Experiments with GloVe embeddings and Domain Adversarial Neural Networks on the Dutch medical domain

Abstract

The focus in this thesis is on developing models and resources that will be useful for the Dutch medical domain. This domain lacks annotated data and domain-specific models. In the fist part of the thesis, GloVe embeddings (Pennington et al., 2014) are developed. However, evaluating the quality of these embeddings is a challenge, given the lack of annotated resources for medical Dutch. The second part of the thesis presents experiments using a novel domain adaptation method, Domain Adversarial Neural Networks, which is getting attention for domain-adaptation problems in NLP. The network is trained on a Named Entity Recognition task and a Part-of-Speech tagging task, with and without (English) medical embeddings. Its performance and suitability for various domain-adaptation scenarios is evaluated.