Browsing by Author "Gundyreva, Elina"

In this thesis, a novel method for linking scientific articles to taxonomy terms in the domain of food systems research is presented. With food systems being in the center of 12 of the 17 United Nations Sustainable Development goals, there has been an ever-growing amount of scientific articles in this field. These articles are vital in understanding the complex nature of food systems and their inter-dependencies. However, finding relevant literature in this field is difficult for decision makers given the interdisciplinary nature of the field and that annotation and expert feedback is expensive. In the thesis, BERT-based models (SBERT, SPECTER and SciBERT) are adapted to the food systems area and fine-tuned for tasks such as text classification and text similarity, which represents a solution to the problem of finding relevant articles in the food systems domain. The proposed search system uses several taxonomies and data augmentation to achieve the results, which are visualized in a created website. Linking food systems research articles to taxonomy terms shows good accuracy, with models finetuned on domain data achieving better performance on classification task. The best fine-tuning strategy for SPECTER and SciBERT is the combination of domain adaptation and classification. Fine-tuning for text similarity for SBERT improves SBERT performance only slightly. The proposed method can be used in other domains than food systems.