Andrea Silvi, Data Science och AI

Language is fundamentally shaped by pressures for efficient communication, typically characterised as a trade-off between simplicity and informativeness. While both information-theoretic models and multi-agent reinforcement learning have successfully captured these dynamics, they predominantly treat linguistic terms as atomic labels. Consequently, these existing frameworks struggle to account for the systematicity and recursiveness that are prevalent across languages.

This thesis extends the study of efficient communication to compositional domains with productive morphosyntax, focusing primarily on recursive numeral systems. Across four contributions, we employ computational modeling to explore how functional and learning pressures shape structured languages. First, we show through multi-agent reinforcement learning that artificial agents optimised solely for communicative success tend to prefer more efficient recursive numeral systems. Second, we argue that previous efficiency measures cannot account for regularity and introduce a different trade-off that can separate human systems from artificial ones that were previously considered optimal but were lacking human-likeness. Third, we connect regularity to learnability, using reinforcement learning to show that human numeral systems exhibit high regularity because they are inherently easier to learn. Finally, we expand this framework to an open-ended collaborative building task, showing that agents utilising procedural abstractions develop languages that minimise similar efficiency trade-offs.

Overall, this work attempts to bridge the gap between efficient communication models and the compositional reality of language, demonstrating how structure is consistently preferred because of communicative and cognitive constraints.