【摘要】 通过评议文档与专家库的专家学术专长匹配,可以输出领域相关的候选专家列表,是同行评议中专家遴选和专家推荐的重要参考依据.针对学术专长匹配存在语义鸿沟、无法反映专家和评议文档多源信息间语义关联的问题,首先对专家信息和评议文档的多源信息进行语义特征抽取,融合多类特征进行表示学习,利用卷积神经网络设计专家特征抽取器ExpFeat和评议特征抽取器RevFeat,采用词嵌入方法和注意力机制对专家专长标签、评议文档关键词、学科分类树语义特征进行抽取和融合,生成具有“小同行”特征的专家和评议文档语义特征向量表示,解决多类信息源间不同学术分类标准造成的语义差异,反映内在语义联系,利用低维稠密向量表达语义信息,降低匹配复杂度.然后,根据专家语义特征表示和评议文档语义特征表示进行学术专长语义匹配,将专家和评议文档特征向量映射到相同语义空间,计算向量间余弦相似度衡量语义相似性,引入负例专家进行模型训练,通过softmax函数计算最大化正例专家概率优化特征提取器参数,进一步提升语义差异的捕捉能力,解决专家信息和评议文档之间的语义鸿沟问题,提升专家匹配效果.最后,在开源的论文评审数据集和项目评审数据集上进行了实验对比和实例分析,结果表明,本文所提方法可以有效提升专家匹配精准度.更多还原

【Abstract】 By matching the review documents with the academic expertise of experts in the expert database,the list of candidate experts related to the field in peer review can be output,which is an important reference basis for expert selection and expert recommendation.Aiming at the problem of semantic gap and unable to reflect semantic association between multi-source information of experts and review documents in academic expertise matching,firstly,multi-source semantic features are extracted from expert information and evaluation documents,and multi-source features are extracted and fused for representation learning.The semantic expression ability of the model can be supplemented by fusing multi-source information,so that the model can fully extract the feature information of experts and review documents for reasonable correlation matching.Convolution neural network is used to design expert feature extractor ExpFeat and review feature extractor RevFeat,and word embedding method and attention mechanism is used to label expert expertise,review document keywords The semantic features of subject classification tree are extracted and fused to generate the semantic feature vector representation of expert and review documents with small peers,by separately modeling the characteristics of "small peers",the academic association information between projects and experts can be captured in a more granular way,solve the differences of academic classification standards among multiple information sources,and reflect the internal semantic connection.In addition,different source information is fused in the feature matching stage,and the original features of different source information are retained as much as possible through vector concatenating operation,so that the model can more fully capture the data distribution differences between information sources,and use low dimensional dense vectors to express semantic information to reduce the matching complexity.Then,the academic expertise matching degree is designed,the academic expertise semantic matching is carried out according to the expert semantic feature representation and the review document semantic feature representation,the expert and review document feature vectors are mapped into the same semantic space,and the cosine similarity between vectors is calculated to measure the semantic similarity.Through end-to-end learning of the neural network,the trainable parameters can converge to the global optimal direction,and the semantic correlation between expert information and review documents can be more accurately captured.And negative case experts are introduced for model training and the expert probability of maximizing positive cases is calculated by softmax function to optimize model parameters and improve the ability to capture semantic differences,so as to solve the semantic gap between expert information and review documents and improve expert matching effect.Finally,experimental comparison and analysis are carried out on the open-source paper review data set.The proposed method can effectively improve the accuracy of expert matching.And when the topK value is 10,compared with the comparison method,the proposed method increases the hit rate by 9% and the gain rate by 10% in the paper review data set,and increases the hit rate by 9% and the gain rate by 6% in the project review data set,which verifies the effectiveness of the proposed method.更多还原