The Efficiency of Ambiguity: Abstract References Emerge in Coarse Contexts

We investigate how context granularity, i.e. whether fine or coarse distinctions need to be made, influences an emerging lexicon. We conduct an agent-based simulation of a concept-level reference game, in which agents learn to communicate about concepts that are operationalized by combining multiple objects. We create three experimental conditions by manipulating the context in which the instances of the target concept appear: In the fine context condition, agents must make precise distinctions between similar targets and distractors. In the coarse context condition, targets are easy to discriminate because they share no overlapping features with the distractors. In the mixed baseline condition, both fine and coarse distinctions are necessary.Our results suggest that agents adapt their communication strategies to the granularity of the context in which they learned the concepts. In the fine context and baseline conditions, agents develop a communication protocol heavily based on one-to-one mappings between messages and concepts. Conversely, in the coarse context condition agents communicate more efficiently by vastly relying on abstract references that may refer to more than a single concept but are unambiguous in context. These results show that ambiguity emerges in coarse contexts and that ambiguous abstract terms are used for more efficient communication.