A novel multi-parameter tuned optimizer for information retrieval based on particle swarm optimization

Abstract

Tuning multi-parameter and parameter optimization in Information Retrieval has been a huge area of research and development, especially with BM25F scoring functions having a 2F+1 feature with F fields in the documents. The scoring and ranking function conventionally uses multiple input parameters, to augment the quality of results even at the value of huge calculation time. The searching and ranking documents in the medical literature encompass high recall rates, which are difficult to satisfy with multiple input parameters. The performance of the BM25F depends upon the choice of these F parameters. Particle Swarm Optimization (PSO) searches through the solution-space independently and discovers an optimal solution as opposed to improving and optimizing the gradient; henceforth it can straightforward optimize Mean Average Precision (MAP) a non-differentiable function. In this paper, the usage of PSO to tune multi-parameters is proposed to deal with the gaps in BM25Fscoring function. Also, the advantage of the proposed technique by directly optimizing the MAP has been discussed. Experimental results of quantitative performance metrics MAP and Mean Reciprocal Rank of the proposed PSO-optimized BM25F and most recent ranking algorithms have been compared. The performance measure results demonstrate that the proposed PSO-optimized BM25F performance measure outclasses the standard ranking methods for the OHSUMED data set.