Nested Named Entity Recognition in Geotechnical Engineering Based on Pre-training and Information Enhancement

Geotechnical engineering entity recognition is the foundation and important prerequisite for text mining and knowledge graph construction in geotechnical engineering. To address the nested named entity recognition problem in geotechnical engineering, a small-scale Chinese nested named entity corpus in geotechnical engineering texts was constructed using a combination of sequence labeling and span labeling methods. A geotechnical engineering named entity recognition deep learning model, BERT-BiLSTM-LIE, was designed: (1) the representation learning layer utilized the BERT pre-trained language model to achieve transfer learning of geotechnical engineering text features; (2) the BiLSTM contextual encoding layer encoded contextual features of geotechnical engineering texts; (3) the LIE local enhancement layer enhanced local correlation information between encoding vectors; (4) the MLP label decoding layer generated geotechnical engineering named entity label sequences conforming to labeling rules. Finally, experimental analysis of the BERT-BiLSTM-LIE model was conducted based on the geotechnical engineering nested named entity corpus. In the controlled experiments, the model achieved good results, with a precision of 89.12%, a recall of 89.77%, and an F1 score of 89.44%. Experiments were also conducted on flat named entity and nested named entity corpora. The experimental results indicate that compared to baseline models and other pre-trained models, BERT-BiLSTM-LIE model performs well in nested named entity recognition in small-scale geotechnical engineering corpora, capturing more detailed semantic information in deep text understanding compared to flat named entity. The construction of the nested named entity corpus effectively improves the model’s performance.