Is Child-Directed Language Optimized for Word Learning? A Computational Study of Verb Meaning Acquisition

Technical Summary: Is Child-Directed Language Optimized for Word Learning? A Computational Study of Verb Meaning Acquisition

Problem Statement
The study investigates whether Child-Directed Language (CDL) is structurally optimized to facilitate language learning, specifically the acquisition of verb meanings. While CDL is traditionally viewed as uniquely conducive to learning, recent critiques suggest its benefits may be limited to specific features (e.g., prosody) and may not generalize to morphosyntax or semantics. Furthermore, existing computational studies have yielded conflicting results: some suggest CDL supports syntax-independent semantic inference (You et al., 2021), while others indicate that syntactic bootstrapping remains critical even in CDL (Zhu et al., 2025). A key gap exists in understanding whether the robustness of verb learning to syntactic degradation is a unique property of CDL or a broader characteristic of the spoken register compared to written language.

Methodology
The authors employ a controlled computational approach using autoregressive Transformer language models (GPT-2 architecture: 12 layers, 12 heads, 768 hidden units) trained on four distinct corpora:

1. CDL: CHILDES transcripts (BabyLM Challenge subset).
2. Spoken ADL: A composite of CANDOR, Switchboard, and spoken BNC.
3. Written ADL (BNC): The written portion of the British National Corpus.
4. Written ADL (Wikipedia): Encyclopedic prose.

To isolate the contributions of syntax and lexical co-occurrence, the authors apply two targeted perturbations to the training data, following the framework of Zhu et al. (2025):

• REPLACE.WORD (Lexical Disruption): Nouns, adjectives, and adverbs co-occurring with verbs are replaced with words of the same Part-of-Speech and fine-grained morphological tag, preserving word order but disrupting verb-specific distributional cues.
• SHUFFLE.ORDER (Syntactic Disruption): Word order within sentences is randomized, destroying syntactic structure while retaining sentence-level co-occurrence statistics.

Evaluation Metrics
Verb meaning acquisition and syntactic proficiency are evaluated using minimal pair paradigms:

• Semantic Minimal Pairs: In-domain pairs where the target verb is substituted with a semantically mismatched alternative from the same frequency bin. Accuracy is measured by the model's ability to assign higher probability to the correct verb.
• Syntactic Minimal Pairs: The FIT-CLAMS benchmark (adapted for all four corpora) tests subject-verb number agreement.

The study addresses three research questions (RQs) regarding the differential impact of these perturbations and the developmental trajectories of semantic versus syntactic learning.

Key Results

• RQ1 (Relative Impact of Perturbations): Across all four domains (CDL, Spoken ADL, and Written ADL), disrupting syntactic structure (SHUFFLE.ORDER) causes a significantly larger drop in verb learning accuracy than disrupting lexical co-occurrences (REPLACE.WORD). This confirms that syntactic cues serve as a universal scaffold for verb learning, regardless of the input domain.
• RQ2 (Robustness to Syntax Ablation): While CDL models show high resilience to syntactic disruption, this resilience is not unique to CDL. Models trained on Spoken ADL (CANDOR) exhibit comparable robustness to CDL. In contrast, models trained on Written ADL (BNC and Wikipedia) show a significantly larger performance decrement when syntax is ablated. This suggests that the advantage lies in the spoken vs. written register distinction rather than a specific CDL vs. ADL optimization.
• RQ3 (Developmental Trajectories): Tracking performance over 20 training epochs reveals a "semantic-first" trajectory across all domains, where semantic accuracy rises before syntactic proficiency. However, the degree of asynchrony varies:
  • CDL maintains the highest and most persistent ratio of semantic-to-syntactic knowledge throughout training.
  • Written ADL shows a rapid "catch-up" of syntactic skills, leading to a tightly coupled development of semantics and syntax.
  • Spoken ADL occupies an intermediate position, initially resembling CDL but eventually converging closer to written patterns, though retaining a distinct spoken-domain profile.

Significance and Claims
The paper argues that the previously attributed advantage of CDL for verb learning may not stem from a unique optimization of child-directed speech, but rather reflects broader properties of the spoken register. The findings challenge the notion that CDL enables syntax-independent semantic inference in a way that ADL does not; instead, both spoken CDL and spoken ADL support more robust semantic representations under syntactic degradation compared to written input.

The study concludes that while syntax is a universal scaffold for verb learning, conversational input (whether child- or adult-directed) allows for more robust semantic acquisition even when syntactic knowledge is limited. The authors caution that these conclusions are specific to English and rely on a coarse-grained operationalization of syntax (word order), suggesting that future work should investigate finer-grained syntactic frames and typologically diverse languages.