Retrieval parameter optimization using genetic algorithms

摘要

This paper describes our experiments on automatic parameter optimization for the Japanese monolingual retrieval task. Unlike regression approaches, we optimized parameters completely independently of retrieval models enabling the optimized parameter set to illustrate the characteristics of the target test collections. We adopted genetic algorithms as optimization tools and cross-validated with four test collections, namely the CLIR-J-J collections for NTCIR-3 to NTCIR-6. The most difficult retrieval parameters to optimize are the feedback parameters, because there are no principles for calibrating them. Our approach optimized feedback parameters and basic scoring parameters at the same time. Using test sets and validation sets, we achieved effectiveness levels comparable with very strong baselines, i.e., the best-performing NTCIR official runs.