Multitask & Meta Learning for Language Models: Enhancing Aspect Based Sentiment Analysis

This chapter presents a comprehensive investigation into improving Aspect Based Sentiment Analysis (ABSA) through multitask learning, meta learning, and task sampling strategies within the framework of language models. Leveraging state-of-the-art models like BERT and XLNet, the study explores the impact of pretraining tasks, particularly Next next-sentence prediction, on ABSA performance. Through meticulous experimentation and analysis, the efficacy of sampling tasks based on an importance subtask hierarchy is demonstrated, showcasing significant enhancements over state-of-the-art benchmarks. The findings underscore the importance of incorporating diverse tasks and sampling strategies for advancing ABSA and related natural language processing tasks, offering valuable insights for future research endeavors.