基于深度语义挖掘的引文推荐多样化研究

People commonly encounter three major issues when employing a citation recommendation system: 1) top-ranked recommendations are often highly similar with each other (we name it as homogenous recommendation), making it difficult to obtain sufficient literatures with a full coverage of all potential subtopics; 2) without a proper diversification mechanism, the homogenous recommendations may easily deviate from the information needs of tail users; and 3) users’ needs are contextual-dependent, highly relying on their current task-at-hand, search purposes, and etc. This cannot be simply satisfied with the homogenous recommendations. ..As a result, this project attempts to propose a framework for citation recommendation diversification, which includes the investigation of corresponding user needs, potential application domains, dimensions for diversification, and evaluation techniques for diversified recommendations. Built on top of the past successes on search result diversification and recommendation result diversification, we are planning to expand and extend related theories/practices into the area of citation recommendation. Moreover, we would like to push the current content recommendation technique from simple text analysis toward semantic understanding; thus, we take into account multiple semantic dimensions for recommendation such as topic analysis, citation functions, and word/term functions. Our ultimate goal is to integrate the above-mentioned factors and construct a highly semantic-based, contextual-aware citation recommendation system for satisfying diversified recommendation needs. We plan to deploy a live recommendation system on a selective domain, for the purpose of exploring potential issues, such as algorithm scalability, when constructing such system, as well as evaluating the validity of our propose framework...We expect both theoretical and practical values of our projects in multiple domains. The proposed recommendation diversification framework could be beneficial for the study of digital library, scholarly information search, and etc. The proposed recommendation algorithms might also be applicable in multiple scenarios for practical use.

在使用引文推荐系统时，学者们常常遇到以下问题：1）推荐结果相似性高，需要查看大量条目才能收集足够的文献；2）极端情况下，同质化的推荐结果可能整体偏离用户需要；2）用户在不同场景下有着不同的文献需求，无法为同质性的推荐结果所满足。.本课题尝试构建学术文献引文多样化推荐理论框架，对引文推荐多样化的用户需求、应用场景、多样化维度、评测方法进行梳理和分析；课题立足于信息检索多样性研究和信息推荐多样性研究的已有成果，在引文推荐多样化中引入主题、引文功能、章节功能、文献类型、词汇功能等多种语义信息，研究基于多维语义特征的全局引文推荐多样化技术和基于上下文需求感知的局部引文推荐多样化技术。课题组还将选取合适的学科领域，构建实际可用的多样化推荐系统。.本课题具有较大的理论意义和应用价值：课题在数字图书馆、学术检索、战略阅读等领域有一定的理论创新意义；提出的算法及构建的系统在多个实际场景中具有潜在应用价值。

引文推荐是信息推荐的重要应用场景，也是学术文本挖掘、科学知识组织与挖掘的重要应用任务。已有的学术文献引文推荐往往直接借鉴通用的方式和模型，对推荐的个性化、场景化和多样化重视不够，导致推荐任务的实际效果不佳。.本研究从学术文献推荐的基础需求和实践范式出发，构建学术文献引文多样化推荐理论框架，对引文推荐多样化的用户需求、应用场景、多样化维度、评测方法进行梳理和分析。课题组立足于信息检索多样性研究和信息推荐多样性研究的已有成果，在引文推荐多样化中引入主题、引文上下文、词汇功能等多种语义信息，构建了一套大规模细粒度标注的引文推荐多样化研究数据集。在此基础上，课题组深入研究基于多维语义特征的全局引文推荐多样化技术和基于上下文需求感知的局部引文推荐多样化技术。在全局引文推荐多样化部分，课题组提出了基于聚类主题多样性指标的引文推荐模型，帮助用户快速生成相关研究；在局部引文推荐多样化部分，课题组构建了一种基于引文上下文理解的多样化结果生成模型，提出了基于分数差的DivScore算法，获得了较CiteSeerX系统更好的多样化结果。课题组还选取了计算机科学学科作为案例，构建了实际可用的引文多样化推荐原型系统。.课题研究产出研究成果包括论文、专著、数据集和软件工具。论文成果12篇，其中1）SCI/SSCI索引论文2篇、CSSCI索引论文5篇、计算机领域顶级会议CCF A类论文2篇，领域权威或核心期刊论文合计7篇；2）构建数据集“细粒度语义信息标注的引文推荐多样化研究数据集”1个；3）开发与设计了“基于学术文本深度语义挖掘的引文多样化推荐原型系统”1个；4）出版专著一部。.本课题具有较大的理论意义和应用价值：课题在数字图书馆、学术检索、智能科学家等研究领域有一定的理论创新意义；提出的算法及构建的系统在多个实际场景中具有应用价值。