PARADE: Passage Representation Aggregation for Document Reranking

Pre-trained transformer models, such as BERT and T5, have shown to be highly effective at ad-hoc passage document ranking. Due the inherent sequence length limits of these they need process passages one a time rather than processing entire once. Although several approaches for aggregating passage-level signals into document-level relevance score been proposed, there has yet an extensive comparison techniques. In this work, we explore strategies from document’s final ranking score. We find that representation aggregation techniques can significantly improve over proposed in prior taking maximum call new approach PARADE. particular, PARADE results on collections with broad information needs where spread throughout (such TREC Robust04 GOV2). Meanwhile, less complex may work better often pinpointed single DL Genomics). also conduct efficiency analyses highlight improving transformer-based aggregation.