Pith. sign in

REVIEW 2 major objections 5 minor 40 references

The paper claims that an automated hybrid parser-based detector can reliably measure children's filler-gap input and production at corpus scale, revealing construction-specific extraction biases and a generalization asymmetry in language mo

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-02 19:26 UTC pith:DVOHNGZK

load-bearing objection A genuinely useful automated detector and corpus-scale measurements, but the child-production numbers need a recall check on child speech before they can be treated as stable. the 2 major comments →

arxiv 2603.02082 v2 pith:DVOHNGZK submitted 2026-03-02 cs.CL

What Exactly do Children Receive in Language Acquisition? A Case Study on CHILDES with Automated Detection of Filler-Gap Dependencies

Zhenghao Herbert Zhou , William Dai , Maya Viswanathan , Simon Charlow , R. Thomas McCoy , Robert Frank This is my paper
classification cs.CL
keywords filler-gap dependencieschild-directed speechCHILDESwh-questionsembedded questionsrelative clausesextraction sitefiltered corpus training
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Filler-gap dependencies—sentences in which a moved wh-word or relative pronoun is interpreted in an empty position—sit at the center of a debate between innate-knowledge and statistical-learning accounts of acquisition. The paper tries to settle the empirical side of that debate by showing that a hybrid detector, built from off-the-shelf constituency and dependency parses plus hand-written heuristics, can label three filler-gap constructions and their extraction sites in large, noisy child-directed corpora. Applying it to 57 CHILDES corpora, it reports that matrix questions are about ten times more frequent than embedded questions or relative clauses, that both wh-question families show a subject/object asymmetry favoring object extraction while relative clauses are roughly balanced, and that children's production rankings mirror adult input. In a filtered-training experiment with small language models, removing matrix questions from the training data degrades performance on embedded questions and relative clauses, but not the reverse. If these measurements hold, they give acquisition theories and computational models a concrete, fine-grained picture of the input children actually receive.

Core claim

The central claim is that the detector is a viable solution for large-scale, fine-grained detection of target filler-gap structures even though it is imperfect. After validation against human annotations, the paper applies it to 2.84 million child-directed and child utterances from 57 CHILDES corpora and reports that matrix questions are an order of magnitude more frequent than embedded questions or relative clauses; object extraction outnumbers subject extraction in both question types; and relative clauses are roughly balanced between subject and object. The same fine-grained labels support a training intervention: removing matrix questions from a child-directed corpus significantly degrad

What carries the argument

The load-bearing tool is a hybrid detection pipeline that combines two parse types: constituency parses supply local structural signatures such as NP -> NP SBAR for relative clauses and SBARQ -> WHNP SQ for matrix questions, while dependency parses validate head-dependent relations such as relcl, nsubj, and dobj to pin down the extraction site. Additional heuristics, including a lexical filter over matrix verbs, separate embedded questions from free relatives. The detector outputs per-utterance labels for three constructions and subtypes (subject, object, adjunct, polar, possessive, reduced), which is the granularity that makes the corpus statistics and the filtered-training experiment possi

Load-bearing premise

The load-bearing premise is that errors made by the automatic parsers on noisy child speech do not systematically shift the counts of constructions or extraction sites; if parser failures are correlated with, say, fragmentary child questions, the reported object-bias and relative-clause balance could be artifacts rather than properties of the input.

What would settle it

Take a fresh sample of child-directed and child utterances from CHILDES, have linguists annotate every utterance for the three constructions and gap sites, then run the detector on the same utterances. If precision or recall on any category—especially adjunct relatives and reduced object relatives, where the paper's own validation is weakest—falls below 0.7, the corpus statistics would not support the conclusions.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X LinkedIn Reddit HN

