A multi-strategy fusion neural network method for named entity recognition in PST
Yongkang Zhu, Peiyan Wang, RuiTing Li, Dongfeng Cai · 2024
This paper presents a novel multi-strategy fusion neural network Named Entity Recognition (NER) method aimed at identifying entities in Process Specification Text (PST). To address the challenges of short sentence length, entity type error, and entity boundary error caused by limited context and semantic information in small sample data, the proposed method incorporates semantic enhancement (BM25, SBERT) and data augmentation (AUG) techniques. Additionally, a constraint matrix (Matrix) is designed to improve the consistency of the model and address the inconsistency caused by the pre-training model in the BERT-CRF model. The proposed method demonstrates significant improvements in both model consistency and F1 value. Specifically, the model consistency is improved from 90.71% to 94.51%, and the F1 value is increased from 76.90% to 77.21%. These results validate the effectiveness of the proposed approach for NER in PST.