While few-shot classification has been widely explored with similarity based methods, few-shot sequence labeling poses a unique challenge as it also calls for modeling the label dependencies. To consider both the item similarity and label dependency, we propose to leverage the conditional random fields (CRFs) in few-shot sequence labeling. It calculates emission score with similarity based methods and obtains transition score with a specially designed transfer mechanism. When applying CRF in the few-shot scenarios, the discrepancy of label sets among different domains makes it hard to use the label dependency learned in prior domains. To tackle this, we introduce the dependency transfer mechanism that transfers abstract label transition patterns. In addition, the similarity methods rely on the high quality sample representation, which is challenging for sequence labeling, because sense of a word is different when measuring its similarity to words in different sentences. To remedy this, we take advantage of recent contextual embedding technique, and further propose a pair-wise embedder. It provides additional certainty for word sense by embedding query and support sentence pairwisely. Experimental results on slot tagging and named entity recognition show that our model significantly outperforms the strongest few-shot learning baseline by 11.76 (21.2%) and 12.18 (97.7%) F1 scores respectively in the one-shot setting.
تحميل البحث