If this is right

  • Large-scale, fine-grained annotation of filler-gap constructions no longer requires manual trace annotation; the released detector and labels can be reused on any English CHILDES-style corpus.
  • Children's filler-gap input is dominated by matrix questions by roughly an order of magnitude, so any frequency-driven account of acquisition must operate with construction- and extraction-site-specific counts rather than collapsed totals.
  • Object-biased extraction is a property of wh-questions but not relative clauses in both adult input and child output, making extraction-site asymmetry construction-specific.
  • In language models trained on child-directed data, matrix questions act as a source of generalization to embedded questions and relative clauses, but the rarer constructions do not feed back into matrix questions.
  • The generated labels support future input-attribution studies that quantify which construction-specific training examples most affect model predictions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural reading of the findings is that the input contains enough distributional evidence for a learner to induce extraction-site biases from frequency, weakening the poverty-of-stimulus argument for these constructions; the paper itself stops short of drawing that conclusion.
  • The matrix-question transfer result suggests a finer-grained, testable prediction: removing object-extracted matrix questions should impair object-extracted embedded questions more than subject-extracted ones, since the shared dependency structure is the object gap.
  • Because child production rankings track adult rankings in these cross-sectional snapshots, the paper establishes correlation, not causation; a longitudinal cross-lagged analysis on the same corpora could test whether adults' construction-specific rates actually predict later child rates.
  • Extending the detector to clefting, topicalization, and free relatives would reveal whether the question/relative-clause split in extraction-site bias is a feature of filler-gap dependencies generally or specific to the three constructions studied.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper develops an automated annotation pipeline that combines constituency and dependency parsing to detect matrix wh-questions, embedded wh-questions, and relative clauses in English, with subtype labels for extraction site (subject, object, adjunct). The detector is evaluated in two ways: a manual evaluation of detector-positive CHILDES sentences (precision-oriented) and a comparison against Pearl and Sprouse's hand-annotated corpus (precision/recall on adult child-directed speech). The tool is then applied to 57 CHILDES corpora to estimate adult-input and child-production rates of filler-gap constructions over development, yielding claims about construction frequency, extraction-site asymmetries, and child-adult mirroring. A BabyLM filtered-corpus-training experiment tests whether removing one construction family from training data degrades performance on the same or other construction families.

Significance. If the detector is sufficiently accurate, this is a useful methodological contribution: it would be the first large-scale, fine-grained automated annotation of filler-gap constructions in CHILDES, enabling descriptive acquisition statistics and controlled LM training interventions that were previously impractical. The strengths of the paper include the release of code and data, validation against an external human-annotated corpus, the hybrid use of two parser types, and the controlled BabyLM ablation with random-sentence controls. The corpus-level statistics and the asymmetry results are also directly relevant to the poverty-of-the-stimulus debate. The main unresolved threat is the unmeasured recall of the detector on child production, which is load-bearing for the child-production trajectory claims.

major comments (2)
  1. [§4.1, §5 (Figs. 4, 6, 7)] The manual evaluation in §4.1 samples only sentences that the detector labeled as belonging to a target category; it therefore estimates precision, not recall, for child speech. The Pearl–Sprouse comparison in §4.2 provides recall estimates only for adult child-directed speech. Yet Figures 4 and 6 report child production rates and subject/object log-ratios, and §5.1 interprets child/adult mirroring. If fragmentary or ungrammatical child utterances fail at the parse stage before the construction heuristics apply, false negatives could be construction- or gap-site-specific and could create the observed trajectories and asymmetries. Please add a recall estimate on child speech—for example, by hand-annotating a random sample of child utterances (or a held-out set of child transcripts) and computing per-category recall—or explicitly restrict the child-production claims to precision-only state
  2. [Appendix C, §5, §6] The embedded-question detector depends on a hand-curated list of 17 matrix verbs derived from the top 50 verb lemmas in BabyLM. The paper does not report how many CHILDES embedded questions are headed by verbs outside this list, or a CHILDES-specific recall check for SEQ/OEQ. Because the embedded-question rates in §5 and the embedded-question ablation in §6 assume reasonably complete detection, a coverage analysis would be needed—for example, listing the matrix verbs that actually occur in potential embedded-question structures in CHILDES and the proportion captured by the 17-verb list—to show that the filter does not systematically undercount embedded questions in the target corpus.
minor comments (5)
  1. [§4.1, Table 2] The text says 'six core categories' and 'sampled 100 sentences from both child-directed speech and child speech,' but Table 2 lists eight categories and appears to report 100 sentences per category. Please clarify the sampling design.
  2. [Table 2] The metric labeled 'accuracy' is actually precision given the sampling of detector positives. Rename it 'precision' or clarify the denominator.
  3. [§5.1, Figure 4] The text claims that 'adjunct and polar questions outnumber subject and object extractions,' but Figure 4 groups these under 'Other,' so the claim is not directly visible. Please provide a separate table or figure showing adjunct and polar rates individually.
  4. [§5.3, Figure 8, Table 6] The significance test underlying the 'n.s.' and '***' markers is not described: the paper does not state whether the pairwise comparisons use bootstrap CIs, a paired test, or something else, nor does it give the minimum-count threshold used for the age bins.
  5. [Figure 6] The caption says 'Sizes of dot shades indicate number of total utterances within each bin,' but there is no legend or scale for the dot shading. Please add one or describe the mapping in the caption.

