Do large language models and humans follow similar learning stages? Assessing GPT-2’s order of Swedish grammar acquisition within the Processability Theory framework

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing & Linguistics · Depth: Expert, quick

Summary

Researchers investigated whether GPT-2 acquires Swedish grammatical structures in a similar implicational order to human second language (L2) learners, as predicted by Processability Theory (PT). They developed SwePT, a minimal pair dataset targeting Swedish syntax and morphology across four human L2 development stages. Evaluating fine-tuned GPT-2 models on SwePT via an acceptability classification task, the study found that observed acquisition orders correlated across models but violated PT's hypothesized implicational sequence. The relationship between classification performance and feature frequency distributions in minimal pairs suggests GPT-2's acquisition order is driven by unigram and n-gram heuristics. While acknowledging the need for further methodological refinement in adapting NLP to PT, the experiments provided no evidence for PT-like grammatical development in GPT-2.

Key takeaway

For NLP engineers evaluating large language models' linguistic capabilities, you should recognize that models like GPT-2 do not necessarily mimic human second language acquisition stages. Your evaluations should account for the strong influence of unigram and n-gram frequency heuristics on model performance, rather than assuming human-like grammatical development. This implies a need to design specific tests that differentiate between frequency-based learning and deeper structural understanding when assessing language acquisition in LLMs.

Key insights

GPT-2's Swedish grammar acquisition does not follow human L2 stages predicted by Processability Theory, instead relying on frequency heuristics.

Principles

• Language model acquisition may differ from human L2.
• Frequency heuristics influence LLM grammar learning.
• Processability Theory might not apply to LLMs.

Method

Developed SwePT, a minimal pair dataset for Swedish grammar stages. Evaluated fine-tuned GPT-2 models using an acceptability classification task to assess acquisition order against Processability Theory.

In practice

• Use minimal pair datasets for grammar evaluation.
• Analyze n-gram heuristics in LLM language tasks.
• Compare LLM learning to human language theories.

Topics

• Large Language Models
• GPT-2
• Processability Theory
• Swedish Grammar
• Second Language Acquisition
• N-gram Heuristics

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Large Language Models in NLP — LLMs integrate NLP advancements with massive scale to create general-purpose, prompt-driven language systems.
• Language in the brain — Statistical learning and synaptic plasticity are insufficient for modeling human language competency.
• Filling in the Mechanisms: How do LMs Learn Filler-Gap Dependencies under Developmental Constraints? — LLMs show shared, item-sensitive filler-gap dependency mechanisms, but require more data than humans.
• Dango: A Strictly L1-Only Large Language Model for Studying Second Language Acquisition — Dango, a 1.8B-parameter LLM, enables controlled study of L1-to-L2 transfer in SLA by mitigating L2 contamination and simulating acquisition.
• Do Large Language Models Possess a Theory of Mind? A Comparative Evaluation Using the Strange Stories Paradigm — GPT-4o demonstrates human-comparable Theory of Mind capabilities in text-based evaluations, outperforming smaller LLMs.
• Multilinguality of Large Language Models From a Structural Perspective — LLMs' multilinguality involves structural differences, with low-resource languages diverging more from English.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.