02279nas a2200373   4500000000100000000000100001008004100002653002400043653002500067653000900092653002100101653001400122653002900136653002700165653002700192653001600219653002500235653003400260653000900294653003100303653002600334653002900360653002100389653001200410100001100422700001300433700001200446245007300458856011800531300001200649490000700661520121700668022002001885                                       d10aAttention mechanism10aAttention mechanisms10aBERT10aComputer science10aComputers10aConditional random field10aDynamic representation10aElectrical engineering10aEngineering10aForward-and-backward10aIntangible cultural heritages10aLSTM10aLong and short term memory10aMulti-neural networks10aNamed entity recognition10aRandom processes10aThangka1 aX. Guo1 aS. Cheng1 aW. Wang00aNamed entity recognition in thangka field based on bert-bilstm-crf-a  uhttps://www.scopus.com/inward/record.uri?eid=2-s2.0-85102971591&partnerID=40&md5=b35d8dbab6b317c09c00b21942ea8c6b  a161-1740 v833 aThangka is one of the precious intangible cultural heritages, which is closely related to Tibetan Buddhism. However, Tibetan Buddhism has a complex system, and the naming patterns of various deities are not fixed and difficult to identify from Chinese texts. In this paper, we propose a multi-neural network fusion named entity recognition model BERT-BiLSTM-CRF-a which is based on the BERT pre-training language model, Bidirectional Long-and-Short Term Memory (BiLSTM) and Conditional Random Field (CRF). Specifically, the model uses the BERT to enhance the dynamic representation ability. Then, a weighting method from attention mechanism is introduced to weight the forward and backward BiLSTM hidden layer vectors before concatenating to further improve the effective utilization of context features. Finally, CRF model is used to output the global optimal annotation results. Experimental results on the test sets show that the recall of the BERT-BiLSTM-CRF-a model is 87.4\%, 8.2\% higher than the traditional named entity recognition model BiLSTM-CRF, and the F1 value is also 4.8\% higher. Therefore, the model we proposed can be effectively used in the task of named entity recognition in thangka field.  a22863540 (ISSN)