Chemistry-specific Features and Heuristics for Developing a CRF-based Chemical Named Entity Recogniser

We describe and compare methods developed for the BioCreative IV chemical compound and drug name recognition (CHEMDNER) task. The presented conditional random fields (CRF)-based named entity recogniser employs a statistical model trained on domain-specific features, in addition to those typically used in biomedical NERs. In order to increase recall, two heuristics-based post-processing steps were introduced, namely, abbreviation recognition and re-labelling based on a token’s chemical segment composition. The chemical NER was used to generate predictions for both the Chemical Entity Mention recognition (CEM) and Chemical Document Indexing (CDI) subtasks of the challenge. Results obtained from training a model on the provided training set and testing on the development set show that employing chemistryspecific features and heuristics leads to an increase in performance in both subtasks.