Rational LAMOL: A Rationale-based Lifelong Learning Framework

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 2942–2953 August 1–6, 2021. ©2021 Association for Computational Linguistics 2942 Rational LAMOL: A Rationale-Based Lifelong Learning Framework Kasidis Kanwatchara1∗, Thanapapas Horsuwan1∗, Piyawat Lertvittayakumjorn2, Boonserm Kijsirikul1, Peerapon Vateekul1† 1 Department of Computer Engineering, Faculty of Engineering, Chulalongkorn University, Thailand 2 Department of Computing, Imperial College London, UK {kanwatchara.k, thanapapas.h}@gmail.com, pl1515@imperial.ac.uk, {boonserm.k, peerapon.v}@chula.ac.th Abstract Lifelong learning (LL) aims to train a neu- ral network on a stream of tasks while retain- ing knowledge from previous tasks. However, many prior attempts in NLP still suffer from the catastrophic forgetting issue, where the model completely forgets what it just learned in the previous tasks. In this paper, we in- troduce Rational LAMOL, a novel end-to-end LL framework for language models. In or- der to alleviate catastrophic forgetting, Ratio- nal LAMOL enhances LAMOL, a recent LL model, by applying critical freezing guided by human rationales. When the human rationales are not available, we propose exploiting unsu- pervised generated rationales as substitutions. In the experiment, we tested Rational LAMOL on permutations of three datasets from the ERASER benchmark. The results show that our proposed framework outperformed vanilla LAMOL on most permutations. Furthermore, unsupervised rationale generation was able to consistently improve the overall LL perfor- mance from the baseline without relying on human-annotated rationales. We made our code publicly available at https://github. com/kanwatchara-k/r_lamol. 1 Introduction The grounds of lifelong learning (LL) stem from the ability of humans to continually acquire, consol- idate, and transfer knowledge and skills throughout their lifespan. This ability is also important for real-world natural language processing (NLP) ap- plications, where autonomous agents are required to interact with users from various domains through continuous streams of information and language ∗ Equal contributions † Corresponding author semantic drifts occur over time. The existing dom- inant paradigm for machine learning, however, is isolated learning (Chen and Liu, 2016). While isolated learning has shown some successes in a variety of domains, their applicability remains lim- ited to the assumption that all samples are available during the learning phase. When a stream of tasks are trained sequentially, machine learning and neu- ral network models face catastrophic forgetting or interference (McCloskey and Cohen, 1989). This occurs due to the non-stationary data distribution that biases the model. We focus on lifelong language learning (LLL), which is lifelong learning on a stream of NLP tasks. To the best of our knowledge, the grounds of LLL are left largely underexplored. LAMOL is an LLL general framework that has garnered recent interest due to its simplicity (Sun et al., 2020). In particular, LAMOL transforms all NLP tasks into the question answering (QA) format according to McCann et al. (2018) and generates pseudo-samples of old tasks using its language modeling (LM) capability to re- fresh the learned knowledge. However, there is still a gap between the performance of LAMOL and the result of multi-task learning which is generally con- sidered as …

Content truncated to 500 words