Continual Learning of Natural Language Processing Tasks: A Survey

TL;DR

This survey analyzes continual learning (CL) in NLP, highlighting how NLP's task diversity necessitates all CL settings and bidirectional knowledge transfer, unlike the more CF-focused emphasis in CV. It organizes approaches into CF prevention, KT, and ICS, detailing NLP-specific methods such as adapters, instruction-based strategies, and domain-adaptive pre-training, while discussing their limitations in scalability and ICS. The work emphasizes that KT and ICS remain active challenges, with domain adaptation and temporal CL identified as nascent areas, and it outlines practical future directions and evaluation considerations. Overall, the paper provides a comprehensive NLP-focused taxonomy, surveys state-of-the-art methods, and suggests avenues to build more robust, transferable, and scalable CL systems for evolving language tasks.

Abstract

Continual learning (CL) is a learning paradigm that emulates the human capability of learning and accumulating knowledge continually without forgetting the previously learned knowledge and also transferring the learned knowledge to help learn new tasks better. This survey presents a comprehensive review and analysis of the recent progress of CL in NLP, which has significant differences from CL in computer vision and machine learning. It covers (1) all CL settings with a taxonomy of existing techniques; (2) catastrophic forgetting (CF) prevention, (3) knowledge transfer (KT), which is particularly important for NLP tasks; and (4) some theory and the hidden challenge of inter-task class separation (ICS). (1), (3) and (4) have not been included in the existing survey. Finally, a list of future directions is discussed.