Contextual information contributes to biomedical named entity normalization

As one of the most crucial upstream tasks in biomedical informatics, biomedical named entity normalization (BNEN) aims to map mentioned named entities to uniform standard identifiers or terms. Most existing methods only consider the similarity between the individual mention itself and its candidates, however, ignore the valuable information of the context around the mention, which is also very important to understand the real semantic of the mention when it is ambiguous.

In this paper, based on IA-BIOSYN, a representative SOTA (state-of-the-art) BNEN method, we propose a novel BNEN method with contextual information fusion, called CIFSYN, where the context of a given mention is comprehensively considered by putting the given mention's candidates in the same context of the mention, and the contextual information fusion module is introduced to capture the relationship among the mention, candidates, and context.

Experiments on five public BNEN datasets show that our proposed method achieves Acc@1 of 0.934, 0.937, 0.969, 0.959, and 0.856 on NCBI-Disease, BC5CDR-Disease, BC5CDR-Chemical, TAC2017-ADR, and COMETA, respectively, significantly better than other existing SOTA methods, and the introduced context information module brings a 0.5% improvement in Acc@1 on average.

Contextual information around the mention improves the performance of biomedical named entity normalization.