摘要
随着大数据时代的到来,信息抽取已成为自然语言处理领域的重要研究方向。信息抽取涉及多项任务,包括命名实体识别、关系抽取和事件抽取等,每项任务通常需要依靠专用模型来应对其特定的挑战。该文提出一种基于提示学习的ERNIE-BiLSTM-PN通用信息抽取方法(EBP-UIE),结合预训练语言模型(ERNIE)、双向长短期记忆网络(BiLSTM)和指针网络(PN),旨在通过一个统一的框架解决信息抽取任务的复杂性,并实现跨任务知识的共享。ERNIE优化了对文本的深层理解和上下文分析,BiLSTM的应用加强了对序列特征的捕捉及长距离依赖关系的解析,PN则提高了对文本中信息元素起止位置的精确标定,提示学习机制灵活实现多个信息抽取任务的统一建模。实验结果显示:在命名实体识别任务,EBP-UIE在MSRA和PeopleDaily数据集上的F1分数比UIE模型分别高出7.12%和0.53%;在关系抽取任务,EBP-UIE在DuIE数据集上的F1分数超过UIE模型6.84%;对于事件抽取任务,EBP-UIE在DuEE数据集上的触发词和论元抽取F1分数分别比UIE模型高出4.49%和0.95%。
With the advent of the big data era,information extraction has become a significant research direction in the field of natural language processing.Information extraction involves multiple tasks,including named entity recognition,relation extraction,and event extraction,each typically relying on specialized models to address its specific challenges.This paper proposes a universal information extraction method based on prompt learning(EBP-UIE),enhanced representation through knowledge integration(ERNIE),bi-directional long short-term memory networks(BiLSTM),and pointer networks(PN),aimed at resolving the complexities of information extraction tasks through a unified framework and facilitating cross-task knowledge sharing.The introduction of the ERNIE model enhances deep text understanding and contextual analysis,the application of BiLSTM strengthens the capture of sequential features and the parsing of long-distance dependencies,and the pointer network improves the precise identification of start and end positions of information elements in text.The experimental results show that on named entity recognition,the F1 scores of EBP-UIE on the MSRA and PeopleDaily datasets are respectively 7.12%and 0.53%higher than those of the UIE models;on relation extraction,the F1 score of EBP-UIE on the DuIE dataset exceeded that of the UIE model by 6.84%;And on the event extraction,the F1 score of EBP-UIE outperforms the UIE model by 4.49%and 0.95%in trigger word and argument extraction performance on the DuEE dataset,respectively.
作者
刘万里
雍新有
曹开臣
陈俞舟
刘禄波
蔡世民
LIU Wanli;YONG Xinyou;CAO Kaichen;CHEN Yuzhou;LIU Lubo;CAI Shimin(Second Laboratory,Southwest Institute of Electronic Technology,Chengdu 610036,China;Big Data Research Center,University of Electronic Science and Technology of China,Chengdu 611731,China)
出处
《电子科技大学学报》
北大核心
2025年第3期411-423,共13页
Journal of University of Electronic Science and Technology of China
基金
国家自然科学基金(T2293771,11975071)。
关键词
通用信息抽取
深度学习
指针网络
提示学习
universal information extraction
deep learning
pointer network
prompt learning
