Adversarial multi-task learning to solve holistic semantic polishing problem of Chinese text

Abstract

For the core obstacles faced by Chinese proofreading, we think that the effective holistic semantic feature representation of a sentence and its characteristics is a key. Here, we propose a novel adversarial multi-task learning framework to realize above point. Wherein, sentence scoring and word prediction tasks are considered as not only a group of multi-tasks, but also a couple of generative adversarial relationship. Moreover, a policy network is introduced to achieve text polishing based on above adversarial multi-task model and Monte Carlo search training strategy. The ablation experiment results on Xuexi and CLUE corpus show that, for text scoring and prediction tasks, adversarial multi-task learning achieves a 10-20% improvement in accuracy and F1-score compared to the baseline. And, in text correction task, our method is significantly better than the baselines.

Tasks

Reproductions