A Word Similarity Feature-based Semi-supervised Approach for Named Entity Recognition

摘要

Named Entity Recognition (NER) is an important branch of Natural Language Processing (NLP). Among the existed NER methods, one of the most advanced and commonly deployed approach is the Long Short Term Memory with a Conditional Random Field layer (LSTM-CRF). However, this supervised method generally requires a large number of labeled corpuses, which is very limited regarding the texts in drug patent of this study. Bearing this in mind, a word similarity feature-based semi-supervised NER approach is proposed in this study. The feature of word similarity with regard to various types of entities are firstly extracted from word embedding to form similarity constraint. Then they are combined with the features computed by supervised LSTM. Finally, the tagged results are obtained through the CRF layer. By introducing the similarity feature of word embedding to LSTM-CRF model, the proposed method can greatly reduce the untagged cases in a large amount of similar entities. Experimental studies demonstrated that the proposed method performs obvious advantages in both the accuracy and comprehensiveness when compared with the traditional baseline model and other semi-supervised methods.