A BERT-Based Hybrid System for Chemical Identification and Indexing in Full-Text Articles

Identification and indexing of chemical compounds in full-text articles are essential steps in biomedical article categorization, information extraction, and biological text mining. BioCreative Challenge was established to evaluate methods for biological text mining and information extraction. Track 2 of BioCreative VII (summer 2021) consists of two subtasks: chemical identification and chemical indexing in full-text PubMed articles. The chemical identification subtask also includes two parts: chemical named entity recognition (NER) and chemical normalization. In this paper, we present our work on developing a hybrid pipeline for chemical named entity recognition, chemical normalization, and chemical indexing in full-text PubMed articles. Specifically, we applied BERT-based methods for chemical NER and chemical indexing, and a sieve-based dictionary matching method for chemical normalization. For subtask 1, we used PubMedBERT with data augmentation on the chemical NER task. Several chemical-MeSH dictionaries including MeSH.XML, SUPP.XML, MRCONSO.RFF, and PubTator chemical annotations are used in a specific order to get the best performance on chemical normalization. We achieved an F1 score of 0.86 and 0.7668 on chemical NER and chemical normalization, respectively. For subtask 2, we formulated it as a binary prediction problem for each individual chemical compound name. We then used a BERT-based model with engineered features and achieved a strict F1 score of 0.4825 on the test set, which is substantially higher than the median F1 score (0.3971) of all the submissions.