DIP Colloquium

Speaker: Peter Gärdenfors (Lund)
Title: Where Do Semantic Domains Come From
Date:
Time: 11:00 - 12:30
Location: ILLC seminar room F1.15
Children learn a language without effort. A fundamental question is how children know what to learn. When a new word is uttered in a situation, there is often a multitude of objects, features of objects and features of the on-going events that could be the meaning of the word. How does the child select the right meaning? This talk argues that early language learning in children emerges from five primary semantic domains: Space, objects, actions, number and events. These are called ’core knowledge domains’ by Spelke and Carey, but unlike their nativist view, I take an epigenetic position and claim that the domains are learned.  My central thesis is that the conceptual structures that emerge in children’s cognitive development are based on detecting invariants in the sensory input. To some extent, my analysis follows the program of Gibson, although my approach is more cognitively focussed. 
I argue that space, object, action and number domains are very natural outcomes of a reduction of sensory information in terms of invariants. Each of these semantic domains corresponds to a separate set of invariants. For example, our representation of space is characterized by the fact that distances between locations remain invariant. In contrast, objects are can have varying locations in space, but are determined by shape invariants, among other things. Actions are determined by invariants of force patterns, which entails that who performs the action or where it is done is irrelevant. The number of a collection is invariant of the locations and identities of the objects in the collection. Finally, events combine information about agent, patient, action and result that contain information from the other domains. 
The semantic domains are naturally modelled by conceptual spaces. Categories can be described as regions of such spaces. The categories then form the basis for the meanings of early words that children learn. I argue the semantic domains can be used to generate an epigenetic model of the acquisition of word classes. Categories based on the object domain are typically expressed by nouns; categories based on the action domain are expressed by verbs; and relational categories based on the space domain are expressed by prepositions; and categories based on the number domain are expressed by quantifiers and numbers. In this way a semantic foundation for the most common word classes is established. Finally, the event domain forms the meanings of declarative sentences.