A Resume Entity Recognition Method Based on Co-Prediction
Purpose:At present,there is a multitude of entity types in individual resumes,with a significant presence of both flat and nested entities intertwined.This complexity has had a considerable negative impact on entity recognition.Method:We designed ajointly predicted named entity recognition framework.First,the pretrained model Mengzi-BERT was used to express the word embedding of the context.To make full use of the features extracted by the pre-trained model,we compressed the network depth,enlarged the receptive field of the convolutional layers,and incorporated a self-attention mechanism.Subse-quently,a new named entity recognition model,TPDCA(triple layers progressive dilated convolutional neural network-atten-tion),is designed.Simultaneously,to prevent excessively large spans between entities and addressing the issue of nested enti-ties in resumes,based on the Biaffine bidirectional affine attention mechanism,a novel model for local relationship entity recog-nition named BCN(biaffine-based local relationship capture network)had been designed.Finally,a joint prediction framework,Mengzi-TPDCA-CRF-BCN,was constructed by separately adjusting the prediction weights of the TPDCA model and the BCN local relation recognition model.This framework exhibits the best overall performance in entity recognition.This method avoids the model losing long-distance dependencies between entities,reduces the negative impact of the intertwined flat and nested en-tities on predictions,and resolves the issue of high coupling between entity types affecting the recognition task.Conclusion:Compared to the current mainstream methods,the model proposed in this paper has shown a 3%improvement in various evaluation metrics,effectively addressing the practical issues of high coupling between entity types and large spans between entities in resumes.
natural language processingpretrained modelnamed entity recognitiondeep learningresume informationco-prediction
NETL
NSTL
万方数据
