From partners to populations: A hierarchical Bayesian account of coordination and convention.

Abstract

Languages are powerful solutions to coordination problems: They provide stable, shared expectations about how the words we say correspond to the beliefs and intentions in our heads. Yet, language use in a variable and nonstationary social environment requires linguistic representations to be flexible: Old words acquire new ad hoc or partner-specific meanings on the fly. In this article, we introduce continual hierarchical adaptation through inference (CHAI), a hierarchical Bayesian theory of coordination and convention formation that aims to reconcile the long-standing tension between these two basic observations. We argue that the central computational problem of communication is not simply transmission, as in classical formulations, but continual learning and adaptation over multiple timescales. Partner-specific common ground quickly emerges from social inferences within dyadic interactions, while community-wide social conventions are stable priors that have been abstracted away from interactions with multiple partners. We present new empirical data alongside simulations showing how our model provides a computational foundation for several phenomena that have posed a challenge for previous accounts: (a) the convergence to more efficient referring expressions across repeated interaction with the same partner, (b) the gradual transfer of partner-specific common ground to strangers, and (c) the influence of communicative context on which conventions eventually form.

Communicative Efficiency or Iconic Learning: Do acquisition and communicative pressures interact to shape colour-naming systems?

“…In order for a language to evolve, there needs to be a bias towards informativity and a loss of information (e.g., from a pressure to simplify). Empirical work in experimental semiotics has demonstrated that dyadic coordination provides a bias for informativity [12,19,20] and the shuffling of conversational partners affords information loss [21]; thus, satisfying the criteria for evolution without appealing to acquisition. At the same time, it is generally agreed that language acquisition has a bias for simplicity [11,66,67] resulting in information loss.…”

“…Several researchers have argued that that speaker-listener interaction is vital for informative signs to emerge [12,19,20]. Information is lost through shifting conversation partners and generational turn-over, which allows alignment on a generalisable language [21]. Thus, language evolution can occur through dyadic coordination alone [for review of the empirical literature see 6].…”

Great ape communication as contextual social inference: a computational modelling perspective

“…Recently, Hawkins et al . [ 99 ] embedded an RSA model of pragmatic in-the-moment inferences in a model of convention formation and showed how signals with vague meanings can give rise to conventional communication systems. The meaning of a signal can get fixed (e.g.…”

Mixed-effects transformers for hierarchical adaptation

“…However, if a new user is introduced from the same population, then the central tendency of the random effect is the best initial guess for their intercept. This hierarchical, mixed-effects model (also known as a partial-pooling model) may be understood as a way of interpolating between two extremes which are each prevalent in machine learning (Hawkins et al, 2022). The first extreme, known as a complete-pooling model, ignores group-specific effects and learns a single monolithic model across the pooled dataset.…”