Edinburgh Virtual Workshop on
Artificial Language Learning
4-6 October, 2021

This talk will discuss both practical and theoretical issues related to the use of artificial language learning (ALL) experiments to study phonological learning biases. I will begin by briefly contextualising this line of research within the field and discussing why ALL experiments can be a useful tool. The remainder of the talk will provide a survey of specific topics/findings from my (and others') research, types of biases studied, and common paradigms and methodologies. Discussion of specific studies will be interwoven with practical discussion of methods and design considerations that will (hopefully) be helpful for researchers looking to add ALL experiments to their repertoire.

In this practical tutorial, we will discuss various factors to consider when designing stimuli for artificial language learning experiments. The discussion will focus on phonology-oriented studies, but some principles will be broadly applicable. Topics will include issues such as choosing the phonemic/phonetic inventory for the language, choosing word shape, designing critical and filler items, considering the L1 of participants, recording/synthesising stimuli, using visual stimuli (or not), making linguistic patterns of interest more or less apparent, and so forth.

A well-known phenomena in language acquisition is that children can acquire rules when they are available (Berko 1955; Brown 1973). But precisely how children acquire such rules under circumstances of language variation is not as frequently discussed. In the sociolinguistics literature, it is widely known that children exposed to stable language variation (e.g. Labov, 1989) can learn and match this variation in their own productions (Labov & Roberts, 1995; Roberts, 1997). However, evidence from a related literature suggests that children exposed to variation that is inconsistent – for example, from late learners or pidgin or creole speakers - do not learn and match the probabilities of the input variation, instead regularizing the language in their own productions (Singleton & Newport 2004, Hudson-Kam & Newport, 2005, 2009). In my work, I ask how we reconcile these two related but seemingly conflicting findings. What are the principles that govern when children learn rules and when they reproduce variation or inconsistency? In a series of artificial language learning experiments, I explore how the structure of the input, the maturational state of the learner, and the social context of the learning could constrain the acquisition of variation.

In this talk I will present two artificial language learning studies which aim to test the relationship between morphology and learning. The first study targets the impact of redundant morphological marking on learning. Redundant marking is found across languages. Its presence is somewhat puzzling in light of the growing evidence on speakers’ tendency to avoid redundant elements in production. I propose that redundant cues may be functional for learning. In particular, I argue that redundant morphological cues can facilitate learning, even when they make the language system more complicated. I test this hypothesis in an artificial language learning study, where either word order alone or both word order and case marking serve as cues for thematic assignment in a novel construction. Results show that child learners learn the language with the redundant cue better despite having to learn an additional morpheme.
The second study uses an artificial language learning paradigm to investigate the emergence of differential object marking systems (where only certain types of grammatical objects are marked with morphological case). In particular, I test a recent hypothesis according to which differential object marking is driven by atypical pragmatic-discourse properties. Results show that as in documented cases of differential object marking emergence, atypical pragmatic-discourse properties can bring about differential object marking systems via word order variation.
Together, these studies illustrate how artificial language learning studies can complement, support, and adjudicate between linguistic theories of language acquisition and change.

While data collection traditionally takes place face-to-face, recent years have seen an explosion in the use of online data collection: participants take part remotely, providing responses through a survey tool or custom experimental software running in their web browser, with surveys or experiments often being advertised on crowdsourcing websites like Amazon Mechanical Turk (MTurk) or Prolific. Online methods potentially allow rapid and low-effort collection of large samples, and are particularly useful in situations where face-to-face data collection is not possible (e.g. during a pandemic); however, building and running these experiments poses challenges that differ from lab-based methods. In this tutorial we’ll give you a brief overview of how to build online artificial language learning experiments that run in a web browser (using jsPsych and javascript, which we think provide a nice balance of flexibility and accessibility), and how to crowdsource participants; we’ll focus on summarising the main technical challenges of this work and sharing some tips and tricks we’ve picked up in the course of doing online data collection in our own research, and we’ll provide pointers to materials that will allow you to work on building your experiments in your own time.

The majority of artificial language learning experiments conducted each year use adult participants, while many fewer are conducted with children. One reason for this imbalance is that children are less willing and able to participate in these lengthy (often boring) experiments that can require hundreds of learning and test trials. In this tutorial, I’ll cover some basic design principles that will allow you to convert your adult artificial language learning experiment into a version suitable for use with children. I’ll include some tips for both live in-person experiments and online unsupervised experiments. Then, you’ll work together in small groups to apply these principles to a real artificial language learning experiment, converting an experiment originally designed for adults into a child-friendly version.

Abstract.

Abstract.