Named entity recognition of legal judgment based on small-scale labeled data

An important task in intelligent justice is to extract the relevant part of the case description, that is, case elements. Correspondingly, named entity recognition technology is an effective method to extract the elements of the cases. However, most of the existing methods for named entity recognition require large-scale training data, which is a very labor-intensive process to label all data manually, especially for different judgement cases as well as their elements. To this end, this paper proposes a method based on small-scale labeled data in legal judgment. Specifically, first we label data with bootstrapped pattern learning from small-scale manual labeling data, expand the seed dictionary during iteration to obtain sufficient labeled data, and then embed the words by Google BERT pre-trained Chinese model. Finally, CRF+Bilstm named entity recognition model is leveraged to extract case elements. Through the detailed experiments, our method can achieve better entity recognition based on small-scale manual labeling data.