turn ID volatility of the nulls (p<.001, d>17). This finding indicates that dialogue canalized the space of possible meanings considered by each dyad. Every dyad reached peak ID volatility within the first third of normalized session time (p<.001), followed by a more stable regime. This finding indicates that canalization of the semantic space unfolded early during the dialogue. Critically, dyad-specific semantic alignment was associated with the degree of conceptual reorganization that occurred in each dialogue after the early canalization event, as indexed by mid-session ID volatility, i.e. the semantic change occurring in the second third of normalized session time (r=.37, p≤.014). This finding indicates that communicatively successful dyads kept the shared space of possible meanings plastic enough to be refined, and stable enough to be mutually accessible. This sustained conceptual reorganization left a spectral signature on the semantic similarity graphs (Ghavasieh and De Domenico 2024), with real dyads retaining differentiated semantic organization and sustained entropy for longer than ER networks (p=.001, d=0.50; p=.005, d=0.41). Real dyads also carried higher entropy than ER networks in 40 of 43 dyads (p<.001, d=1.56), indicating that the relational organization of meaning was not random connectivity. These findings indicate that, during face-to-face referential interactions that evolve over time, interlocutors jointly build a conceptual structure that supports semantic alignment. Successful communication requires sustained plastic reorganization of this structure, balanced with a corresponding entropic price for stabilizing it.
Topic Areas: Meaning: Discourse and Pragmatics, Computational Approaches