面向中文科学数据集的句子级语义匹配模型  

Semantic Matching Model for Chinese Scientific Datasets

作　　者：刘建平 初新涛 王健[3] 顾勋勋 王萌 王影菲 LIU Jianping;CHU Xintao;WANG Jian;GU Xunxun;WANG Meng;WANG Yingfei(College of Computer Science and Engineering,North Minzu University,Yinchuan 750021,China;The Key Laboratory of Images and Graphics Intelligent Processing of State Ethnic Affairs Commission,North Minzu University,Yinchuan 750021,China;Agricultural Information Institute,Chinese Academy of Agricultural Sciences,Beijing 100081,China)

机构地区：[1]北方民族大学计算机科学与工程学院,宁夏银川750021 [2]北方民族大学图像图形智能处理国家民委重点实验室,宁夏银川750021 [3]中国农业科学院农业信息研究所,北京100081

出　　处：《郑州大学学报（工学版）》2024年第6期56-64,共9页Journal of Zhengzhou University（Engineering Science）

基　　金：宁夏回族自治区重点研发计划(2022BSB03044);宁夏回族自治区自然科学基金资助项目(2021AAC03205);北方民族大学科研启动金项目(2020KYQD37)。

摘　　要：针对现有以词为粒度的语义匹配模型难以理解句子级科学数据集元数据的问题,提出了一个面向中文科学数据集的句子级语义匹配(CSDSM)模型。该模型使用CSL数据集对SimCSE和CoSENT进行训练生成CoSENT预训练模型。基于CoSENT模型,引入多头自注意力机制进行特征提取,通过余弦相似度与KNN分类结果加权求和得到最终输出。以国家地球系统科学数据中心开放的语义元数据信息作为自建科学数据集进行实验,实验结果表明:与中文BERT模型相比,所提模型在公共数据集AFQMC、LCQMC、Chinese-STS-B和PAWS-X上的Spearman指标ρ分别提升了0.0448,0.0290,0.1777和0.0509;在自建科学数据集上的F 1和Acc分别提升了0.0788和0.0634,所提模型能够有效地解决科学数据集句子级语义匹配问题。In order to address the difficulty of existing word-level semantic matching models in understanding sentence-level scientific dataset metadata,a sentence-level semantic matching(CSDSM)model for Chinese scientific datasets was proposed.The model used the CSL dataset to train and generate the CoSENT pre-training model based on SimCSE and CoSENT.Building upon the CoSENT model,a multi-head self-attention mechanism was introduced for feature extraction,and the final output was obtained by weighting the cosine similarity and KNN classification results.Experimental data from the National Earth System Science Data Center′s open semantic metadata information was used as a self-built scientific dataset.The experimental results showed that compared to the Chinese BERT model,the proposed model improved the Spearman′sρindex by 0.0448,0.0290,0.1777 and 0.0509 on the public datasets AFQMC,LCQMC,Chinese-STS-B,and PAWS-X,respectively.Additionally,F 1 and Acc on the self-built scientific dataset were improved by 0.0788 and 0.0634 respectively.The proposed model effectively addresses the problem of sentence-level semantic matching in scientific datasets.

关 键 词：文本匹配 语义匹配 预训练模型 科学数据集 自然语言处理 

分 类 号：TP3-05[自动化与计算机技术—计算机科学与技术] TP391.1