融合多类型软提示与全局指针网络的古籍命名实体识别OA

Named entity recognition(NER)for ancient texts aims to identify previously unseen entities using a limited set of annotated examples.The annotation scarcity,ambiguous entity boundaries,and complex contextual scenarios character-istic of ancient texts pose significant challenges for traditional methods that fine-tune pre-trained language models(PLMs).These methods often struggle with the detection of nested entities and are prone to error propagation in their two-stage processes.Although prompt-based learning shows promise in few-shot learning tasks,it has not been fully opti-mized to accommodate the unique linguistic features of ancient texts,leading to suboptimal information extraction.Further-more,these methods necessitate intricately designed templates and multiple rounds of inference,resulting in high com-putational costs and reduced stability.To address these challenges,this paper proposes a novel method for ancient text NER that integrates multi-type soft prompts with an efficient global pointer(EGP)network.The method begins by initial-izing multi-type soft prompts using various entity type labels,from which several learnable entity category soft prompt vectors are generated through embedding matrix mappings to enhance semantic recognition capabilities.An attention mechanism is then employed to optimize the interaction between the soft prompts and the input texts,thereby improving the comprehension of the model for ancient text semantics.By inputting multiple soft prompts in parallel to the EGP,the model can predict all entities concurrently and delineate their boundaries precisely,resolving the issues of error propaga-tion and overlapping entities of conventional two-stage models.Compared to the BERT-BiLSTM-CRF methodology,the proposed method demonstrates significant improvements,achieving increases of 11.56%,9.52%,and 13.66%in F1 scores across the Thirty Biographies,ritual texts,and CHisIEC datasets,respectively.Moreover,the inference efficiency of the proposed model is 37.65 times greater than that of the BERT-BiLSTM-CRF approach,validating its efficiency in the domain of ancient text NER.