Context-Dependent Conceptualization

Conceptualization seeks to map a short text to a set of concepts as a mechanism of understanding text. Most of prior research in conceptualization uses human-crafted knowledge bases that map instances to concepts. Such approaches to conceptualization have the limitation that the mappings are not context sensitive. To overcome this limitation, we propose a framework in which we harness the power of a probabilistic topic model which inherently captures the semantic relations between words. By combining latent Dirichlet allocation, a widely used topic model with Probase, a large-scale probabilistic knowledge base, we develop a corpus-based framework for context-dependent conceptualization. Through this simple but powerful framework, we improve conceptualization and enable a wide range of applications that rely on semantic understanding of short texts, including frame element prediction, word similarity in context, ad-query similarity, and query similarity.