Exploring Semantic Links in Bibliographic Network-based Citation Recommendation

Citation recommendations hold substantial academic importance in the era of big scholarly data. The tasks not only depend on the semantic expression of publications, but also are associated with structural information. Recently, many recommendation tasks have been modeled using network-based representations, where nodes signify entities and edges indicate their interactions. Creating node representations with fusing both semantic and structural information in network and considering citation recommendation as publication node similarity searching task is an effective solution. In order to achieve this goal, existing works consistently interact two publications. This paper addresses this challenge by introducing semantic links within a bibliographic network for citation recommendations. The basic motivation is to construct semantic links on network instead of the role of keywords. The semantic links denote two publications sharing similar semantic meanings and provide valuable supervision to learn publication representations. To obtain high-quality semantic links, we create them on network with high-similarity other publications' text representations captured from the two-level hierarchical and transparent information (word-sentence-document). This word-sentence-document information reflects the hierarchical structure of publication texts, accurately capturing their semantics. Finally, we verify the effectiveness of semantic link in bibliographic network-based citation recommendation using two real-world datasets.