基于文献网络节点属性的论文重要性评价模型的构建及评估  

Construction and Assessment of Paper Importance Evaluation Model Based on the Node Attributes of Paper Similarity Network

作　　者：潘现伟[1] 崔雷[1] Pan Xianwei

机构地区：[1]中国医科大学健康管理学院,辽宁沈阳110122

出　　处：《情报理论与实践》2023年第9期130-141,共12页Information Studies:Theory & Application

摘　　要：[目的/意义]学术论文评价是科研评价的基础,是科研管理和评价的刚需。目前基于引文和论文内容视角构建的论文评价模型效果仍有提升的空间。[方法/过程]首先,利用复杂网络分析法,从文献相似性网络的节点属性构建论文重要性评价模型,探讨从文献网络的角度评价论文质量的可行性。其次,选择8个医学相关学科,下载数据形成8个文献数据集,根据论文被Faculty Opinions数据库收录的情况,事先标记为重要论文和普通论文。再次,从论文的主题词、题目摘要和参考文献3种信息源分别构建基于医学主题词表树状结构、Doc2Vec算法和文献耦合的3种文献相似性网络,并对每一种文献相似性网络,利用复杂网络分析方法对网络中的节点进行特征计算,选择具有统计学差异的节点属性指标作为区别两类论文的评价指标。最后,采用4种机器学习算法对数据集中的论文进行二分类学习,构建并评估论文重要性评价模型。[结果/结论]基于文献网络进行论文评价的方法是可行的,3种文献网络构建算法效果差异较小,语义相似性文献网络和文献耦合两种算法略优于基于Doc2Vec的算法。BP神经网络算法在基于文献网络构建的论文重要性评价模型中性能最好。[Purpose/significance] Evaluation of academic papers is the basis of scientific research evaluation,and it is a rigid need for scientific research management and evaluation.At present,there is still something to do to improving the effect of paper evaluation model based on citations and contents of papers.[Method/process] In this study,we constructed a paper importance evaluation model based on the node attributes of the article similarity network using complex network analysis,and explore the feasibility of evaluating the quality of papers from the perspective of the article similarity network.We selected eight medical disciplines,retrieved and downloaded academic papers to form eight datasets,and marked papers as important papers and general papers in advance respectively according to whether the papers were included by the faculty opinions database.three types of article similarity networks were constructed based on mesh structure,Doc2Vec algorithm and bibliographic coupling respectively by extracting mesh terms,titles/abstracts and references of included papers.For each type of article similarity network,we calculated node attributes in the network using complex network analysis,and selected node attributes with statistical differences as evaluation indicators of distinguishing two types of papers.Finally,four types of machine learning algorithms were conducted for classification learning of important papers and general papers in our dataset to build and assess the paper importance evaluation model.[Result/conclusion] We found that the method evaluating papers based on article similarity network is feasible.All three algorithms to construct paper similarity networks had little difference,the algorithm based on MeSH structure and the bibliographic coupling method were better than the Doc2Vec algorithm.BP Neural Network had the best performance in the article importance evaluation model based on the article similarity network.

关 键 词：论文评价 文献相似性网络 复杂网络分析 Doc2Vec算法 机器学习 

分 类 号：G353.1[文化科学—情报学]