Learning term weights by overfitting pairwise ranking loss

A search engine strikes a balance between effectiveness and efficiency to retrieve the best documents in scalable way. Recent deep learning-based ranker methods are proving be effective improving state-of-the-art relevancy metrics. However, as opposed index-based retrieval methods, neural rankers like bidirectional encoder representations from transformers (BERT) do not scale large datasets. In this article, we propose query term weighting method that can used with standard inverted index without modifying it. Query weights learned using relevant irrelevant document pairs for each query, pairwise ranking loss. The prove more than recall which is probabilistic relevance feedback, previously task. We further show these predicted BERT regression model improve performance of both BM25 based an already optimized function.