Lexical search and social reasoning jointly explain communication in associative reference games

Abstract

Effective linguistic communication depends upon many different cognitive processes working together in concert. Yet our computational models of these processes are often developed in isolation, without considering how these processes fit together. In this work, we study a simplified variant of the popular board game Codenames, which highlights the integration of two important processes: (1) lexical search and (2) pragmatic reasoning. In this open-ended task, speakers must generate a 'clue' word from their vast lexicon that allows their partner to select a pair of target words from a context of distractors. In Experiment 1, we examine a large corpus of associative reference games to evaluate models of how speakers search for information in their lexicon and ultimately produce informative clues. We find that a two-stage model of pragmatic reasoning built on top of semantic search was best able to predict behavior in the game. In Experiment 2, we explicitly elicit lists of candidate clues that speakers are considering prior to selection. We use these data to evaluate models of semantic search initialized from multiple nodes in the lexical network, and find models that jointly model the search process best account for the behavioral patterns. Finally, in Experiment 3, we examine targeted endorsements for a particularly diagnostic set of clues, which provide further evidence for this integrated model of search and selection. Taken together, our findings suggest that context-sensitivity is infused into both retrieval processes (generating promising candidates) and selection processes (evaluating the likely communicative success of these candidates).