How young children integrate information sources to infer the meaning of words

Abstract

Before formal education begins, children typically acquire a vocabulary of thousands of words. This learning process requires the use of many different information sources in their social environment, including their current state of knowledge and the context in which they hear words used. How is this information integrated? We specify a developmental model according to which children consider information sources in an age-specific way and integrate them via Bayesian inference. This model accurately predicted 2-5-year-old children's word learning across a range of experimental conditions in which they had to integrate three information sources. Model comparison suggests that the central locus of development is an increased sensitivity to individual information sources, rather than changes in integration ability. This work presents a developmental theory of information integration during language learning and illustrates how formal models can be used to make a quantitative test of the predictive and explanatory power of competing theories.

Fig. 1. Experimental task and model.

a,b, Screenshots from the experimental task showing the condition of the experiment in which common ground information is congruent (that is, points to the same object) with speaker informativeness (a) and also showing the incongruent condition (b). The congruent and incongruent conditions are each paired with the 12 known objects, resulting in 24 unique conditions. Steps shown are: the speaker encounters one object and then leaves the scene (1), while the speaker is away (2), a second object appears (3), when returning, the speaker uses a new word to request an object (4). Steps (1) to (3) establish common ground between the speaker and the listener, in that one object is new in context (red). The request in (4) licences an inference based on expectations about how informative speakers are (blue). Listeners’ semantic knowledge enters the task because the identity of the known object on one of the tables is varied from well-known objects such as a duck to relatively unfamiliar objects such as a chess pawn (total of 12 objects, yellow). c–e, Developmental trajectories are shown for sensitivity to common ground (c), speaker informativeness (d) and semantic knowledge (e), estimated on the basis of Experiments 1 and 2 (main text). f, The model equation for the rational-integration model, linking information sources to model parameters.

Fig. 2. Predicting information integration.

a,b, Correlation between model predictions and child inference data for all 24 conditions and for each age group (binned by year) for the rational-integration model (a) and the three lesioned models (b). Horizontal and vertical error bars show 95% HDI. Inset in a shows an example of model predictions as developmental trajectories (Fig. 3). BF₁₀ gives the Bayes factor in favour of the integration model based on the marginal likelihood of the data under each model. c, Predictions from all models considered alongside the data (with 95% HDI) for two experimental conditions (familiar word: “duck”).

Fig. 3. Explaining information integration.

a, Model predictions from the rational-integration model (coloured lines) next to the behavioural data (dotted black lines with 95% CI in grey) for all 24 experimental conditions. Top row shows the congruent condition, while bottom row shows the incongruent condition. Familiar objects are ordered on the basis of their rated age of acquisition (left to right). Light dots represent individual data points. b,c, Correlations between model predictions binned by age and condition for the integration model (b) and the two biased models (c). Vertical and horizontal error bars show 95% HDIs. BF₁₀ gives the Bayes factor in favour of the rational-integration model on the basis of the marginal likelihood of the data under each model. d, Posterior distribution of the bias parameter in the biased-integration model and developmental trajectories for the bias parameter in the developmental-bias model. e, Predictions from all models considered alongside the data (with 95% HDI) for two experimental conditions (familiar word: “duck”).