Reproduce “Continual Learning for Temporal-Sensitive Question Answering” on OpenResearch

About this paper

Wanqi Yang, Yunqiu Xu, Yanda Li, Kunze Wang, Binbin Huang, Ling Chen

This paper introduces the Continual Learning for Temporal-Sensitive Question Answering (CLTSQA) task and a corresponding framework, allowing QA models to continuously update their knowledge about temporal facts without forgetting previously learned information. The framework, featuring Temporal Memory Replay and Temporal Contrastive Learning, demonstrates the ability to maintain performance on older datasets while integrating new temporal data.

Key insights

• Catastrophic forgetting in temporal QA can be mitigated by selectively replaying easily learnable samples from previous time periods.
• Enhancing a model's sensitivity to temporal information through contrastive learning, by generating questions with altered or similar temporal references, improves its ability to handle evolving facts.
• Framing temporal question answering as a continual learning problem allows models to adapt to new information over time while retaining the ability to answer questions about past events.

Summary by alphaXiv.