Document Ranking for Curated Document Databases Using BERT and Knowledge Graph Embeddings: Introducing GRAB-Rank

Curated Document Databases (CDD) play an important role in helping researchers find relevant articles in scientific literature. Considerable recent attention has been given to the use of various document ranking algorithms to support the maintenance of CDDs. The typical approach is to represent the update document collection using a form of word embedding and to input this into a ranking model; the resulting document rankings can then be used to decide which documents should be added to the CDD and which should be rejected. The hypothesis considered in this paper is that a better ranking model can be produced if a hybrid embedding is used. To this end the Knowledge Graph And BERT Ranking (GRAB-Rank) approach is presented. The Online Resource for Recruitment research in Clinical trials (ORRCA) CDD was used as a focus for the work and as a means of evaluating the proposed technique. The GRAB-Rank approach is fully described and evaluated in the context of learning to rank for the purpose of maintaining CDDs. The evaluation indicates that the hypothesis is correct, hybrid embedding outperforms individual embeddings used in isolation. The evaluation also indicates that GRAB-Rank outperforms a traditional approach based on BM25 and and a ngram-based SVR document ranking approach.