Artificial Grammar Learning in children, adults, animals and machines

Carel ten Cate, Judit Gervain, Clara C. Levelt, Christopher I. Petkov, Willem ZuidemaHAL 2020

TL;DR

This paper is an editorial introduction to a special issue on Artificial Grammar Learning (AGL), a field that investigates how humans, animals, and machines learn structured sequences. It synthesizes recent advances, highlights open questions, and argues that AGL is a powerful interdisciplinary tool for understanding the evolution and cognitive underpinnings of language and other complex learning abilities.

What Problem Does This Paper Address?

Human languages are characterized by grammar, a system of rules for combining symbols into meaningful expressions. Despite extensive research, the evolutionary, developmental, cognitive, and computational bases of grammatical abilities remain poorly understood. A key question is whether abstract rule acquisition relies on general learning mechanisms or language-specific ones. Artificial Grammar Learning (AGL) studies offer a way to investigate these questions by using simplified, artificial languages to study how rules and structured sequences are learned across different species and systems. This special issue aims to synthesize the growing body of AGL literature, critically assess its contributions, and identify future research directions.

Methods

The paper describes the Artificial Grammar Learning (AGL) paradigm, which involves exposing participants (humans, animals, or machines) to sequences of arbitrary items (sounds, visual stimuli) that follow specific, pre-defined rules. After exposure, participants' knowledge of the grammar is tested through recall or discrimination tasks. The paper highlights various adaptations of this paradigm for different populations, including:

  • Infants: Behavioral (head-turn, looking time) and neurophysiological (EEG, fNIRS) measures are used to assess implicit processing of rule-based strings.
  • Non-human animals: Familiarization/habituation tasks and operant discrimination tasks are employed.
  • Machines: Computational modeling and machine learning techniques are used to simulate and understand learning mechanisms.

The paper also discusses methodological advancements, such as using graph theory to explicitly define hierarchical structures and the importance of considering whether participants are exposed to both legal and illegal strings during training.

Datasets

Not discussed in the paper. The paper focuses on the AGL paradigm itself, which uses artificial stimuli and grammars rather than pre-existing datasets.

Key Insights and Ideas

  • AGL as a unifying framework: The AGL paradigm provides a common ground for studying rule learning across diverse species (humans, infants, animals) and systems (machines), facilitating cross-disciplinary comparisons.
  • Beyond "rule learning": AGL tasks can reveal insights into broader perceptual and cognitive abilities, such as pattern recognition and the detection of dependencies, which may predate and bootstrap the evolution of grammatical structure learning.
  • Species similarities and differences: While humans generally excel at complex grammatical structures, evidence suggests that many underlying learning mechanisms and biases are shared with other animals, implying that the gap might be one of degree rather than kind.
  • Importance of explicit methodology: Explicitly defining assumptions, theoretical considerations, and using tools like graph theory can lead to sharper research questions and more robust findings in AGL.
  • The role of constraints vs. rules: The paper raises the possibility that learning might be driven by detecting constraints rather than explicit rules, and that incorporating "semantics" (meaning) into AGL tasks could be crucial for understanding more complex learning.
  • Developmental parallels: Similarities in the development of structured vocalizations in songbirds and human infants suggest potential evolutionary pathways for sequence learning.

Key Results

The paper synthesizes findings from various studies within the special issue, rather than presenting novel experimental results. Key qualitative findings include:

  • Infants as young as 4 months can extract patterns from input after brief familiarization.
  • Non-human animals can detect certain types of regularities and dependencies in structured sequences, though their capabilities may differ from humans.
  • Non-adjacent dependencies, which are crucial for complex grammars, are difficult to learn even for adult humans in AGL tasks without specific cues.
  • Auditory biases and grouping mechanisms that facilitate rule learning in humans are also present in other species.

Limitations and Future Work

  • Limitations of AGL: The paper acknowledges that AGL tasks often use simplified stimuli and may not fully capture the complexity of natural language. The interpretation of results can be contentious, and there's a need for more direct cross-species comparisons. Methodological limitations, such as potential memory constraints in animals, can affect outcomes.
  • Future Work:
    • Broader species testing: More species need to be tested with comparable designs and stimulus sets to better assess species differences.
  • Refined methodologies: Developing more explicit frameworks (e.g., graph theory) and incorporating semantic information into AGL tasks are crucial.
  • Computational modeling: Integrating computational models with empirical AGL approaches can help formalize theories, design experiments, and analyze data more effectively.
  • Evolutionary pathways: Further comparative studies can shed light on the evolutionary origins of sequence and grammar learning mechanisms.
  • Understanding limits: Continued research is needed to fully understand the limits of animal processing capacities and evolutionary patterns in rule learning.

Summary updated

Abstract

International audience
BibTeX
@inproceedings{cate2020artificial,
  author = {Carel ten Cate and Judit Gervain and Clara C. Levelt and Christopher I. Petkov and Willem Zuidema},
  title = {{Artificial Grammar Learning in children, adults, animals and machines}},
  booktitle = {HAL},
  year = {2020},
}

Loading related content...

Figure