Circularity Check

0 steps flagged

No significant circularity: detector is validated against external gold annotations and the LM generalization study uses independent synthetic evaluations.

full rationale

The paper's claimed derivation is not circular. The detection system in §3 is a set of hand-written heuristics over constituency and dependency parses; no parameter is fitted to the corpus statistics that the paper later reports. The viability claim in §4.2 is supported by two external validations: a manual precision sample (§4.1, Table 2) and a comparison against the independently human-annotated Pearl & Sprouse (2013) corpus (§4.2, Table 3), which provides precision and recall. The CHILDES statistics in §5 are descriptive measurements produced by applying this detector, not predictions of those measurements. The BabyLM ablation in §6 uses the detector only to filter training corpora; the evaluation is on 3,432+5,000+5,000 synthetically constructed minimal pairs whose grammaticality judgments are defined independently of the detector. The paper explicitly acknowledges its limitations ('Our detectors are imperfect... the datasets we have extracted may not be suitable for research questions that require perfect identification'), which is an honest measurement-error caveat rather than an admission of circularity. No load-bearing self-citation or imported uniqueness theorem appears; the cited prior work (e.g., Pearl & Sprouse 2013; Boguraev et al. 2025) provides independent evidence or is used for comparison. The only substantive risk—that detector errors (especially unmeasured recall on child speech) could bias the §5 child/adult comparisons—is an empirical artifact risk, not a definitional reduction of the paper's outputs to its inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 8 axioms · 0 invented entities

The central empirical contribution rests on the detector's heuristic rules and the validity of the corpora and gold labels. Most assumptions are domain assumptions about parse reliability and corpus representativeness; the one clearly ad hoc element is the matrix-verb list that separates embedded questions from free relatives.

free parameters (4)
  • Embedded-question matrix verb list = [know, see, tell, look, remember, wonder, guess, ask, say, forget, figure, understand, decide, show, watch, hear, think]
    Manually compiled from the top-50 verb lemmas covering 95% of potential embedded-question sentences (Appendix C); determines which sentences count as embedded questions, directly affecting SEQ/OEQ/AEQ/PEQ frequencies and LM filtering.
  • Age range and bin widths = 3-80 months; 6-month bins for Figure 4, 3-month bins for Figure 6; Laura 15-42 months
    Chosen after data-sparsity checks; affects the shape of developmental trajectories and the number of bins entering statistical comparisons.
  • Minimum-count threshold for subject-share comparisons = not specified
    Bins below threshold are excluded in §5.3 and Appendix F, but the threshold value is not stated, making the pairwise tests not fully reproducible.
  • Epsilon in log ratio = unspecified
    Smoothing constant in log((#SUBJ+ε)/(#OBJ+ε)) used for Figure 6; the value is not given, so readers cannot reproduce the exact plotted ratios.
axioms (8)
  • domain assumption CHILDES transcripts are a representative sample of children's linguistic input and output.
    The entire English-NA CHILDES database is used as the empirical basis in §5; representativeness is assumed.
  • domain assumption Speaker metadata in CHILDES allows reliable separation of adult vs. child speech after age corrections.
    Appendix E describes age corrections; the reliability of the metadata is assumed.
  • domain assumption spaCy and Berkeley Neural parser outputs are accurate enough for the heuristic detection rules.
    All detection steps in §3 depend on parse correctness; validation measures errors but does not provide an error model for corpus-level statistics.
  • domain assumption Pearl & Sprouse (2013) trace annotations, converted by Appendix D heuristics, are valid gold labels.
    Used as the reference standard for precision/recall in §4.2.
  • domain assumption Filler-gap theory (movement/traces) and extraction-site inference from surface configurations are the correct linguistic model.
    The target typology in Table 1 assumes trace theory and does not consider non-movement alternatives.
  • ad hoc to paper The hand-curated list of 17 question-embedding verbs is sufficient to distinguish embedded questions from free relatives in CHILDES/BabyLM.
    Appendix C; all other matrix verbs are excluded, which can depress embedded-question counts and alter LM filtering.
  • domain assumption Continuation probability on synthetic minimal pairs is a valid measure of FGD knowledge in language models.
    Section 6 relies on the grammatical/ungrammatical contrast in each minimal pair being the only relevant difference.
  • domain assumption Filtered-corpus control conditions (same token count, random sentence removal) isolate the effect of construction absence.
    Section 6.2 assumes random removal is a neutral control despite potentially removing other constructions and affecting the model in other ways.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of What Exactly do Children Receive in Language Acquisition? A Case Study on CHILDES with Automated Detection of Filler-Gap Dependencies." pith.science (2026). https://pith.science/paper/DVOHNGZK

