A Light Transfer Model for Chinese Named Entity Recognition for Specialty Domain

Abstract

Named entity recognition (NER) for specialty domain is a challenging task since the labels are specific and there are not sufficient labelled data for training. In this paper, we propose a simple but effective method, named Light Transfer NER model (LTN), to tackle this problem. Different with most traditional methods that fine tune the network or reconstruct its probing layer, we design an additional part over a general NER network for new labels in the specific task. By this way, on the one hand, we can reuse the knowledge learned in the general NER task as much as possible, from the granular elements for combining inputs, to higher level embedding of outputs. On the other hand, the model can be easily adapted to the domain specific NER task without reconstruction. We also adopt the linear combination on each dimension of input feature vectors instead of using vector concatenation, which reduces about half parameters in the forward levels of network and makes the transfer light. We compare our model with other state-of-the-art NER models on real datasets against different quantity of labelled data. The experimental results show that our model is consistently superior than baseline methods on both effectiveness and efficiency, especially in case of low-resource data for specialty domain.