无监督re-ranker,Improving Passage Retrieval with Zero-Shot Question Generation

论文：Improving Passage Retrieval with Zero-Shot Question Generation

作者：Devendra Singh Sachan1,2∗, Mike Lewis3, Mandar Joshi4, Armen Aghajanyan3, Wen-tau Yih3, Joelle Pineau1,2,3, Luke Zettlemoyer3,4 1McGill University; 2Mila - Quebec AI Institute 3Meta AI; 4University of Washington

源码：GitHub - DevSinghSachan/unsupervised-passage-reranking: Code and datasets for the paper "Improving Passage Retrieval with Zero-Shot Question Generation"

任务：re-ranking模型，改善open QA中的passage retrieval阶段

方法：使用预训练模型GPT,T5使用zero-shot手段，在候选段落后添加prompt: “Passage: { pi }. Please write a question based on this passage”,从而基于段落生成问题，并以真正的问题作为模型输入，计算基于段落生成问题token的平均概率。以此概率作