Expansion via Prediction of Importance with Contextualization

• DOI: 10.1145/3397271.3401262
• Publication Date: 2020-07-25T07:50:08Z
• Authors: Sean MacAvaney; Franco Maria Nardini; Raffaele Perego; Nicola Tonellotto; Nazli Goharian; Ophir Frieder
• Keywords: Contextualization; Pruning

The identification of relevance with little textual context is a primary challenge in passage retrieval. We address this problem representation-based ranking approach that: (1) explicitly models the importance each term using contextualized language model; (2) performs expansion by propagating to similar terms; and (3) grounds representations lexicon, making them interpretable. Passage can be pre-computed at index time reduce query-time latency. call our EPIC (Expansion via Prediction Importance Contextualization). show that significantly outperforms prior importance-modeling document approaches. also observe performance additive current leading first-stage retrieval methods, further narrowing gap between inexpensive cost-prohibitive Specifically, achieves [email protected] 0.304 on MS-MARCO dataset 78ms average query latency commodity hardware. find reduced 68ms pruning representations, virtually no difference effectiveness.