About this Event
250 Hutchison Rd, Rochester, NY 14620
The ability to efficiently represent space and relations in it is, undoubtedly, one of the crucial faculties of the human brain. Since direct observation of internal spatial representations in the brain is not yet possible, we can try to infer as much as possible by using locative expressions in natural languages as a proxy. In English and many other languages, prepositions form one of the most ambiguous classes of words and have rich spatial senses, thus serving as a dominant class of locative predicates. In the case of spatial prepositions, one would expect them to be defined by geometrical constraints. However, the correct use of spatial prepositions in describing actual object configurations is heavily influenced by non-geometric factors such as argument types and their real-world relations, e.g., whether one argument denotes a proper part of the second argument, what is the standard use of these objects, etc.
In this dissertation, we describe our attempts at modeling spatial relations from a cognitive linguistics-inspired perspective that takes into account not only geometric relations, such as distances, but also important contextual factors, including objects' shapes, sizes, types, and distribution in the scene. We improve our models using empirical data and enable them to be explainable by incorporating rules that correspond to common-sense intuitions about applicability of various spatial predicates. We apply our models to two separate domains, the physical blocks world as well as a set of 3D-modeled "room worlds" and demonstrate that they achieve a good degree of agreement with human judgements in the corresponding domains. In conclusion, we present suggestions for future improvements of our cognitively-inspired approach.
Advisor: Prof. Lenhart Shubert (Computer Science)
Committee: Prof. Dan Gildea (Computer Science), Prof. James Allen (Computer Science),
Prof. Aaron White (Linguistics)
Chair: Prof. Jens Kipper (Philosophy)
0 people are interested in this event