Structural Priors for Sample-Efficient Grammar Learning: A Neurosymbolic Approach

This study investigates whether explicit structural grammatical priors improve neural language model learning from child-scale linguistic input. We combine a BabyBERTa-style masked language model with a differentiable probabilistic context-free grammar (PCFG) inspired by Minimalist Grammar, using a "grammar-as-auxiliary-loss" architecture: L = L_MLM + λ·L_grammar, where L_grammar = −log P(sentence | grammar) is computed via a differentiable CKY inside algorithm. The model is trained on the AO-CHILDES corpus (~5M words from child-directed speech, ages 1–6) and evaluated on the Zorro grammaticality benchmark (23 paradigms). We test four hypotheses: (H1) the grammar-constrained model outperforms the BabyBERTa baseline on Zorro by ≥3 percentage points or p < 0.05; (H2) the constrained model trained on 50% of data matches baseline performance at 100%; (H3) improvement localizes to phenomena explicitly encoded in the grammar rather than reflecting general regularization; and (H3b) hand-crafted linguistic constraints outperform data-derived statistical grammars. Exploratory analyses examine whether the constrained model's acquisition trajectory correlates with documented child developmental order (H4) and whether its errors resemble attested child language errors (H5). Controls include random grammar baselines (matched structure, shuffled probabilities), frequency-based constraints, and data-derived PCFGs induced from CHILDES. All conditions are run across 5 random seeds with Bonferroni-corrected lambda search over 7 values. We commit to reporting all results regardless of outcome.