BERT_BLSTM_CRF entity recognition algorithm

• 1、School of Computer, Beijing University of Posts and Telecommunications，Beijing 100876

Abstract：Aiming at the problems of various indicators, classification subdivision, difficult to complete coverage and accurate extraction of existing science and technology resources, this paper proposes an entity recognition algorithm based on BERT_BLSTM_CRF to extract the scientific and technological resource entities and construct the scientific and technological big data portrait. In this paper, we use the best pre training language model instead of word2vec language model. By adjusting the context of all layers, we use the bidirectional transformer network structure with strong semantic ability to pre train the language model. We can enrich the semantic dimension of words according to the context information of words. Then we introduce the current mainstream deep learning sequence annotation model blstm-crf model as the benchmark model The output word vector sequence after the pre training of Bert is input into blstm-crf for training. Firstly, the text to be recognized is input into the bidirectional LSTM model to obtain the corresponding label of each character. These tags have strong dependence. The CRF layer linked to the two-way LSTM is used to learn the dependency relationship between tags. Finally, the global optimal sentence level tag sequence is obtained. Experimental results show that compared with the traditional entity recognition algorithm model, the proposed entity recognition algorithm model based on BERT_BLSTM_CRF has higher accuracy, which verifies the effectiveness of the proposed method.

Keywords： BERT；LSTM；CRF；Entity Recognition