Ask Question First for Enhancing Lifelong Language Learning

Proceedings of the 29th International Conference on Computational Linguistics, pages 4610–4621 October 12–17, 2022. 4610 Ask Question First for Enhancing Lifelong Language Learning Han Wang1, 2 and Ruiliu Fu1, 2 and Xuejun Zhang1, 2 and Jun Zhou1, 2 and Qingwei Zhao1, 2 1Institute of Acoustics, Chinese Academy of Sciences, Beijing, China. 2University of Chinese Academy of Sciences, Beijing, China {wanghan, furuiliu, zhangxuejun, zhoujun, qzhao}@hccl.ioa.ac.cn Abstract Lifelong language learning aims to stream learning NLP tasks while retaining knowledge of previous tasks. Previous works based on the language model and following data-free constraint approaches have explored format- ting all data as "begin token (B) + context (C) + question (Q) + answer (A)" for different tasks. However, they still suffer from catastrophic for- getting and are exacerbated when the previous task’s pseudo data is insufficient for the follow- ing reasons: (1) The model has difficulty gener- ating task-corresponding pseudo data, and (2) A is prone to error when A and C are separated by Q because the information of the C is dimin- ished before generating A. Therefore, we pro- pose the Ask Question First and Replay Ques- tion (AQF-RQ), including a novel data format "BQCA" and a new training task to train pseudo questions of previous tasks. Experimental re- sults demonstrate that AQF-RQ makes it easier for the model to generate more pseudo data that match corresponding tasks, and is more robust to both sufficient and insufficient pseudo-data when the task boundary is both clear and un- clear. AQF-RQ can achieve only 0.36% lower performance than multi-task learning. 1 Introduction Lifelong learning is the capacity of human beings to acquire, reconstruct, strengthen, and transfer knowledge (Ring, 1997). Human beings can learn new knowledge while consolidating old knowledge by first detecting and learning the distinctions be- tween old and new knowledge and then simplifying the old and new knowledge based on leveraging the common points. This concept is critical for en- couraging machines to learn NLP tasks in a similar way that people do. In the application of NLP, new data are continuously acquired and categorized as either new data for existing tasks or new data for new tasks. For new data of existing tasks, the tra- ditional method, known as isolated learning (Chen ? Ask questions first before reading Ask questions after reading ---------------------------------------------------------------------------------------------------- Figure 1: Illustration of the difference between asking questions before reading and asking after reading. and Liu, 2018), is to retrain the model with the old data appended with new data. For data from new tasks, multi-task learning (MTL) integrates data from previous and new tasks to retrain the model. Both of these methods are limited to assuming that all new and old data can be obtained during train- ing. However, the reality is that tasks are acquired and trained in the stream, which makes the model suffer from catastrophic forgetting (Ring, 1997; McCloskey and Cohen, 1989; French, 1999) (i.e. forgetting previously learned tasks/knowledge). Lifelong language learning (LLL), which this pa- per focuses on, aims to learn …

Content truncated to 500 words