Co-evolution of language and agents in referential games

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 2993–3004 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 2993 Co-evolution of language and agents in referential games Gautier Dagan University of Amsterdam gautierdagan@gmail.com Dieuwke Hupkes Facebook AI Research dieuwkehupkes@fb.com Elia Bruni University of Osnabr¨uck elia.bruni@gmail.com Abstract Referential games offer a grounded learning environment for neural agents which accounts for the fact that language is functionally used to communicate. However, they do not take into account a second constraint considered to be fundamental for the shape of human lan- guage: that it must be learnable by new lan- guage learners. Cogswell et al. (2019) introduced cultural transmission within referential games through a changing population of agents to constrain the emerging language to be learnable. How- ever, the resulting languages remain inherently biased by the agents’ underlying capabilities. In this work, we introduce Language Transmis- sion Simulator to model both cultural and ar- chitectural evolution in a population of agents. As our core contribution, we empirically show that the optimal situation is to take into ac- count also the learning biases of the language learners and thus let language and agents co- evolve. When we allow the agent popula- tion to evolve through architectural evolution, we achieve across the board improvements on all considered metrics and surpass the gains made with cultural transmission. These re- sults stress the importance of studying the un- derlying agent architecture and pave the way to investigate the co-evolution of language and agent in language emergence studies. 1 Introduction Human languages show a remarkable degree of structure and complexity. In the evolution of this complex structure, several different intertwined pressures are assumed to have played a role. The first of these pressures concerns the function of lan- guage: as language is to communicate, it should allow effective communication between proficient language users (e.g. Smith and Kirby, 2012). This pressure is strongly intertwined with the nature of the proficient user: what features of language allow effective communication depends on the abilities of the user to use the language. A second pressure on the shape of human lan- guage stems from the fact that language must be learnable. Unlike animal languages, which are taken to be mostly innate, human languages must be re-acquired by each individual (Pinker and Bloom, 1990; Hurford, 1998). A language can only survive if it can successfully be transmit- ted to a next generation of learners. In the field of language evolution, this transmission process is referred to as cultural transmission, while the process of change that occurs as a consequence is called cultural evolution.1 Like the pressures arising from the function of language, the way that cultural evolution shapes language also depends on the language users: what is learnable depends on the inductive biases of the learner. Computationally, the emergence of language can be studied through simulation with artificial agents and by investigating the resulting languages for structure, level of compositionality, and mor- phosyntactic properties …

Content truncated to 500 words