DAML: Chinese Named Entity Recognition with a fusion method of data-augmentation and meta-learning

Overfitting is still a common problem in NER with insufficient data. Latest methods such as Transfer Learning, which focuses on storing knowledge gained while solving one task and applying it to a different but related task, or Model-Agnostic Meta-Learning (MAML), which learns a model parameter initialization that generalizes better to similar tasks. However, these methods still need rich resources to pre-train. In this work, we present new perspectives on how to make the most of in-domain and out-domain information. By introducing a fusion method of data augmentation and MAML, we first use data augmentation to mine more information. With the augmented resources, we directly utilize out-domain and in-domain data with MAML, while avoiding performance degradation after domain transfer. To further improve the model’s generalization ability, we proposed a new data augmentation method based on a generative approach. We conduct experiments on six open Chinese NER datasets (MSRANER, PeopleDairyNER, CLUENER, WeiboNER, Resume NER, and BOSONNER). The results show that our method significantly reduces the impact of insufficient data and outperforms the state-of-the-art.