Representing anything from scholar papers

Many scholar knowledge mining tasks depend on the representations of different scholar entities and their compositions, including author, publication venue and paper. However, the existing methods always tried to learn different kinds of scholar representations individually. The citation relationships among different types of entities are sometimes ignored. Moreover, the separated learned representations are in separate vector space and hence cannot be operated together. Indeed, representations for different kinds of entities are constructed based on the papers and the citation relationships, which provide the opportunity for developing an unified representation method. In this paper, we propose a novel method for mapping scholar entities and their compositions to a unified distributed vector space. Each paper is firstly defined as a full composition of entities, including title, author, publication venue and so on. Then, with the citation relationships among papers, we train a model that tries to maximize the likelihood of references of papers. After training, the model can map all kinds of entities and their compositions to a unified vector space. We quantitatively evaluate our vectors on two tasks of collaborate prediction and related paper recommendation. The results show our method can learn good representations for scholar entities and their compositions. Since our method allows different types of representations to be operated together, we also show some interesting applications, such as publication venue recommendation and reviewer recommendation. (C) 2019 Published by Elsevier B.V.