End-to-End Open-Domain Question Answering with BERTserini
Wei Yang,1,2 Yuqing Xie,1,2††Aileen Lin,2 Xingyu Li,2 Luchen Tan,2 Kun Xiong,2 Ming Li,1,2 equal contribution Jimmy Lin1,2 1 David R. Cheriton School of Computer Science, University of Waterloo 2 RSVP.ai
Abstract
We demonstrate an end-to-end question answering system that integrates BERT with the open-source Anserini information retrieval toolkit. In contrast to most question answering and reading comprehension models today, which operate over small amounts of input text, our system integrates best practices from IR with a BERT-based reader to identify answers from a large corpus of Wikipedia articles in an end-to-end fashion. We report large improvements over previous results on a standard benchmark test collection, showing that fine-tuning pretrained BERT with SQuAD is sufficient to achieve high accuracy in identifying answer spans.
中文速览
开放域问答系统需要在海量文档中找到准确答案，但现有大多数模型只处理少量输入文本，无法直接面对大规模语料库。BERTserini 把经典信息检索工具 Anserini（基于 BM25 的检索引擎）和预训练语言模型 BERT 拼成一条两阶段流水线：先用 Anserini 从 500 余万篇维基百科文章中检索出相关段落，再用在 SQuAD 上微调过的 BERT 从这些段落里抽取具体答案。实验表明，以段落为检索粒度、取前 100 个候选段落时，系统在标准测评集上的精确匹配得分大幅超越此前最优结果，而且仅凭 SQuAD 微调就足以达到很高精度，无需为开放域场景专门设计复杂的训练策略。这项工作证明，将成熟 IR 工程实践与大规模预训练模型相结合，是构建实用端到端问答系统的有效路径，也为后续在检索重排和多语言扩展方向的改进提供了清晰的分析基础。
原文 arXiv:1902.01718;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1902.01718v2