On the Separation of Logical and Physical Ranking Models for Text Retrieval Applications


Text retrieval using bags of words is typically formulated as inner products between vector representations of queries and documents, realized in query evaluation algorithms that traverse postings in an inverted index. Viewed in database terms, this captures a tight coupling between the ‘logical’ aspects of ranking (i.e., term weighting) and the ‘physical’ aspects of ranking (query evaluation). We argue that explicitly decoupling these two aspects offers a framework for thinking about the relationship between sparse retrieval techniques and the rapidly growing literature on dense retrieval techniques.

In Proceedings of DESIRES 2021