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 & tables
Construction
Extraction Site
Label
Sample Sentence
Matrix wh-questions
Subject
SMQ
Who praised the student?
Object
OMQ
Who did the professor praise ?
Adjunct
AMQ
When did the professor praise the student ?
Polar
PMQ
Did the professor praise the student?
Plain/Fragment
PlainMQ
Who ?
Cross-clausal
CC_
Who did the student say [that the professor praised ]?
Table 1: Target filler-gap constructions and subtypes by extraction site, with gap positions underlined.
Figure 1: Dependency (top) and constituency (bottom) parses of a sentence containing a subject relative clause.
Figure 2: Dependency parses of an embedded question (left) and a matrix question (right).
Category
Child-directed
Child speech
SMQ
1.00
0.83
OMQ
0.99
0.89
SEQ
0.95
0.95
OEQ
0.94
0.97
SRC
0.98
0.98
SRC reduced
0.98
0.92
Table 2: Human annotated detector accuracy by construction and gap site (out of 100 sentences per category) for child-directed speech and child speech.
Category
Precision (total)
Recall (total)
F1
SMQ
0.827 (712)
0.792 (716)
0.809
OMQ
0.908 (4950)
0.977 (4464)
0.941
AMQ
0.921 (1839)
0.957 (1746)
0.939
SEQ
0.925 (146)
0.894 (236)
0.909
OEQ
0.958 (642)
0.895 (1325)
0.925
AEQ
0.808 (339)
0.905 (924)
0.854
Table 3: Precision (against parser labels), recall (against annotation labels), and F1 scores (bolded if >0.8 ) by construction and extraction site.
Figure 4: Adult (solid) and child (dashed) speech distributions across time by constructions and extraction sites. Uncertainty is shown as 95% Wilson intervals (wider in sparser bins).
Figure 5: Per-1,000 utterance rates of all filler-gap dependency sentences received and produced by Laura.
Figure 6: Subject- versus object-extracted log ratios for each construction across ages, binned by 3 months. Sizes of dot shades indicate number of total utterances within each bin.
Figure 7: Number of utterances with targeted constructions received and produced by Laura in log scale.
Figure 8: Cross-construction comparisons of subject share across age.
Filtered Construction
# Sentences
# Tokens
Matrix questions
69,214
502,555
Embedded questions
3,765
51,930
Relative clauses
5,850
62,637
Table 4: Numbers of tokens and sentences removed from the training corpus for each construction family.
Construction × Extraction Site
Sample Evaluation Pairs
Matrix Question Object Gap
What will you build today You will build today You will build it What will you build it
Matrix Question Subject Gap
Who will chase the doctor will chase the doctor It will chase the doctor Who will it chase the doctor
Embedded Question Object Gap
I knew what you built today I knew that you built today I knew that you built it I knew what you built it
Embedded Question Subject Gap
I knew who chased the doctor I knew that chased the doctor I knew that they chased the doctor I knew who they chased the doctor
Relative Clause Object Gap
I knew the cake that you made I knew that you made I knew that you made it I knew the cake that you made it
Table 5: Example minimal pairs. Gap positions are marked by underline and continuations marked by italics.
Figure 9: Cross-construction evaluations on filtered corpus training results of 15 Llama models. Evaluation accuracy means accuracy on the minimal pair tests.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 10: Dependency parse (top) and constituency parse (bottom) of object matrix question What’s your name?
Pair
Adult Δsubj [95% CI]
Child Δsubj [95% CI]
MatrixQ ∼ EmbQ
[-0.057, 0.013] (n=25)
[-0.032, 0.102] (n=19)
MatrixQ ∼ RC
[-0.321, -0.256] (n=25)
[-0.309, -0.193] (n=20)
EmbQ ∼ RC
[-0.288, -0.239] (n=25)
[-0.322, -0.258] (n=19)
Appendix
Table 6: Across-age summary of Δsubj for each construction pair. Intervals are bootstrap 95% CIs; parentheses give the number of age bins passing the minimum-count filter.
(Hyper)parameter
GPT-2
Llama
Data
Sequence length
128
128
Eval samples
8192
8192
Model
# Parameters
705M
360M
Hidden size
1536
1024
Appendix
Table 7: Summary of configuration and training hyperparameters.
Figure 12: Cross-construction evaluations on filtered corpus training results of 15 GPT-2 models.
Figure 13: A flowchart for the relative clause detection process.
For humans, filler-gap dependencies require a shared representation across different syntactic constructions. Although causal analyses suggest this may also be true for LLMs (Boguraev et al., 2025), it is still unclear if such a representation also exists for language models trained on developmentally feasible quantities of data. We applied Distributed Alignment Search (DAS, Geiger et al. (2024)) to LMs trained on varying amounts of data from the BabyLM challenge (Warstadt et al., 2023), to evaluate whether representations of filler-gap dependencies transfer between wh-questions and topicalization, which greatly vary in terms of their input frequency. Our results suggest shared, yet item-sensitive mechanisms may develop with limited training data. More importantly, LMs still require far more data than humans to learn comparable generalizations, highlighting the need for language-specific biases in models of language acquisition.
CHILDES is a paramount resource for language acquisition studies -- yet computational tools for analyzing its syntactic structure remain limited. Leveraging the recent release of the UD-English-CHILDES treebank with gold-standard Universal Dependencies (UD) annotations, we train a state-of-the-art dependency parser specifically tailored to CHILDES. The parser more accurately captures syntactic patterns in child-adult interactions, outperforming widely used off-the-shelf English parsers, including SpaCy and Stanza. Alongside the parser, we also release a Part-of-Speech tagger and an utterance-level construction tagger, which together form the open-source Syntactic Parsing Toolkit for Child-Adult InTeractions (CAIT). Through a detailed error analysis and a case study tracking the distribution of syntactic constructions across developmental time in CHILDES, we demonstrate the practical utility of the toolkit for large-scale, reproducible research on language acquisition.
Center for Language and Cognition (CLCG), University of Groningen · Georgetown University · Computational Linguistics, Department of Linguistics, Bielefeld University
Automatically distinguishing child-directed speech from adult-directed speech in long-form recordings is key to scalable analyses of children's language environments. Existing approaches process utterances in isolation and have been evaluated primarily on English. We address these gaps along three dimensions. First, we fine-tune and evaluate six-self supervised models on a multilingual dataset of 182 children, showing that in-domain pre-training on child-centered recordings substantially outperforms models trained on adult speech. Second, we demonstrate that incorporating surrounding context substantially improves classification, with an absolute gain of 13.8% in average F1-score. Third, we evaluate our model in a realistic end-to-end pipeline, from adult speech detection to addressee classification, showing that performance drops under automatic segmentation but still consistently outperforms a rule-based baseline.
Théo Charlot, Tarek Kunze, Kaveri K. Sheth +2
LSCP, DEC, ENS, EHESS, CNRS, PSL University, France · Laboratoire d’Informatique et Syst`emes, Universit´e Aix-Marseille, CNRS, France