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

The paper addresses how young children infer the meanings of new words by integrating multiple information sources available in social contexts. The core research question is how children arbitrate among cues such as assumptions about cooperative, informative speech, shared common ground, and their own semantic knowledge to identify a speaker’s intended referent. Prior accounts have described separate mechanisms (a ‘bag of tricks’) without specifying how they are integrated, leading to a lack of explicit, quantitative theories of integration. The authors propose and test a formal, developmental cognitive model in which children integrate information via Bayesian inference, with development reflected in increasing sensitivity to individual sources while the integration mechanism remains stable.

Prior work shows that children use multiple social-contextual cues from early in development, including expectations that speakers use new words for unknown objects (mutual exclusivity), talk about relevant or new things in context, and maintain discourse continuity. Classic paradigms (e.g., the 2×2 shelf task) reveal that preschoolers integrate utterance semantics with a speaker’s visual perspective, but these studies have not specified the computational process of integration. Social-pragmatic theories assume information is integrated via social inference, yet without quantitative formalization. The authors build on this literature by formalizing integration within a pragmatic Bayesian framework that captures informativeness, common ground, and semantic knowledge, extending related work in adult pragmatic inference to developmental word learning.

Design and model: The authors formalize a rational-integration model based on Bayesian inference in which a listener (L1) infers the referent of a speaker’s (S1) utterance using three information sources: (1) expectations that speakers are cooperative and informative (speaker informativeness, parameter α), (2) common ground established in the conversation (prior over referents, parameter ρ or p), and (3) the child’s semantic knowledge of familiar word-object mappings (object-specific knowledge, parameters θj). The model assumes a rational speaker chooses informative utterances given a literal listener and the listener integrates likelihood (informativeness and semantics) with priors (common ground). Development is captured by age-sensitive parameters for α, ρ, and θ. Alternative models include ‘lesioned’ variants that selectively ignore one source (no-word-knowledge; no-common-ground; no-speaker-informativeness) and ‘biased’ integration models that compute separate inferences and combine them via a weighting parameter φ, with a developmental variant allowing φ to vary with age.

Experimental paradigm: Three experiments using an interactive tablet-based storybook task with animal speakers who request objects using a nonce word. Children select the referent, operationalized as choosing the novel object when appropriate. Stimuli included pairs of familiar objects (varying in age-of-acquisition familiarity from items like “duck” to low-familiarity items like a chess pawn) and unfamiliar novel objects. Common ground was manipulated by making one object new to the speaker in the discourse context.

Participants: Total N=368 across experiments. Experiment 1: n=90 (30 2-year-olds, 30 3-year-olds, 30 4-year-olds). Experiment 2: n=58 (18 2-year-olds, 19 3-year-olds, 21 4-year-olds). Experiment 3: n=220 (76 2-year-olds, 72 3-year-olds, 72 4-year-olds). Children were recruited at a museum in San José, CA; English exposure and completion criteria applied.

Procedures:

• Experiment 1 (mutual exclusivity/informativeness and semantic knowledge): On each trial, one familiar and one unfamiliar object were shown; the speaker used a novel word to request an object. Children completed 12 trials with 12 different familiar objects spanning a range of expected familiarity. Correct choice was selecting the novel object. These data estimated age-sensitive α and object-specific θj.
• Experiment 2 (common ground sensitivity): The speaker first observed one table (object present or absent) and left; while away, a second unfamiliar object appeared. Upon return, the speaker requested an object with a novel word. Correct choice was selecting the object that was new to the speaker. Ten trials with unfamiliar object pairs. These data estimated ρ (age-sensitive common ground prior).
• Experiment 3 (integration test): Combined Exp. 1 and 2 manipulations with familiar vs. unfamiliar objects and discourse novelty, creating 24 conditions: 12 familiar objects × two alignment conditions (congruent: both cues point to novel object; incongruent: cues conflict). Each child received up to 12 trials (six per alignment), counterbalanced.

