TILDE: Term Independent Likelihood moDEl for Passage Re-ranking
Shengyao Zhuang, Guido Zuccon · 2021
Deep language models (deep LMs) are increasingly being used for full text retrieval or within cascade retrieval pipelines as later-stage re-rankers. A problem with using deep LMs is that, at query time, a slow inference step needs to be performed -- this hinders the practical adoption of these powerful retrieval models, or limits sensibly how many documents can be considered for re-ranking.