Multiple Linear Combination Approaches for Information Search in Ranking

Since the well-known BM25 [1] was proposed, BM25 and its enhanced version [2] – [4] have dominated the document/passage ranking task for a long time. However, with the advent of deep learning models like BERT [5] , these pre-trained models have achieved noticeable progress in various information retrieval (IR) tasks. But, as BM25 is a "bag of words" retrieval method by matching keywords, it remains a better option for passage ranking in some exceptional cases, like identifying names [6] . Therefore, fusing BM25 with deep learning models is a natural idea to benefit the ranking results. This paper discusses various linear methods of combing BM25 with BERT to see how they affect the final results of the models. We conduct experiments on the MS MARCO V2 dataset, which show convincing results.