Model fitting and prediction: Parameters for α (linear regression submodel), θ (logistic regression with object-level variation), and ρ (logistic regression) were estimated from Experiments 1–2 to generate parameter-free a priori predictions for Experiment 3 (prediction phase). For explanatory comparisons, parameters were additionally constrained using Experiment 3 (fully Bayesian estimation). Models were implemented in WebPPL; analyses used Pearson correlations between predictions and binned data and marginal likelihood-based Bayes Factors for model comparison.

• Predictive accuracy: Using parameters estimated only from Experiments 1–2, the rational-integration model’s a priori predictions closely matched Experiment 3 behavior across ages 2–5, explaining 79% of the variance when predictions and data were binned by age (years) and condition.
• Children used all information sources: Lesioned models that ignored one source performed substantially worse. Bayes Factors favored the rational-integration model over each heuristic alternative by several orders of magnitude: vs. no-word-knowledge BF10 = 3.9 × 10^9; vs. no-common-ground BF10 = 2.6 × 10^10; vs. no-speaker-informativeness BF10 = 4.8 × 10^110. Heuristic models systematically underestimated performance in congruent conditions and showed characteristic mispredictions in incongruent conditions.
• Explanatory comparisons: When all data constrained parameters, the rational-integration model explained 87% of variance and outperformed biased-integration alternatives. Against a biased-integration model (φ fixed), BF10 = 2.1 × 10^10; biased model explained 78% variance with φ MAP ≈ 0.65 (95% HDI: 0.60–0.71), over-weighting mutual exclusivity and failing when that inference was weak. A developmental-bias model (φ varies with age) also explained 78% but was decisively worse than rational integration (BF10 = 1.4 × 10^6), tending to underestimate performance due to constrained interplay among inferences.
• Developmental locus: Evidence supports increasing sensitivity to individual information sources with age while the integration mechanism remains constant, indicating developmental continuity in integration from ages 2 to 5.

Findings demonstrate that even very young children integrate multiple social and lexical information sources to infer novel word meanings, and that a unified Bayesian pragmatic model captures this process quantitatively. The rational-integration model, derived from the Rational Speech Acts framework, generalized across conditions where cues aligned and conflicted, outperforming models that ignored cues or combined them modularly with ad-hoc weights. This supports the view that children’s word learning involves integrated social inference rather than isolated ‘bag of tricks’ mechanisms. Importantly, development appears to reflect heightened sensitivity to informativeness, common ground, and semantic knowledge rather than a qualitative change in how cues are combined. The framework connects developmental pragmatics with broader theories of rational social action and suggests natural extensions to additional cues (e.g., gaze, gesture) and contexts.

The paper contributes a formal, developmental theory of how children integrate informativeness, common ground, and semantic knowledge during word learning via Bayesian inference. Empirically, a tablet-based paradigm manipulating these sources showed that the rational-integration model makes accurate quantitative predictions across ages 2–5 and explains behavior better than lesioned or biased-integration alternatives. The work illustrates how computational cognitive models can test and compare psychological theories. Future directions include: (1) cross-cultural and cross-linguistic studies to assess variability in developmental sensitivities while testing for a shared integration architecture; (2) deeper modeling of how common ground expectations are computed over discourse and differentiated from other sources of contextual salience; and (3) linking momentary referent identification to learning processes unfolding over longer timescales, integrating cognitive and cross-situational mechanisms.

• Population and generalizability: Participants were recruited from a single museum population with high parental education/SES; results may not generalize across cultures and languages. The authors anticipate substantial cross-cultural variability in developmental trajectories.
• Common ground representation: The model treats common ground as a conversational prior without specifying the cognitive processes by which these priors are formed or distinguishing them from other contextual salience factors.
• Computational-level abstraction: Parameters (α, ρ, θ) are unidimensional summaries of complex processes; the model does not claim psychological reality of these parameters or specify underlying mechanisms.
• Scope: The model targets in-the-moment referent identification; it does not directly model how informativeness expectations or lexical knowledge evolve across interactions and developmental timescales.