A Bayesian Computational Model of Cross-Situational Word Learning with Phonetic Constraints

Abstract

Young children solve the referential ambiguity problem in word learning through cross-situational statistics, yet computational models often ignore phonetic information. We present a Bayesian hierarchical model that integrates phonetic similarity (generalised linear mixed effects on phoneme inventory) with cross-situational co-occurrence matrices. Fitted to a corpus of naturalistic parent-child interactions (N = 18 dyads, 25,000+ word tokens), the model predicts child vocabulary growth and word learning errors with significantly higher accuracy than phonetic-only or statistics-only baselines (McFadden's pseudo-R² = .68 vs. .51–.53). Simulations suggest phonetic constraints enable learners to rapidly rule out implausible referents, reducing the sample complexity problem by approximately 40%.

Introduction

The logical problem of word learning is well-known: in a given moment, a child hears a word and observes multiple potential referents, yet must somehow assign meaning to that word despite this pervasive ambiguity (Quine, 1960). Over the past two decades, psycholinguists have demonstrated that children solve this problem in part by tracking the co-occurrence statistics across situations: if a word appears repeatedly when a particular object is present, the child's cross-situational learner hypothesis is reinforced (Smith & Yu, 2008). However, most computational models of cross-situational learning operate over discrete objects or visual features without considering the linguistic signal itself. This is a curious omission, because children's own phonetic inventories, preferences, and constraints likely influence word recognition and learning.

The goal of the present work is to develop and test a unified Bayesian model that treats phonetic information and cross-situational statistics as complementary constraints on word learning. Such a model should be better equipped to explain both the trajectory of vocabulary growth and the systematic errors children make during learning. We hypothesize that phonetic similarity between a novel word and known words constrains the candidate set of possible referents, thereby reducing the sample complexity burden on the learner and facilitating faster convergence to correct mappings.

Method

Participants and Corpus

Eighteen parent-child dyads (child ages 18–36 months; M = 26.5 months, SD = 7.2 months) participated in this longitudinal naturalistic observation study. Families were recruited through community childcare centres in Toronto and gave written informed consent. Weekly 90-minute video-recorded play sessions were collected over 6 months, yielding approximately 25,000 word tokens and 312 unique word types in child-directed speech. Maternal socioeconomic status (via parental education and occupational codes) was balanced across dyads. All families were English-speaking. The study was approved by the University of Toronto Research Ethics Board. Data will be made available on the Open Science Framework upon publication of this manuscript.

Computational Model

We developed a hierarchical Bayesian model formalised as a dynamic topic model (Blei & Lafferty, 2006) with phonetic priors. At each time point t, the model maintained a joint distribution P(referent | word, phonetic similarity, co-occurrence history). Phonetic similarity was computed as the Levenshtein distance between the novel word's phoneme sequence and the phoneme sequences of all known words in the child's input, weighted by word frequency. Cross-situational statistics were derived from manual annotations of objects present in each interaction, coded by two independent coders (inter-rater agreement Cohen's κ = .87 for object presence; κ = .79 for object identity). The model was implemented in Stan (Gelman et al., 2013) using Hamiltonian Monte Carlo sampling. Hyperpriors were set to weakly informative defaults and sensitivity analyses confirmed robustness to hyperparameter choices. Model selection was conducted using leave-one-out cross-validation (Vehtari, Gelman, & Gabry, 2015).

Analysis

The model was fitted to the corpus with half the data (50% of word tokens from each dyad) and evaluated on held-out test data. Predictive accuracy was assessed using McFadden's pseudo-R² for the multiclass categorical distribution of referents. We compared three models: (1) a full model integrating phonetic and cross-situational statistics, (2) a phonetic-only model, and (3) a cross-situational-only model. Additionally, we simulated learning trajectories by iteratively presenting the model with word-referent pairs drawn from the corpus and measuring convergence time and learning errors. Simulations examined how much phonetic information reduces the need for observations to learn mappings (sample complexity).

Results

The full model significantly outperformed both baselines in held-out prediction: McFadden's pseudo-R² = .68 (95% CI [.65, .71]) for the full model, .53 (95% CI [.50, .56]) for the phonetic-only model, and .51 (95% CI [.48, .54]) for the cross-situational-only model. The difference between full and phonetic-only models was substantial (ΔR² = .15, corresponding to a 28% reduction in error). Likelihood ratio tests confirmed that the interaction terms between phonetic and cross-situational predictors were significant (χ²(4) = 18.7, p < .001). In simulation, children equipped with phonetic constraints achieved 80% accuracy in referent identification with a median of 3.2 exposures per word (IQR: 2–5), whereas the cross-situational-only model required a median of 5.4 exposures (IQR: 4–8), representing a 40% reduction in sample complexity. The model successfully predicted 76% of vocabulary composition errors observed in the children's speech (e.g., over-extensions, under-extensions), with 89% specificity.

Discussion

These results demonstrate that a unified Bayesian framework capturing both phonetic and cross-situational constraints provides a powerful account of word learning in naturalistic settings. The substantial improvement in predictive accuracy over single-constraint models suggests that children are indeed sensitive to phonetic similarity and use it to constrain referential ambiguity. The simulation finding that phonetic information reduces sample complexity by 40% has important implications for understanding how learners bootstrap language without requiring implausibly large numbers of exposures—a longstanding puzzle in language acquisition (Bloom, 2000).

The phonetic contribution to word learning likely arises through multiple mechanisms: phonotactic constraints based on the child's native language phonology (Nazzi, Iakimova, Perrier, Métivier, & Loevenbruck, 2011), statistical regularities in the phonetic forms of object categories within the input (e.g., diminutives, onomatopoeia), and possibly implicit reasoning about the structure of the lexicon itself. Future work should investigate how this model scales to larger vocabularies, whether it captures individual differences in learning trajectories, and whether it predicts atypical development in children with language delay.

References

Blei, D. M., & Lafferty, J. D. (2006). Dynamic topic models. Proceedings of the 23rd International Conference on Machine Learning, 113–120.
Bloom, P. (2000). How children learn the meanings of words. MIT Press.
Gelman, A., Carlin, J. B., Stern, H. S., & Rubin, D. B. (2013). Bayesian data analysis (3rd ed.). Chapman and Hall/CRC.
Nazzi, T., Iakimova, G., Perrier, P., Métivier, V., & Loevenbruck, H. (2011). Phoneme inventories and phonotactic constraints in infants' first words. Developmental Science, 14(6), 1446–1456.
Quine, W. V. O. (1960). Word and object. MIT Press.
Smith, L. B., & Yu, C. (2008). Infants rapidly learn word-referent mappings via cross-situational statistics. Cognition, 106(3), 1558–1568.
Vehtari, A., Gelman, A., & Gabry, J. (2015). Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC. Statistics and Computing, 27(5), 1413–1432.