Year: 2023
Type: dissertation
Abstract: The biomedical field is a critical area for natural language processing (NLP) applications because it involves a vast amount of unstructured data, including clinical notes, medical publications, and electronic health records. NLP techniques can help extract valuable information from these documents, such as disease symptoms, medication usage, and treatment outcomes, which can improve patient care and clinical decision-making. MAPS S.p.A. currently produces Clinika, a software that extracts knowledge from clinical corpora. Clinika performs the task of Named Entity Recognition (NER) by linking entities to medical concepts from an established knowledge base, in this case, the Unified Medical Language System (UMLS). This dissertation details how we approached designing and implementing a component for the new version of Clinika, specifically a mention embedder that uses embeddings to perform entity linking with UMLS concepts. We focused on enhancing existing dense contextual embeddings by injecting ontological knowledge, using two parallel approaches: (1) taking the embeddings as a by-product of an entity alignment model aided by an ontology, and (2) fine-tuning a contextual language model with custom sampling strategies. We evaluated both approaches with suitable experiments from the relevant literature. After testing, we discontinued the first approach but found more significant results using the second. The results on the tasks chosen to evaluate the embeddings were not promising, we address the causes in the Error Analysis section, and discuss further work on this topic.
Author: Niccolai, Lorenzo
Institution: University of Bologna
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7029489977
