Improving the learning of chemical-protein interactions from literature using transfer learning and specialized word embeddings

In this paper, we explore the application of artificial neural network ('deep learning') methods to the problem of detecting chemical-protein interactions in PubMed abstracts. We present here a system using multiple Long Short Term Memory layers to analyse candidate interactions, to determine whether there is a relation and which type. A particular feature of our system is the use of unlabelled data, both to pre-train word embeddings and also pre-train LSTM layers in the neural network. On the BioCreative VI CHEMPROT test corpus, our system achieves an F score of 61.51% (56.10% precision, 67.84% recall).