| dc.description.abstract | Detecting negation and resolving its scope is an essential task in NLP and a priority area in the clinical subfield of NLP. For larger languages such as English, there has been much research in the development of datasets and models for negation resolution, including efforts targeting the medical and clinical domains. Neural methods have come to dominate negation modeling in recent years, but simpler, rule-based approaches are still popular in medical applications. For Norwegian, the availability of resources for negation resolution has until recently been quite sparse. In this thesis, we train an end-to-end negation resolution system utilizing a negation dataset of Norwegian review articles and a simple neural approach inspired by previous work. Using standardized evaluation metrics, the models achieve good results on in-domain test data. Furthermore, we evaluate the applicability of the existing dataset and its guidelines to future projects. Our review shows that better specification of the guidelines is desirable and reveals inconsistencies and annotation errors in the dataset. Building on a previously released dataset, we present NorMed_neg, a publicly available Norwegian negation dataset of biomedical journal articles annotated according to an adjusted version of the mentioned guidelines. The transfer of our models to the medical domain represented by NorMed_neg leads to poor performance, but we find that this can be compensated for by further training on parts of NorMed_neg. A positive effect is observed even with small amounts of training data. Considering the focus on negated symptoms and findings in clinical NLP, we provide our thoughts on the use of models trained according to the existing annotation scheme in a clinical setting. We conclude that adjustments are necessary if the goal is to identify the specific clinical entities described as absent. | eng |