The Effect of Efficient Messaging and Input Variability on Neural-Agent Iterated Language Learning

TL;DR

This paper investigates how neural agents in iterated language learning adapt their communication strategies, revealing they tend to preserve input distributions rather than optimize for efficiency or systematic structure, influenced by input variability and learning constraints.

Contribution

It re-evaluates previous findings by incorporating factors like speaker bias, input variability, and learning bottlenecks, showing these influence language evolution in neural agents.

Findings

Neural agents maintain input utterance distributions rather than optimize for efficiency.

Input variability and learning bottlenecks affect language development in neural simulations.

Agents do not naturally develop systematic or more efficient languages under tested conditions.

Abstract

Natural languages display a trade-off among different strategies to convey syntactic structure, such as word order or inflection. This trade-off, however, has not appeared in recent simulations of iterated language learning with neural network agents (Chaabouni et al., 2019b). We re-evaluate this result in light of three factors that play an important role in comparable experiments from the Language Evolution field: (i) speaker bias towards efficient messaging, (ii) non systematic input languages, and (iii) learning bottleneck. Our simulations show that neural agents mainly strive to maintain the utterance type distribution observed during learning, instead of developing a more efficient or systematic language.