Rank-at-a-Time Query Processing
Ahmed Elbagoury, Matt Crane, Jimmy Lin · 2016
Query processing strategies for ranked retrieval have been studied for decades. In this paper we propose a new strategy, which we call rank-at-a-time query processing, that evaluates documents in descending order of quantized scores and is able to directly compute the final document ranking via a sequence of boolean intersections. We show that such a strategy is equivalent to a second-order restricted composition of per-term scores. Rank-at-a-time query processing has the advantage that it is anytime score-safe, which means that the retrieval algorithm can self-adapt to produce an exact ranking given an arbitrary latency constraint. Due to the combinatorial nature of compositions, however, a naive implementation is too slow to be of practical use. To address this issue, we introduce a hybrid variant that is able to reduce query latency to a point that is on par with state-of-the-art retrieval engines.