BanditMTL: Bandit-based Multi-task Learning for Text Classification

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 5506–5516 August 1–6, 2021. ©2021 Association for Computational Linguistics 5506 BanditMTL: Bandit-based Multi-task Learning for Text Classification Yuren Mao1, Zekai Wang2, Weiwei Liu2∗, Xuemin Lin1, Wenbin Hu2 1School of Computer Science and Engineering, University of New South Wales 2School of Computer Science, Wuhan University yuren.mao@unsw.edu.au, {wzekai99,liuweiwei863}@gmail.com lxue@cse.unsw.edu.au, hwb@whu.edu.cn Abstract Task variance regularization, which can be used to improve the generalization of Multi- task Learning (MTL) models, remains unex- plored in multi-task text classification. Ac- cordingly, to fill this gap, this paper investi- gates how the task might be effectively regular- ized, and consequently proposes a multi-task learning method based on adversarial multi- armed bandit. The proposed method, named BanditMTL, regularizes the task variance by means of a mirror gradient ascent-descent al- gorithm. Adopting BanditMTL in the multi- task text classification context is found to achieve state-of-the-art performance. The re- sults of extensive experiments back up our the- oretical analysis and validate the superiority of our proposals. 1 Introduction Multi-task Learning (MTL), which involves the si- multaneous learning of multiple tasks, can achieve better performance than learning each task indepen- dently (Caruana, 1993; Ando and Zhang, 2005). It has achieved great success in various applications, ranging from summary quality estimation (Kriz et al., 2020) to text classification (Liu et al., 2017). In the multi-task text classification context, MTL simultaneously learns the tasks by minimizing their empirical losses together; for example, by mini- mizing the mean of the empirical losses for the in- cluded tasks. However, it is common for these tasks to be competing. Minimizing the losses of some tasks increases the losses of others, which accord- ingly increases the task variance (variance between the task-specific loss). Large task variance can lead to over-fitting in some tasks and under-fitting in others, which degenerates the generalization per- formance of an MTL model. To illustrate this issue, *Corresponding author. it is instructive to consider a case of two-task learn- ing, where task 1 and task 2 are conflicting binary classification tasks. When the task variance is un- controlled, it is possible that the empirical loss of task 1 will converge to 0, while the empirical loss of task 2 will converge to 0.5. In such a case, although the mean of the empirical losses is decreasing, task 1 overfits and task 2 underfits, which leads to poor generalization performance. To address the problem caused by uncontrolled task variance, it is necessary to implement task vari- ance regularization, which regularizes the variance between the task-specific losses during training. However, existing deep MTL methods, including both adaptive weighting sum methods (Kendall et al., 2018; Chen et al., 2018; Liu et al., 2017) and multi-objective optimization-based methods (Sener and Koltun, 2018; Mao et al., 2020b), ig- nore the task variance. Overlooking task variance degenerates an MTL model’s generalization ability. To fill this gap and further improve the general- ization ability of MTL models, …

Content truncated to 500 words