Exploration and adaptation of large language models for specialized domains

Abstract

Large language models have transformed the field of natural language processing (NLP). Their improved performance on various NLP benchmarks makes them a promising tool—also for the application in specialized domains. Such domains are characterized by highly trained professionals with particular domain expertise. Since these experts are rare, improving the efficiency of their work with automated systems is especially desirable. However, domain-specific text resources hold various challenges for NLP systems. These challenges include distinct language, noisy and scarce data, and a high level of variation. Further, specialized domains present an increased need for transparent systems since they are often applied in high stakes settings. In this dissertation, we examine whether large language models (LLMs) can overcome some of these challenges and propose methods to effectively adapt them to domain-specific requirements. We first investigate the inner workings and abilities of LLMs and show how they can fill the gaps that are present in previous NLP algorithms for specialized domains. To this end, we explore the sources of errors produced by earlier systems to identify which of them can be addressed by using LLMs. Following this, we take a closer look at how information is processed within Transformer-based LLMs to better understand their capabilities. We find that their layers encode different dimensions of the input text. Here, the contextual vector representation, and the general language knowledge learned during pre-training are especially beneficial for solving complex and multi-step tasks common in specialized domains. Following this exploration, we propose solutions for further adapting LLMs to the requirements of domain-specific tasks. We focus on the clinical domain, which incorporates many typical challenges found in specialized domains. We show how to improve generalization by integrating different domain-specific resources into our models. We further analyze the behavior of the produced models and propose a behavioral testing framework that can serve as a tool for communication with domain experts. Finally, we present an approach for incorporating the benefits of LLMs while fulfilling requirements such as interpretability and modularity. The presented solutions show improvements in performance on benchmark datasets and in manually conducted analyses with medical professionals. Our work provides both new insights into the inner workings of pre-trained language models as well as multiple adaptation methods showing that LLMs can be an effective tool for NLP in specialized domains.