BANNER-CHEMDNER: Incorporating Domain Knowledge in Chemical and Drug Named Entity Recognition

Exploiting unlabeled text data to leverage the system performance has been an active and challenging research topic in text mining, due to the recent growth of the amount of biomedical literature. Named entity recognition is an essential prerequisite task before effective text mining of biomedical literature can begin. The participants of the CHEMDNER task of the BioCreative IV challenge are asked to develop a chemical compounds and drugs mention recognition system and are given a set of annotated PubMed documents for training and a set of selected PubMed documents for evaluation of their systems by the organizers. Our primary goal is to develop a named entity recognition system that can scale well over millions of documents and can easily be plugged in a biomedical text mining system, while exploiting unlabeled data to leverage the system performance. We extracted Brown cluster labels and word embeddings, both induced from a large unlabeled document corpus as the word representation features in addition to the word, the word and character n-gram, and the traditional orthographic features. During the training, 2 ± order CRF model was built with varied feature spaces and the influence of the different groups of features defining the baselines of the experiment was observed in terms of the F-score, recall, precision as well as processing time. Our system achieves 81.41% and 82.58% of Fscores on CHEMDNER development set for CEM and CDI sub-tasks, processing ~530 documents per minute.