Multi-information fusion based on dual attention and text embedding network for local citation recommendation

Abstract

Local citation recommendation is a list of references that researchers need to cite based on a given context, so it could help researchers produce high-quality academic writing quickly and efficiently. However, existing citation recommendation methods only consider contextual content or author information, ignore the critical influence of historical citation information on citations, and learn the paper embedding at a coarse-grained level, resulting in lower-quality recommendations. To solve the above problems, we propose a novel two-stage citation recommendation model with multiple information fusion (MICR). The first stage is to enhance the target paper’s representation learning of the MICR model. To achieve the above goal, three encoders, which contain context information encoder, historical citation encoder, and author information encoder, are designed to learn rich representations of the target paper. The second stage is to select appropriate recommendation strategies for the target paper and candidate papers to achieve the goal of efficient citation recommendation. Experiments on two public citation datasets show that our model outperforms several competitive baseline methods on citation recommendation.