Artificial learners often behave differently from human learners in the context of neural agent-based emergent communication (Rita et al., 2022). The lack of human-like cognitive biases is one of the prevailing explanations (Chaabouni et al., 2019). Besides that, it has been proposed that more naturalistic settings of language learning and use could lead to more human-like patterns of language emergence (Lazaridou and Baroni, 2022). In this work, we explore the latter account in the context of an emerging word-order/case-marking trade-off, inspired by human experiments (Fedzechkina et al., 2017). We expose agents to artificial languages in a supervised language learning protocol and combine this with a communication task. We replicate the emergence of a trade-off during the communication phase without the need to hard-code a specific cognitive bias. We see this as an essential step towards the investigation of language universals with neural agents.

Natural languages tend to minimize the linear distance between heads and their dependents in a sentence, known as dependency length minimization (DLM). Such a preference, however, has not been consistently replicated with neural agent simulations. Comparing the behavior of models with that of human learners can reveal which aspects affect the emergence of this phenomenon. This work investigates the minimal conditions that may lead neural learners to develop a DLM preference. We add three factors to the standard neural-agent language learning and communication framework to make the simulation more realistic, namely: (i) the presence of noise during listening, (ii) context-sensitivity of word use, and (iii) incremental sentence processing. While no preference appears in production, we show that the proposed factors contribute to a small but significant learning advantage of DLM for listeners of verb-initial languages. Our findings offer insights into essential elements contributing to DLM preferences in purely statistical learners.

The mainstream approach in NLP is to train systems on large amounts of data. Such passive learning contrasts with the waylanguage is learnt by humans. Human language is acquired within communities, it is culturally transmitted and changesdynamically. These evolutionary mechanisms have been extensively studied in the field of Language Evolution. Despitelimited prior interaction between fields, such mechanisms are now increasingly incorporated into NLP systems. Suchmodels have the potential to both study the evolution of language in multi-agent simulations with state-of-the-art (deep)learning systems in more naturalistic settings and improve NLP systems by having language emerge organically. Weexamine how findings from a model by Havrylov & Titov (2017) compare to those from traditional Language Evolutionmodels and quantify the emerging compositionality using an existing Language Evolution method (Tamariz, 2011). Thisapproach reveals novel insights into the generated data, the applied methodology and the nature of compositionality.