Multitask Fine-Tuning for Passage Re-Ranking Using BM25 and Pseudo Relevance Feedback

Passage re-ranking is a machine learning task that estimates relevance scores between given query and candidate passages. Keyword features based on the lexical similarities queries passages have been traditionally used for passage models. However, such approaches limitation; it difficult to find semantic contextual beyond word-matching information. Recently, several studies neural pre-trained language models as BERT overcome limitations of traditional keyword-based they show significant performance improvements. Such ranking advantage finding documents better than methods. these deep learning-based require large amounts data training. training usually manually labeled with high cost, how utilize efficiently an important issue. This paper proposes fine-tuning method efficient model. The proposed model utilizes augmentation by simultaneously MLM tasks during stages. For task, different parts are masked at each epoch. Even if only one pair given, exposed diverse cases dynamically from one. Also, probability distribution term importance trained model.We calculate weight two novel methods using BM25 pseudo feedback. Terms sampled according weight. learns representation executing task. A feedback applied calculating importance. It enables form feedbacks initial search stage. MS MARCO leaderboard Our achieves state-of-the-art MRR@10 score in except ensemble-based method. In addition, our demonstrates three evaluation metrics: MRR@10, Mean Rank, Hit@(5,10,20,50).