Entity Recognition and Linking for Biomedical Documents Applying recent Transformer-based Entity Recognition and Linking Algorithms for the Biomedical Domain, to a Multi-Lingual Scenario
Year: 2024
Type: dissertation
Abstract: In the ever expanding domain of Biomedicine, a wealth of crucial information is embedded within an extensive number of free-text documents. However, the unstructured nature of this textual reservoir, coupled with the intricacy of biomedical terminology, poses a significant challenge for automated systems to extract valuable insights in an efficient way. Furthermore, the majority of resources are allocated to English, but other lower- resource languages also contain valuable information. The current state-of-the-art models in multilingual biomedical Natural Language Processing lag behind their general domain and English-specific counterparts. This disparity emphasizes the need for approaches that can tackle the complexity of biomedical literature across languages. This dissertation focuses on two Natural Language Processing tasks within the infor- mation extraction realm: Named Entity Recognition (NER) and Entity Linking (EL). To tackle these tasks, we used the following methodology: first, we investigated the efficacy of transfer learning approaches, by adapting pre-trained Transformer-based models to the complexities of multilingual biomedical texts. Second, we employed a data augmentation technique to enrich the training data and try to enhance the performance of those models. We evaluated our approaches on the SympTEMIST, CANTEMIST and MultiCardioNER shared tasks - competitions that provide a benchmark for evaluating NER and EL tech- niques within the biomedical domain for various languages. We obtained competitive results for both NER and EL, consistently surpassing the mean and median results for those shared tasks, and even establishing a new state-of-the-art score for DrugTEMIST English in NER. Our methodology can be easily extended to other languages and datasets.
Author: Gonçalves, Rodrigo Miguel Gameiro Vilhena
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7028069805