Principled Optimization Frameworks for Query Reformulation of Database Queries

Abstract

Traditional databases have traditionally supported the Boolean retrieval model, where a query returns all tuples that match the selection conditions specified -- no more and no less. Such a query model is often inconvenient for naive users conducting searches that are often exploratory in nature, since the user may not have a complete idea, or a firm opinion of what she may be looking for. This is especially relevant in the context of the Deep Web, which offers a plethora of searchable data sources such as electronic products, transportation choices, apparel, investment options, etc. Users often encounter two types of problems: (a) they may under-specify the items of interest, and find too many items satisfying the given conditions (the many answers problem), or (b) they may over-specify the items of interest, and find no item in the source satisfying all the provided conditions (the empty answer problem). In this talk, I discuss our recent efforts in developing techniques for iterative "query reformulation" by which the system guides the user in a systematic way through several small steps, where each step suggests slight query modifications, until the query reaches a form that generates desirable answers. Our proposed approaches for suggesting query reformulations are driven by novel probabilistic frameworks based on optimizing a wide variety of application-dependent objective functions.