Document classification experiments on TREC precision medicine data, used as a feature for information retrieval.
Final design is an adaption of Yang et al. (NAACL 2016): "Hierarchical Attention Networks for Document Classification" with additional structured information (vectors representing entities/keywords) added to the document level representations. Accuracy on 2017 PubMed during 10-fold crossvalidation was 78.14 (versus 74.96 for logistic regression with BoW and structured information); on 2018 data 75.98 (74.40 baseline) could be achieved.
Training code and models can be found inside precision_medicine_scripts, Notebooks were used during development and for evaluation.