To learn verbs, children must solve several inference problems simultaneously. On hearing a verb used for the first time (break), they must decide which of any number of events that the speaker could be talking about the speaker is actually talking about (Sally running around the room, Sally breaking a vase, Sally's mother yelling angrily, etc.). They must also infer the range of events that the verb can be used to talk about (any time Sally breaks something, any time a vase breaks, any time anyone breaks anything, etc.). Finally, they must infer the mapping between the participants in the event (Sally, the vase) and grammatical position in the sentence (subject, object): that is, does Sally broke the vase mean that the vase ended up broken or that Sally did? Although children do apparently learn the correct generalizations, researchers have had more difficulty identifying them, particularly with regards to the mapping patterns (the final inference listed above). A number of theories have been suggested as to the nature of the mapping patterns and how they and the other inferences might be learned. However, these theories are complex, and it has proven difficult to directly compare them. In this project, we develop formal computational models of these learning processes based on different theories proposed in the literature. By building computational models, we can empirically determine what the predictions of each theory would be under various assumptions, allowing us to directly compare the accuracy of different theories. We propose three modeling projects. In the first project, we model only the first two inferences: which event does a verb refer to in a specific instance, and what range of events can be referred to. In the second, we model only the third: identifying the mapping pattern for the verb. The third project combines the first two: the models attempt to learn all three inferences simultaneously. This type of joint inference problem has been the focus of a considerable amount of computational modeling research in recent years, particularly within the Bayesian modeling framework, and there are now extremely powerful computational techniques that can be applied to the problem. These models will provide precision in an area of language acquisition that has long been central to the field but which has been in need of more clarity in terms of how theory and data relate. By helping further characterize normal language acquisition, this clarity should lead to more precise diagnostics and treatment interventions for disorders of language development and linguistically related cognitive development.
The proposed project attempts to develop explicit computational models of verb learning in naturalistic conditions, exploring and comparing different theories for how the meanings of verbs are represented. These models will provide precision in an area of language acquisition that has long been central to the field but which has been in need of more clarity in terms of how theory and data relate. This clarity should lead to more precise diagnostics and treatment interventions for disorders of language development and linguistically related cognitive development.