Lingual-Agnostic Meta-Learning for Low-Resource Part-of-Speech Tagging
Xuejun Zhang, Yujiang Li, Pengyuan Zhang, Yonghong Yan · 2020
Current deep learning based cross-lingual Part-of-Speech (POS) tagging methods are limited by their ability to achieve fast learning and generalization when the data in the target language is scarce. In this paper, we integrate a meta-learning procedure that uses the knowledge learned across many tasks as an inductive bias towards better POS tagging. Based on the Model-Agnostic Meta-Learning framework (MAML), we propose a Lingual-Agnostic Meta-Learning (LAML) for cross-lingual low-resource POS tagging. The proposed LAML models cross-lingual POS tagging as a meta-learning problem, and we learn to adapt to low-resource languages based on multilingual high-resource languages. Also, different from the original MALM, LAML distinguishes parameters into shared parameters for represent learning and parameters to be adapted to different languages. We demonstrate that the proposed model outperforms the multilingual, joint-learning based approaches and enables us to train a competitive POS tagging system with only a fraction of samples.