摘要
针对现有基于大规模预训练语言模型(large-scale pre-trained language models,LLMs)的实体共指消解(entity coreference resolution,ECR)方法在处理长文本和复杂情境时性能受限,且全参数微调计算开销大的问题,进行了一项研究。提出了一种基于提示学习的主题结构增强型ECR模型。此模型利用上下文中的主题结构信息,以增强模型捕捉长距离共指关系的能力;同时,设计了一种可学习的提示模板,显著降低了模型微调所需的计算资源。在三个公开数据集上的实验结果表明,所提方法相较于基准模型,性能分别提升了2.3、0.5和2.6个百分点。并且与当前先进的Link-Append、Seq2seqCoref等模型相比,该方法在仅使用约1.1%参数量的情况下,达到了其约98%的性能水平,证明了该方法在提升长文本ECR任务效果的同时,具备显著的计算效率优势。
To address the suboptimal performance of large-scale pre-trained language model(LLM)-based entity coreference resolution(ECR)methods on long texts and the high computational cost associated with full-parameter fine-tuning,this study developed a topic-structure-enhanced ECR model leveraging prompt-based learning.The model utilized contextual topic structure information to improve the capture of long-range coreference relations.Additionally,a learnable prompt template significantly reduced the computational overhead during fine-tuning.Experimental results demonstrated that the proposed method outperformed baseline models by 2.3,0.5,and 2.6 percentage points on three respective public datasets.Furthermore,compared to state-of-the-art models such as Link-Append and Seq2seqCoref,the proposed method achieved approximately 98%of their performance level while using only about 1.1%of the parameters.This demonstrates the model’s effectiveness and significant computational efficiency for long-text ECR tasks.
作者
刘小明
吴彦博
杨关
刘杰
吴佳昊
Liu Xiaoming;Wu Yanbo;Yang Guan;Liu Jie;Wu Jiahao(School of Artificial Intelligence,Zhengzhou 450007,China;School of Computer Science,Zhongyuan University of Technology,Zhengzhou 450007,China;Zhengzhou Key Laboratory of Text Processing&Image Understanding,Zhengzhou 450007,China;School of Information Science,North China University of Technology,Beijing 100144,China;Research Center for Language Intelligence of China,Beijing 100089,China)
出处
《计算机应用研究》
北大核心
2025年第9期2621-2630,共10页
Application Research of Computers
基金
“新一代人工智能”国家科技重大专项资助项目(2020AAA0109703)
国家自然科学基金联合基金重点项目(U23B2029)
国家自然科学基金资助项目(62076167,61772020)
河南省高等学校重点科研项目(24A520058,24A520060,23A520022)
河南省研究生教育改革与质量提升工程项目(YJS2024AL053)。
