bigIR at TREC 2020: Simple but Deep Retrieval of Passages and Documents

Abstract

In this paper, we present the participation of the bigIR team at Qatar University in the TREC Deep Learning 2020 track. We participated in both document and passage retrieval tasks, and each of its subtasks, full ranking and reranking. As it is our first participation in the track, our primary goal is to experiment with the latest approaches and pre-trained models for both tasks. We used Anserini IR toolkit for indexing and retrieval, and experimented with different techniques for passage expansion and reranking, which are either BERT-based or sequence-to-sequence based. All our submitted runs for the passage retrieval task, and most of our submitted runs for the document retrieval task outperformed TREC median submission. We observed that BERT reranker performed slightly better than T5 reranker when expanding passages with sequence-to-sequence based models. However, T5 achieved better results than BERT when passages were expanded with DeepCT, a BERT-based model. Moreover, the results showed that combining the title and the head segment as document representation for reranking yielded significant improvement over each separately.