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

Abstract

Human communication has been described as a contextual social inference process. Research into great ape communication has been inspired by this view to look for the evolutionary roots of the social, cognitive and interactional processes involved in human communication. This approach has been highly productive, yet it is partly compromised by the widespread focus on how great apes use and understand individual signals. This paper introduces a computational model that formalizes great ape communication as a multi-faceted social inference process that integrates (a) information contained in the signals that make up an utterance, (b) the relationship between communicative partners and (c) the social context. This model makes accurate qualitative and quantitative predictions about real-world communicative interactions between semi-wild-living chimpanzees. When enriched with a pragmatic reasoning process, the model explains repeatedly reported differences between humans and great apes in the interpretation of ambiguous signals (e.g. pointing or iconic gestures). This approach has direct implications for observational and experimental studies of great ape communication and provides a new tool for theorizing about the evolution of uniquely human communication. This article is part of the theme issue 'Revisiting the human 'interaction engine': comparative approaches to social action coordination'.

Keywords: communication; computational modelling; evolution; primates; social cognition.

Figure 1.

Schematic overview of the computational model. The sender (right) is producing an utterance and the receiver (left) tries to infer the intention of the sender based on the information sources available. The model takes in information provided by the utterance (gesture and facial expression) and the interactional history (immediate social context and dominance relation).

Figure 2.

Model predictions compared to data from [50]. (a) The mean proportion (bars) of affiliative and avoidant reactions for combinations of gesture, facial expression, relationship and social context in the data. Only combinations with more than five observations are shown. Error bars are 95% confidence intervals based on a non-parametric bootstrap. Red crosses show model predictions. (b) Correlations between model prediction and data for avoidant reactions. The size of each point is proportional to the number of observations for a particular combination in the data. (c) Correlations for reduced models that focus only on a single component (with all other parameters set to 0.5). (Online version in colour.)

Figure 3.

Schematic depiction of the added pragmatic reasoning component. The literal receiver (a) only reasons about the gesture whereas the pragmatic receiver (b) reasons about why the sender produced that particular gesture. The pragmatic receiver further expects the sender to produce the gesture with the goal of being informative.

Figure 4.

Application of the pragmatically enriched model to an object-choice task with pointing gestures. (a) The context with the two locations (L = left and R = right) that can be referred to. Panel (b) gives the interpretation probabilities of a literal receiver. (c) The production probabilities for the pragmatic sender for values of α = 1, 5 and 10. (d) The interpretation probabilities of the pragmatic sender based on the production probabilities in (c). Coloured bars visualize the probabilities in reference to chance (grey dashed line). Different shades in (c,d) correspond to the magnitude of α. (Online version in colour.)