Year: 2023
Type: dissertation
Abstract: The advent of pre-trained language models (LMs) has enabled unprecedented advances in the Natural Language Processing (NLP) field. In this respect, various specialised \nLMs for the biomedical domain have been introduced, and similar to their general purpose counterparts, these models have achieved state-of-the-art results in many biomedical NLP tasks. Accordingly, it can be assumed that they can perform medical reasoning. However, given the challenging nature of the biomedical domain and the scarcity \nof labelled data, it is still not fully understood what type of knowledge these models encapsulate and how they can be enhanced further. This research seeks to address \nthese questions, with a focus on the task of interpreting patient case descriptions, which \nprovides the means to investigate the model’s ability to perform medical reasoning. In \ngeneral, this task is concerned with inferring a diagnosis or recommending a treatment \nfrom a text fragment describing a set of symptoms accompanied by other information. \nTherefore, we started by probing pre-trained language models. For this purpose, we \nconstructed a benchmark that is derived from an existing dataset (MedNLI). Following \nthat, to improve the performance of LMs, we used a distant supervision strategy to \nidentify cases that are similar to a given one. We then showed that using such similar cases can lead to better results than other strategies for augmenting the input to \nthe LM. As a final contribution, we studied the possibility of fine-tuning biomedical \nLMs on PubMed abstracts that correspond to case reports. In particular, we proposed \na self-supervision task which mimics the downstream tasks of inferring diagnoses and \nrecommending treatments. The findings in this thesis indicate that the performance of the considered biomedical LMs can be improved by using methods that go beyond \nrelying on additional manually annotated datasets.
Author: Alghanmi, Israa
Institution: Cardiff University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W6982604694
