Wednesday, 6th September - 4pm Doug Oard : Seminar

Title:  Pareto Optimality for Neural Cross-Language Information Retrieval

Abstract:

Information retrieval systems have always needed to balance effectiveness with efficiency, but the computational complexity of highly effective neural ranking models has expanded the dynamic range of that trade space.  Cross-Language Information Retrieval (CLIR) further extends that dynamic range, since techniques for transcending language barriers may be more or less computationally complex, and some of that that additional complexity can be allocated between training time, indexing time and query time.  In this talk, I will introduce Pareto analysis as a way of thinking about the nature of this trade space, and then I will illustrate the value of that perspective using two examples.  In the first, I will use a Pareto analysis of effectiveness with query latency to explore the design space for an alternative to the widely used MaxP heuristic for estimating document relevance from passage relevance.  We call this alternative CREPE, the basic idea of which is to use guided rather than exhaustive search to select the best-scoring passage.  In CLIR, the guided search also incurs translation costs, which thus need to be factored into the analysis.  As my second example, I will use a Pareto analysis of efficiency and indexing latency to explore the design space for SPLADE-like neural term expansion techniques.  In CLIR applications, performing neural document translation to convert the task to monolingual retrieval can be a hard baseline to beat, but Pareto analysis shows that such an approach occupies just one part of a far more extensive Pareto-optimal design space. The ideas and results I will present in this talk are the result of joint work with Dawn Lawrie, Jim Mayfield, Paul McNamee, Suraj Nair and Eugene Yang.

Bio:

Douglas W. Oard is a Professor at the University of Maryland, College Park (USA), with joint appointments in the College of Information Studies and the University of Maryland

Institute for Advanced Computer Studies, am Affiliate Professor at the Johns Hopkins University Human Language Technology Center of Excellence (USA), and a Visiting Professor at the National Institute of Informatics (Japan).  With a Ph.D. in Electrical Engineering, his research interests center on the use of emerging technologies to support information seeking, and he is perhaps best known for his work on cross-language information retrieval. Additional information is available at http://terpconnect.umd.edu/~oard/.

 vCal  iCal