Research On Elementary Mathematical Named Entity Recognition Based On Deep Learning

With the rapid development of artificial intelligence and Internet,the informatization and technology in the field of education have also been continuously improved.Due to the large and complex resources,how to use artificial intelligence technology to clean,integrate and use these information is becoming more and more important.The named entity recognition task in the field of elementary mathematics refers to identify mathematical entities from elementary mathematical texts,such as triangles,equations,etc.,which lays the foundation for subsequent tasks such as building elementary mathematical knowledge graph,understanding of elementary mathematical problems and automatic solutions to elementary mathematical problems.Aiming at this,two elementary mathematical named entity recognition models based on deep learning are proposed,and experiments are designed to verify the effectiveness of the two models.The paper mainly includes the following three parts:(1)This paper briefly analyzes the characteristics of elementary mathematics problems.According to this,then determines the categories of elementary mathematics named entities,which includes eight entity types,namely equations,functions,algebraic formulas,points,lines,angles,triangles and quadrilaterals.About 300,000 words of elementary math problem text was crawled from Baidu library,question bank and question bank website as the experimental data set,which labeled manually by the {B,I,O} notation method.(2)A LEBER-based elementary mathematical named entity recognition model-LEBERT-Bi LSTM-CRF is proposed,which adopts LEBERT as the pre-training model.It can fuse vocabulary information into the underlying encoding process of BERT through lexicon adapter,making full use of the encoding ability of transformer.After obtaining the word vector by LEBERT,the Bi LSTM network is used to extract its features.And then the extracted features are labeled through the CRF model.Finally,the efficiency of LEBERT-Bi LSTM-CRF model is verified by experiments,F1-score reached94.30%.(3)An improved LEBERT-based elementary mathematical named entity recognition model-LEBERT-Bi LSTM-IDCNN-CRF is proposed.It obtains word vectors fused with lexical information by the LEBERT model first.Then enters it into the Bi LSTM-IDCNN network for feature extraction,and uses the attention mechanism to integrate features,highlighting key features and weakening or even ignoring irrelevant features.Finally,CRF with Focal Loss is used to label the entities,which effectively improves the accuracy of entity recognition in elementary mathematical texts.Compared to LEBERT-Bi LSTM-CRF,the recall of this model is improved by 2.07%,and the F1 value is improved by 0.92%.