@misc{pith2026260302082,
  author       = {Pith},
  title        = {Pith review of: What Exactly do Children Receive in Language Acquisition? A Case Study on CHILDES with Automated Detection of Filler-Gap Dependencies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DVOHNGZK}},
  note         = {Machine review of arXiv:2603.02082}
}
Share X LinkedIn Reddit HN
read the original abstract

Children's acquisition of filler-gap dependencies has been argued by some to depend on innate grammatical knowledge, while others suggest that the distributional evidence available in child-directed speech suffices. Unfortunately, the relevant input is difficult to quantify at scale with fine granularity, making this question difficult to resolve. We present a system that identifies three core filler-gap constructions in spoken English corpora -- matrix wh-questions, embedded wh-questions, and relative clauses -- and further identifies the extraction site (i.e., subject vs. object vs. adjunct). Our approach combines constituency and dependency parsing, leveraging their complementary strengths for construction classification and extraction site identification. We validate the system on human-annotated data and find that it scores well across most categories. Applying the system to 57 English CHILDES corpora, we are able to characterize children's filler-gap input and their filler-gap production trajectories over the course of development, including construction-specific frequencies and extraction-site asymmetries. The resulting fine-grained labels enable future work in both acquisition and computational studies, which we demonstrate with a case study using filtered corpus training with language models.

Figures

Figures reproduced from arXiv: 2603.02082 by Maya Viswanathan, Robert Frank, R. Thomas McCoy, Simon Charlow, William Dai, Zhenghao Herbert Zhou.

Figure 1
Figure 1. Figure 1: Dependency (top) and constituency (bottom) [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Dependency parses of an embedded question (left) and a matrix question (right). [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Constituency parses of an embedded question (left) and a matrix question (right). [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Adult (solid) and child (dashed) speech distributions across time by constructions and extraction sites. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Per-1,000 utterance rates of all filler-gap dependency sentences received and produced by Laura. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Subject- versus object-extracted log ratios for each construction across ages, binned by 3 months. Sizes [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Cross-construction comparisons of subject [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Cross-construction evaluations on filtered cor [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Dependency parse (top) and constituency parse (bottom) of object matrix question What’s your name? give different labeling if they choose different in￾terpretations. An additional practical motivation is imperfec￾tions in the source data: child speech can be noisy, ungrammatical, and subject to transcription errors, which could lead to parsing errors. Combining in￾formation from both parses makes the dete… view at source ↗
Figure 11
Figure 11. Figure 11: Sample annotation with trace (bolded) from [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Cross-construction evaluations on filtered [PITH_FULL_IMAGE:figures/full_fig_p014_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: A flowchart for the relative clause detection process. [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 2 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Ben Ambridge, Evan Kidd, Caroline F Rowland, and Anna L Theakston. 2015. The ubiquity of frequency effects in first language acquisition. Journal of Child Language, 42(2):239--273

    2015

  4. [4]

    Emily Atkinson, Matthew W Wagers, Jeffrey Lidz, Colin Phillips, and Akira Omaki. 2018. Developing incrementality in filler-gap dependency processing. Cognition, 179:132--149

    2018

  5. [5]

    Lois Bloom, Susan Merkin, and Janet Wootten. 1982. ``wh''- Questions : Linguistic Factors That Contribute to the Sequence of Acquisition . Child development, pages 1084--1092

