The social transmission of linguistic alignment: The role of observational reinforcement learning

Abstract

People align their linguistic use in conversations, and it has been proposed that such linguistic alignment facilitates mutual understanding and social rapport. However, the mechanisms through which perspective alignment is socially transmitted across individuals remain unclear. Here, we combined behavioural experimentation and computational learning models to investigate whether, and if so, how observing aligned and unaligned responses in others affects people's perspective alignment in subsequent language use. Participants completed a three‐session design. First, in a baseline session, they played a picture matching and naming game with a child interlocutor, in which they took turns with the child participant (in fact, just recordings) to name and match pictures. Second, in an observational learning session, on each trial, they first observed the child and an adult (a demonstrator) describing pictures from an appearance perspective or an occupation perspective, with the adult often or rarely using the same perspective as the child; in particular, participants were asked to explicitly predict the expression of the adult. Afterwards, they played a round with the child themselves. Finally, in the generalization session, they played the same picture matching and naming game as in the baseline session, but with a new child. Observing high alignment increased observers' own alignment, whereas observing low alignment decreased it, and these changes carried over to the new child, indicating a communicative strategy that generalizes across interlocutors of the same category rather than a partner‐specific adjustment. To identify the underlying mechanism, we modelled behaviour trial by trial. A Rescorla–Wagner model first confirmed that participants learned to predict the demonstrator's choices. We then used Bayesian model comparison to adjudicate three accounts of how alignment changed. A model driven by observational prediction errors (i.e. the discrepancy between participants' predictions and the demonstrator's actual choices) outperformed both an imitation account (copying the most recent choice) and a priming account (the cumulative frequency of the demonstrator's aligned responses). In the high alignment group, both positive and negative prediction errors in the second half of the observational learning session predicted alignment changes, whereas in the low alignment group, negative prediction errors predicted decreased alignment. Exploratory analyses further suggested that alignment may function as a social signal: observers rated high‐alignment demonstrators as more perspective‐taking than low‐alignment demonstrators. Together, these findings show that perspective alignment can be socially transmitted through prediction‐error‐driven observational learning, neither passive copying nor mere exposure, and offer a computational account of how communicative strategies spread across individuals.