Event-Specific Document Ranking Through Multi-stage Query Expansion Using an Event Knowledge Graph

Event-specific document ranking is a crucial task in supporting users when searching for texts covering events such as Brexit or the Olympics. However, the complex nature of events involving multiple aspects like temporal information, location, participants and sub-events poses challenges in effectively modelling their representations for ranking. In this paper, we propose MusQuE (Multi-stage Query Expansion), a multi-stage ranking framework that jointly learns to rank query expansion terms and documents, and in this manner flexibly identifies the optimal combination and number of expansion terms extracted from an event knowledge graph. Experimental results show that MusQuE outperforms state-of-the-art baselines on MS-MARCOEVENT, a new dataset for event-specific document ranking, by $$9.1\%$$ and more.