Learning to Generalize Unseen Domains via Multi-source Meta Learning for Text Classification

With the rapid development of deep learning methods, there have been many breakthroughs in the field of text classification. Models developed for this task have achievedhigh accuracy. However, most of these models are trained using labeled data from seen domains. It is difficult for these models to maintain high accuracy in a new challenging unseen domain, which is directly related to the generalization of the model. In this paper, we study the multi-source Domain Generalization for text classification and propose a framework to use multiple seen domains to train a model that can achieve high accuracy in an unseen domain. Specifically, we propose a multi-source meta-learning Domain Generalization framework to simulate the process of model generalization to an unseen domain, so as to extract sufficient domain-related features. We introduce a memory mechanism to store domain-specific features, which coordinate with the meta-learning framework. Besides, we adopt a novel “jury” mechanism that enables the model to learn sufficient domain-invariant features. Experiments demonstrate that our meta-learning framework can effectively enhance the ability of the model to generalize to an unseen domain and can outperform the state-of-the-art methods on multi-source text classification datasets.