ConTinTin: Continual Learning from Task Instructions

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3062 - 3072 May 22-27, 2022 c⃝2022 Association for Computational Linguistics ConTinTin: Continual Learning from Task Instructions Wenpeng Yin∗ Temple University wenpeng.yin@temple.edu Jia Li Salesforce Research jia.li@salesforce.com Caiming Xiong Salesforce Research cxiong@salesforce.com Abstract The mainstream machine learning paradigms for NLP often work with two underlying pre- sumptions. First, the target task is predefined and static; a system merely needs to learn to solve it exclusively. Second, the supervision of a task mainly comes from a set of labeled examples. A question arises: how to build a system that can keep learning new tasks from their instructions? This work defines a new learning paradigm ConTinTin (Continual Learning from Task Instructions), in which a system should learn a sequence of new tasks one by one, each task is explained by a piece of textual instruc- tion. The system is required to (i) generate the expected outputs of a new task by learn- ing from its instruction, (ii) transfer the knowl- edge acquired from upstream tasks to help solve downstream tasks (i.e., forward-transfer), and (iii) retain or even improve the perfor- mance on earlier tasks after learning new tasks (i.e., backward-transfer). This new problem is studied on a stream of more than 60 tasks, each equipped with an instruction. Techni- cally, our method InstructionSpeak con- tains two strategies that make full use of task instructions to improve forward-transfer and backward-transfer: one is to learn from nega- tive outputs, the other is to re-visit instructions of previous tasks. To our knowledge, this is the first time to study ConTinTin in NLP. In addition to the problem formulation and our promising approach, this work also contributes to providing rich analyses for the community to better understand this novel learning problem. 1 Introduction The main goal of machine learning algorithms lies in seeking supervision for solving a target task. Tra- ditionally, the supervision is extracted from a set of labeled examples. The learner constructs a decision function that generalizes beyond the seen examples. ∗Work was done at Salesforce Research. While this paradigm has been tremendously suc- cessful for many NLP problems, an inherent draw- back exists in it: the learner can only be as good as the provided data (Goldwasser and Roth, 2014). Learning, therefore, relies on annotating a large volume of training data, an expensive and time- consuming process. To alleviate the costly demand for task-specific annotation (referred as S0 here- after), the human learning process suggests at least two sources of alternative supervision: one is to ac- cumulate knowledge from tasks learned in the past (S1) (Richard, 1970; Thrun and Mitchell, 1995; Chomsky, 2002); the other is to learn from natu- ral instructions (S2) describing a high-level story about target tasks (Goldwasser and Roth, 2014). Unfortunately, we rarely see the joint power of S1 and S2. In this work, we present a new learning paradigm ConTinTin – continual Learning from task instructions. In ConTinTin, each task is …

Content truncated to 500 words