Training NER Models: Knowledge Graphs in the Loop

Motivated by the need of annotated data for training named entity recognition models, in this work we present our experiments on a distantly supervised approach using domain specific vocabularies and raw texts in the same domain. In the experiments we use MeSH vocabulary and a random sample of PubMed articles to automatically create an annotated corpus and train a named entity recognition model capable to identify diseases in raw text. We evaluate method against the manually curated CoNLL-2003 corpus and the NCBI-disease corpus.