    1982

  6. [6]

    Sasha Boguraev, Christopher Potts, and Kyle Mahowald. 2025. Causal Interventions Reveal Shared Structure Across English Filler -- Gap Constructions . In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 25032--25053

    2025

  7. [7]

    Susan R Braunwald. 1971. Mother-child communication: The function of maternal-language input. Word, 27(1-3):28--50

    1971

  8. [8]

    Roger Brown. 1973. A first language: The early stages. Harvard University Press, Cambridge, MA

    1973

  9. [9]

    Iona Carslaw, Sivan Milton, Nicolas Navarre, Ciyang Qing, and Wataru Uegaki. 2025. Automatic extraction of clausal embedding based on large-scale E nglish text data. In Proceedings of the Society for Computation in Linguistics 2025, pages 322--332

    2025

  10. [10]

    Chi-Yun Chang, Xueyang Huang, Humaira Nasir, Shane Storks, Olawale Akingbade, and Huteng Dai. 2025. Mind the Gap : How BabyLMs Learn Filler - Gap Dependencies . In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 15060--15076

    2025

  11. [11]

    Noam Chomsky. 1977. On W h-movement. In Formal Syntax , pages 71--132. Academic Press, New York

    1977

  12. [12]

    Noam Chomsky. 1986. Knowledge of Language : Its Nature , Origin , and Use . Convergence. Praeger, New York

    1986

  13. [13]

    Atrey Desai and Sathvik Nair. 2026. https://arxiv.org/abs/2604.14459 Filling in the Mechanisms : How do LMs Learn Filler - Gap Dependencies under Developmental Constraints ? Preprint, arXiv:2604.14459

    Pith/arXiv arXiv 2026

  14. [14]

    Holger Diessel and Michael Tomasello. 2005. A New Look at the Acquisition of Relative Clauses . Language, 81(1):1--25

    2005

  15. [15]

    Michael C Frank. 2023. Bridging the data gap between children and large language models. Trends in Cognitive Sciences, 27(11):990--992

    2023

  16. [16]

    Annie Gagliardi, Tara M Mease, and Jeffrey Lidz. 2016. Discontinuous development in the acquisition of filler-gap dependencies: Evidence from 15-and 20-month-olds. Language Acquisition, 23(3):234--260

    2016

  17. [17]

    Roger Grosse, Juhan Bae, Cem Anil, Nelson Elhage, Alex Tamkin, Amirhossein Tajdini, Benoit Steiner, Dustin Li, Esin Durmus, Ethan Perez, Evan Hubinger, Kamilė Lukošiūtė, Karina Nguyen, Nicholas Joseph, Sam McCandlish, Jared Kaplan, and Samuel R. Bowman. 2023. https://arxiv.org/abs/2308.03296 Studying Large Language Model Generalization with Influence Func...

    Pith/arXiv arXiv 2023

  18. [18]

    Yiding Hao. 2020. https://doi.org/10.18653/v1/2020.blackboxnlp-1.28 Evaluating Attribution Methods using White - Box LSTM s . In Proceedings of the Third BlackboxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, pages 300--313, Online. Association for Computational Linguistics

    work page doi:10.18653/v1/2020.blackboxnlp-1.28 2020

  19. [19]

    Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. 2020. https://doi.org/10.5281/zenodo.1212303 spaCy: Industrial-strength Natural Language Processing in Python

    doi:10.5281/zenodo.1212303 2020

  20. [20]

    Katherine Howitt, Sathvik Nair, Allison Dods, and Robert Melvin Hopkins. 2024. Generalizations across filler-gap dependencies in neural language models. In Proceedings of the 28th conference on computational natural language learning, pages 269--279

    2024

  21. [21]

    Yaling Hsiao, Nicola J Dawson, Nilanjana Banerji, and Kate Nation. 2023. The nature and frequency of relative clauses in the language children hear and the language children read: A developmental cross-corpus analysis of English complex grammar. Journal of Child Language, 50(3):555--580

    2023

  22. [22]

    Rodney D Huddleston, Geoffrey K Pullum, and Laurie Bauer. 2002. The Cambridge grammar of the English language . Cambridge University Press Cambridge

    2002

  23. [23]

