“…Many biomedical relation extraction systems have often relied hand-crafted linguistic features (Gu et al, 2016;Peng et al, 2016) but recently also convolutional neural networks (Nguyen and Verspoor, 2018;Choi, 2018), LSTM (Li et al, 2017;Sahu and Anand, 2018) or a combination of machine learning models and neuralnetwork-based encoders (Zhang et al, 2018;Peng et al, 2018). A recent paper (Verga et al, 2018) achieves state-of-the-art results on biomedical relation classification for chemically-induced diseases (CDR (Li et al, 2016)) and ChemProt (CPR (Krallinger M., 2017)), by using a Transformer encoder (Vaswani et al, 2017) and end-to-end Named Entity Recognition and relation extraction, without, however, leveraging transformer-based language model pre-training.…”