LAMAL: LAnguage Modeling Is All You Need for Lifelong Language Learning

  • 2019-09-07 20:17:34
  • Fan-Keng Sun, Cheng-Hao Ho, Hung-Yi Lee
  • 6

Abstract

Most research on lifelong learning (LLL) applies to images or games, but notlanguage. Here, we introduce LAMAL, a simple yet effective method for LLL basedon language modeling. LAMAL replays pseudo samples of previous tasks whilerequiring no extra memory or model capacity. To be specific, LAMAL is alanguage model learning to solve the task and generate training samples at thesame time. At the beginning of training a new task, the model generates somepseudo samples of previous tasks to train alongside the data of the new task.The results show that LAMAL prevents catastrophic forgetting without any signof intransigence and can solve up to five very different language taskssequentially with only one model. Overall, LAMAL outperforms previous methodsby a considerable margin and is only 2-3\% worse than multitasking which isusually considered as the upper bound of LLL. Our source code is available athttps://github.com/xxx.

 

Quick Read (beta)

loading the full paper ...