    Jaap Jumelet, Milica Denic, Jakub Szymanik, Dieuwke Hupkes, and Shane Steinert-Threlkeld. 2021. Language models use monotonicity to assess NPI licensing. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 4958--4969

    2021

  24. [24]

    Nikita Kitaev, Steven Cao, and Dan Klein. 2019. https://doi.org/10.18653/v1/P19-1340 Multilingual constituency parsing with self-attention and pre-training . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3499--3505, Florence, Italy. Association for Computational Linguistics

    doi:10.18653/v1/p19-1340 2019

  25. [25]

    Nikita Kitaev and Dan Klein. 2018. https://doi.org/10.18653/v1/P18-1249 Constituency parsing with a self-attentive encoder . In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2676--2686, Melbourne, Australia. Association for Computational Linguistics

    doi:10.18653/v1/p18-1249 2018

  26. [26]

    Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via influence functions. In International conference on machine learning, pages 1885--1894. PMLR

    2017

  27. [27]

    Nur Lan, Emmanuel Chemla, and Roni Katzir. 2024. Large language models and the argument from the poverty of the stimulus. Linguistic Inquiry, pages 1--28

    2024

  28. [28]

    Julie Anne Legate and Charles D Yang. 2002. Empirical re-assessment of stimulus poverty arguments. The Linguistic Review, 19(1-2):151--162

    2002

  29. [29]

    Brian MacWhinney. 2000. The CHILDES Project : Tools for Analyzing Talk . V olume II : The Database , 3rd edition. Lawrence Erlbaum Associates, Mahwah, NJ

    2000

  30. [30]

    Kanishka Misra and Kyle Mahowald. 2024. Language models learn rare phenomena from less rare phenomena: The case of the missing AANN s. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 913--929

    2024

  31. [31]

    Abhinav Patil, Jaap Jumelet, Yu Ying Chiu, Andy Lapastora, Peter Shen, Lexie Wang, Clevis Willrich, and Shane Steinert-Threlkeld. 2024. Filtered corpus training ( FiCT ) shows that language models can generalize from indirect evidence. Transactions of the Association for Computational Linguistics, 12:1597--1615

    2024

  32. [32]

    Lisa Pearl and Jon Sprouse. 2013. Computational models of acquisition for islands. Experimental Syntax and Islands Effects, pages 109--131

    2013

  33. [33]

    Geoffrey K Pullum and Barbara C Scholz. 2002. Empirical assessment of stimulus poverty arguments. The Linguistic Review, 19(1-2):9--50

    2002

  34. [34]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9

    2019

  35. [35]

    Caroline F Rowland, Julian M Pine, Elena VM Lieven, and Anna L Theakston. 2003. Determinants of acquisition order in wh-questions: Re-evaluating the role of caregiver speech. Journal of Child Language, 30(3):609--635

    2003

  36. [36]

    Alessandro Sanchez, Stephan C Meylan, Mika Braginsky, Kyle E MacDonald, Daniel Yurovsky, and Michael C Frank. 2019. childes-db: A flexible and reproducible interface to the child language data exchange system. Behavior Research Methods, 51(4):1928--1941

    2019

  37. [37]

    Jon Sprouse, Ivano Caponigro, Ciro Greco, and Carlo Cecchetto. 2016. Experimental syntax and the variation of island effects in English and Italian . Natural Language & Linguistic Theory, 34(1):307--344

    2016

  38. [38]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. https://arxiv.org/abs/2302.13971 LLaMA : Open and Efficient Foundation Language Models . Preprint, arXiv:2302.13971

    Pith/arXiv arXiv 2023

  39. [39]

    Virginia Valian. 1991. https://doi.org/10.1016/0010-0277(91)90046-7 Syntactic subjects in the early speech of American and Italian children . Cognition, 40(1-2):21--81

    work page doi:10.1016/0010-0277(91)90046-7 1991

  40. [40]

    Alex Warstadt, Aaron Mueller, Leshem Choshen, Ethan Wilcox, Chengxu Zhuang, Juan Ciro, Rafael Mosquera, Bhargavi Paranjabe, Adina Williams, Tal Linzen, and Ryan Cotterell. 2023. Findings of the BabyLM challenge: Sample -efficient pretraining on developmentally plausible corpora. In Proceedings of the BabyLM Challenge at the 27th Conference on Computationa...

    2023

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.