Improving Passage Retrieval Using Interactive Elicition and Statistical Modeling

摘要

The University of Maryland and Johns Hopkins University worked together in the 2004 High Accuracy Retrieval from Documents (HARD) track to explore design options for interactive passage retrieval systems. HARD assessors responded to clarification forms by (1) selected additional search terms from an automatically constructed list of potentially discriminating terms, (2) selected relevant passages from an automatically constructed list of possibly relevant passages, and (3) entered additional search terms. Query expansion based on these three types of elicited information yielded statistically significant improvements in R-precision over baselines with and without blind relevance feedback. For topics that requested passages as answers, a preliminary analysis shows that statistical models for passage extent trained on HARD 2003 data yielded a significant improvement over a replication of the University of Maryland's hard-2003 technique for passage extent determination, and the results of the new technique appear to generally be well above the median for HARD 2004 systems.