Mitigating Data Scarcity for Large Language Models (2302.01806v1)

Abstract: In recent years, pretrained neural LLMs (PNLMs) have taken the field of natural language processing by storm, achieving new benchmarks and state-of-the-art performances. These models often rely heavily on annotated data, which may not always be available. Data scarcity are commonly found in specialized domains, such as medical, or in low-resource languages that are underexplored by AI research. In this dissertation, we focus on mitigating data scarcity using data augmentation and neural ensemble learning techniques for neural LLMs. In both research directions, we implement neural network algorithms and evaluate their impact on assisting neural LLMs in downstream NLP tasks. Specifically, for data augmentation, we explore two techniques: 1) creating positive training data by moving an answer span around its original context and 2) using text simplification techniques to introduce a variety of writing styles to the original training data. Our results indicate that these simple and effective solutions improve the performance of neural LLMs considerably in low-resource NLP domains and tasks. For neural ensemble learning, we use a multilabel neural classifier to select the best prediction outcome from a variety of individual pretrained neural LLMs trained for a low-resource medical text simplification task.