Learning filler-gap dependencies with neural language models: Testing island sensitivity in Norwegian and English
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Human linguistic input is often claimed to be impoverished with respect to linguistic evidence for complex structural generalizations that children induce. The field of language acquisition is currently debating the ability of various learning algorithms to accurately derive target generalizations from the input. A growing body of research explores whether Neural Language Models (NLMs) can induce human-like generalizations about filler-gap dependencies (FGDs) in English, including island constraints on their distribution. Based on positive results for select test cases, some authors have argued that the relevant generalizations can be learned without domain-specific learning biases (Wilcox et al., 2023), though other researchers dispute this conclusion ((Lan et al., 2024b; Howitt et al.,2024). Previous work focuses solely on English, but broader claims about filler-gap dependency learnability can only be made based on multiple languages and dependency types. To address this gap, we compare the ability of NLMs to learn restrictions on FGDs in English and Norwegian. Our results are mixed: they show that although these models acquire some sophisticated generalizations about filler-gap dependencies in the two languages, their generalizations still diverge from those of humans. When tested on structurally complex environments, the models sometimes adopt narrower generalizations than humans do or overgeneralize beyond their input in non-human-like ways. We conclude that current evidence does not support the claim that FGDs and island constraints on them can be learned without domain-specific biases.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it