Infants use cross-situational statistics to learn action verbs in daily contexts

University Of Texas At Austin, Austin TX

Investigators

Abstract

PROJECT SUMMARY Research on early language acquisition has long focused on how infants learn nouns, but far less is known about how they acquire verbs, which are critical for structuring sentences and conveying relational meaning. Verbs present unique learning challenges because actions are transient, context-dependent, and highly variable across instances. To successfully acquire verbs, infants must resolve two computational problems: 1) word-referent ambiguity—determining which action a verb refers to; and 2) word-meaning ambiguity — inferring the meaning of a verb and generalizing its meaning across different objects and motion variations. Despite these challenges, verbs—particularly action verbs with concrete meanings—are a significant part of early vocabulary. The overarching objective of the present study is to understand how infants learn verbs from everyday contexts with noisy, natural input. The proposed research tests the hypothesis that infants resolve word-referent and word-meaning ambiguity by tracking statistical regularities across verb instances. The project takes a three-pronged approach—naturalistic, experimental, and computational—to determine how the structure of real-world input, and infants’ attentional and cognitive abilities, support verb learning. Aim 1 is to quantify the input statistics that can be used to resolve word- referent ambiguity in one everyday activity – meal preparation. I will analyze how often verbs in parent speech are aligned with their manual actions, and more critically, I will use head-mounted eye tracking to measure whether infants attend to the relevant referent when hearing an action verb. Aim 2 is to test infants’ ability to extract verb meaning via cross-situational learning. I will design and conduct a cross-situational learning experiment by systematically manipulating the co-occurrence statistics of manual actions and target objects across several experimental conditions. Aim 3 is to use computational models to test verb-learning mechanisms. The models will help to determine whether statistical input alone is sufficient for verb learning or whether additional cognitive mechanisms, such as attentional biases, are required. By comparing model performance to infant learning, we will identify which aspects of the learning environment best support generalization. Uncovering the mechanisms that support early verb learning will bridge the gap between controlled experimental findings and real-world language acquisition. By identifying how infants track statistical regularities to resolve word-referent and word-meaning ambiguity, this work will advance theories of word learning and contribute to a broader understanding of how children extract meaning from complex natural input. The findings will also have implications for parenting and behavioral interventions aimed at supporting language development in children at risk for language delays, such as those with developmental disorders.