Learning to communicate about conceptual hierarchies

Natural languages provide speakers with remarkable flexibility in the labels they may use to refer to things (Brown, 1958). In addition to the combinatorial explosion of modifiers afforded by compositionality (Partee, 1995), we have a number of lexicalized nominal terms at our disposal. Dalmatian, dog, and animal can all truthfully be used to talk about the same Dalmatian at different levels of specificity, with one level of the conceptual hierarchy – the basic-level – generally privileged over the others (Rosch, Mervis, Gray, Johnson, & Boyes-Braem, 1976). How these overlapping meanings are learned, and why speakers choose different levels of specificity in different contexts, is increasingly accounted for by probabilistic models of pragmatic language use (e.g. Xu & Tenenbaum, 2007; Graf, Degen, Hawkins, & Goodman, 2016) but there remains a more fundamental evolutionary question: how do multiple levels of reference come to coexist in the lexicon to begin with? Our hypothesis, motivated both by classic work on concept representations and contemporary work on the selective pressures induced by communication, is that lexicalization of conceptual hierarchies is a function of (1) the structure and statistics of entities in the environment, and (2) the particular contexts in which communication occurs. In particular, we expect hierarchical lexica to form when features can be encoded as predictable clusters and communicative goals require distinctions to be drawn at multiple levels. To test this hypothesis, we designed a repeated reference game in which pairs of participants interactively created an artificial language to communicate with each other about objects in context (e.g. Winters, Kirby, & Smith, 2014; Galantucci & Garrod, 2011). In this game, participants were paired over the web and placed in a shared environment containing a grid of four objects (Fig. 1A) and a ‘chatbox’ to send messages from a pre-specified vocabulary of sixteen words (Fig. 1B). On each of ninety trials, one player — the ‘speaker’ — was privately shown a highlighted target object and allowed to send a single word to help their partner select this object 158