Meta-learning driven approach for transparent automated natural language processing

Abstract

This thesis addresses the challenges of Automated Machine Learning (AutoML) for Natural Language Processing (NLP) tasks, emphasizing the creation of models and facilitating the analysis, visualization, and interpretability of the results. Leveraging concepts from AutoML, the study explores meta-learning and hyperparameter optimization to develop a flexible and transparent approach – aiming to streamline the NLP pipeline and accommodate users with varying levels of expertise, from beginners to experts. As part of the contributions, it was developed MetaText-ClassifAI, an end-to-end AutoML framework for text classification that integrates data characterization, vectorization optimization, meta-learning, and automatic pipeline recommendation, achieving state-of-thepractice performance while reducing computational cost, execution time, and environmental impact, maintaining transparency and interpretability in the recommendation process. Furthermore, the work contributes with a analysis of the use of Large Language Models (LLMs) in NLP tasks, demonstrating that, although these models are effective for automating tasks such as exploratory data analysis, traditional statistical methods can still outperform modern LLM-based approaches in specific scenarios, achieving competitive performance with lower computational cost and greater efficiency.