Study of a lifelong learning scenario for question answering
作者：
Highlights：
• Question Answering systems suffer over time several shifts in data distribution.
• Performance after learning a new dataset is sensitive to previous training sequences.
• We find a strategy to avoid the need of testing all possible training sequences.
• Random merge of new training material with previous one avoids performance drop.
摘要
•Question Answering systems suffer over time several shifts in data distribution.•Performance after learning a new dataset is sensitive to previous training sequences.•We find a strategy to avoid the need of testing all possible training sequences.•Random merge of new training material with previous one avoids performance drop.
论文关键词：Question answering,Lifelong learning,Transfer learning,Deep learning
论文评审过程：Received 6 July 2021, Revised 10 May 2022, Accepted 22 July 2022, Available online 27 July 2022, Version of Record 10 August 2022.
论文官网地址：https://doi.org/10.1016/j.eswa.2022.118271
