Sequence Labeling with Meta-Learning

Recent neural architectures in sequence labeling have yielded state-of-the-art performance on single domain data such as newswires. However, they still suffer from (i) requiring massive amounts of training to avoid overfitting; (ii) huge degradation when there is a shift the distribution between and testing. In this paper, we investigate problem adaptation for under homogeneous heterogeneous settings. We propose MetaSeq, novel meta-learning approach labeling. Specifically, MetaSeq incorporates adversarial strategies encourage robust, general transferable representations The key advantage that it capable adapting new unseen domains with small amount annotated those domains. extensively evaluate named entity recognition, part-of-speech tagging slot filling tasks experimental results show achieves against eight baselines. Impressively, surpasses in-domain using only 16.17% 7% target average settings, 34.76%, 24%, 22.5%