UMass Machine Learning and Friends Lunch | Main / Parameterized Concept Weighting For Information Retrieval

Abstract: The majority of the current information retrieval models weight the query concepts (e.g., terms or phrases) in an unsupervised manner, based solely on the collection statistics. In this talk, I will present a parameterized concept weighting model that goes beyond the unsupervised estimation of concept weights. In this model, the weight of each query concept is determined using a parameterized combination of diverse sources of evidence. In addition, I will show that the parameterized concept weighting model can be further extended to incorporate expansion concepts, as well as to model dependencies between query concepts using a hypergraph structure.

Bio: Michael Bendersky is a PhD candidate at the Center for Intelligent Information Retrieval. His current research combines insights from information retrieval, natural language processing and statistical machine learning to improve the effectiveness of search with complex natural language queries. He is excited about this research, since it has the potential to revolutionize the way people search on the web, in the enterprise and on mobile devices, to name just a few possible applications.