Emergence of Hierarchical Reference Systems in Multi-agent Communication

Proceedings of the 29th International Conference on Computational Linguistics, pages 5689–5706 October 12–17, 2022. 5689 Emergence of hierarchical reference systems in multi-agent communication Xenia Ohmer and Marko Duda and Elia Bruni Osnabrück University {xenia.ohmer,mduda,elia.bruni}@uos.de Abstract In natural language, referencing objects at dif- ferent levels of specificity is a fundamental pragmatic mechanism for efficient communica- tion in context. We develop a novel communica- tion game, the hierarchical reference game, to study the emergence of such reference systems in artificial agents. We consider a simplified world, in which concepts are abstractions over a set of primitive attributes (e.g., color, style, shape). Depending on how many attributes are combined, concepts are more general (“circle”) or more specific (“red dotted circle”). Based on the context, the agents have to communi- cate at different levels of this hierarchy. Our results show that the agents learn to play the game successfully and can even generalize to novel concepts. To achieve abstraction, they use implicit (omitting irrelevant information) and explicit (indicating that attributes are irrele- vant) strategies. In addition, the compositional structure underlying the concept hierarchy is reflected in the emergent protocols, indicating that the need to develop hierarchical reference systems supports the emergence of composi- tionality. 1 Introduction Humans excel at using language to convey informa- tion efficiently in context. A speaker does not have to communicate every detail. Rather, a listener can infer the intended meaning of an utterance by as- suming that sufficient information was provided. This idea was first explicitly formulated by Grice (1975) in his conversational maxims, in particular the Maxim of Quantity: “1. Make your contribu- tion as informative as is required (for the current purposes of the exchange). 2. Do not make your contribution more informative than is required.” An illustration of this mechanism can be given in the form of a simple referential context. In a scene with a red circle and a green triangle, “circle” is enough information to identify the referent, whereas more complex scenes may require the speaker to name both object attributes—shape and color—to allow for an unambiguous interpretation. The Maxim of Quantity requires a hierarchical reference system, that allows the selection of the most appropriate level of specificity for a given context. In this paper, we follow the proposal by Higgins et al. (2018) and define concepts as compositional abstractions over a set of primitive attributes (e.g., color, style, shape), see Figure 1a. The concepts are maximally specific at the leaf nodes, where all attribute values are determined. Moving from the subordinate levels up to the superordinate levels, the number of concept-defining attribute values de- creases. Thus, each parent concept is an abstraction (i.e. a subset) over its children and over the original set of attribute values. Given this definition of a concept hierarchy, we use a language emergence paradigm with artificial agents to study whether a corresponding reference system can emerge given a structured perception of the world. In most language emergence simulations, a sender and a receiver agent are trained on a …

Content truncated to 500 words