Evaluating Adapter-based Knowledge-enhanced Language Models in the Biomedical Domain

Motivation

Knowledge-enhanced Language Models (KELMs) are a promising next step in the advancement of Language Models (LMs) with a steadily rising amount of research. By injecting data from knowledge graphs and making use of expert knowledge from large ontologies, they bring structure into the unstructured nature of LMs. KELMs can be used in the biomedical domain to assist medical professionals, accelerate research, and help provide medical advice for people in remote areas.

Working together with industry partners of the chair, the thesis draws on data from a private ontology leveraging the “SciWalker” platform. The thesis contributes to the ongoing research in three ways: (1) A comprehensive literature review of adapter-based KELMs, (2) the Development of KELMs based on data from SciWalker, and (3) the conduction of a survey addressed to medical students and professionals. All three contributions are novel to the field of biomedical KELMs and are based on the research questions listed below.

Research Questions

• What adapter-based approaches to knowledge-enhancement exist, and how do they compare to each other?

• Can we improve existing approaches with new methods and data from a private ontology?

• Is the research on biomedical KELMs relevant to medical professionals, and what factors hinder or support the deployment of the technology in practice?

Research Design

• Systematic literature review

• Design and development of model training pipeline

• Model training

• Model evaluation and comparison to related work
  - Quantitative
  - Qualitative probing

• Research survey