Towards Named Entity Disambiguation with Graph Embeddings

Extracting structured knowledge from scientific literature is crucial for helping professionals make well-informed decisions. This paper presents an approach to distilling knowledge from biomedical documents within the context of Named Entity Disambiguation (NED). The proposed method leverages a joint representation of biomedical entities, combining pre-trained language models with graph machine learning techniques. A Siamese Neural Network (SNN) is trained to optimize this joint representation by integrating the contextual text embeddings of entity mentions with the graph embeddings of corresponding canonical entities in a biomedical Knowledge Graph (KG). During the inference phase, the SNN model assigns a score to this joint representation to disambiguate the target entity among a set of candidates. To the best of our knowledge, this is the first NED method in the biomedical domain that incorporates graph embeddings using a neural model. We empirically evaluated the effectiveness of our approach against well-known biomedical datasets, such as MedMentions and BC5CDR. The results demonstrate a promising direction in utilizing the relational knowledge captured by graph embeddings for the NED task.