The Study Of Mathematical Named Entity Recognition Method Based On Neural Network

Named Entity Recognition(NER)is an important foundational task in natural language processing(NLP)techniques,and it plays a crucial role in constructing knowledge graph and enabling intelligent application in the field of mathematics.However,due to the specialized content and diverse structure of mathematical texts,NER in the mathematical domain faces additional challenges.As a result,there is limited research specifically focused on NER for mathematical knowledge.Through research,it has been found that the current NER method based on the model combining bi-directional long short-term memory with conditional random field(BiLSTM-CRF)shows promising application prospects.However,this method suffers from high computational cost because of the complexity of BiLSTM neural network structure.Additionally,using only one type of neural network combined with CRF is insufficient for fully exploiting the features of the text,leading to inadequate feature extraction accuracy and low recognition precision.Consequently,it is challenging to apply this method effectively to NER in the mathematical domain.Therefore,this thesis aims to address these issues and focuses on the following research areas:(1)Building upon the BiLSTM-CRF model,this thesis proposes a mathematical NER method based on the model combining bi-directional gated recurrent unit and self-attention mechanism with conditional random field(BiGRU-ATT-CRF),which replaces the BiLSTM neural network with BiGRU and introduces the self-attention mechanism.Compared to the BiLSTM-CRF model,this method offers several advantages: the BiGRU neural network has a simpler structure,which can reduce computational complexity,and effectively utilize contextual information to extract global features.The introduction of the self-attention mechanism enhances the efficiency of extracting dependency features from mathematical texts.Experiment results demonstrate that the proposed BiGRU-ATT-CRF-based mathematical NER model outperforms the comparison models,achieving precision,recall,and F1 scores of 93.05%,93.30%,and 93.17% respectively.(2)Building upon the aforementioned research on the BiGRU-ATT-CRF model,this thesis addresses the issue of the BiGRU neural network overlooking certain local features during the mathematical NER process.To overcome this,the model is improved by incorporating iterated dilated the convolutional neural network(IDCNN)and proposing a mathematical NER method that incorporates local features,referred to as the IDCNN-BiGRU-ATT-CRF model.This model combines the local information extracted by IDCNN with the output features of the BiGRU neural network.The fused features are then weighted using the self-attention mechanism to strengthen the key features.Finally,the CRF is employed to learn the optimal labeling sequence by enforcing labeling constraints.By combining the strengths of the IDCNN,BiGRU,and self-attention mechanism,this model effectively integrates the advantages of local features,global features,and key features from the contextual information in mathematical texts.This further enhances the feature extraction capability and improves the accuracy of mathematical NER.Experimental results demonstrate that the improved mathematical NER model achieves an increase of 1.06%,0.97%,and 1.02% in precision,recall,and F1 score respectively,compared to the previous model.These results suggest that the proposed method of incorporating local features can accurately and effectively identify mathematical named entities.