LLM Ensembles through Graphs for Text Classification without Labeled Data

• Grant number: 25/24151-5
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: April 01, 2026
• End date: June 30, 2029
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Ricardo Marcondes Marcacini
• Grantee: João Lucas Luz Lima Sarcinelli
• Host Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil

Abstract

Automatic Text Classification (ATC) has become increasingly relevant for both academic and industrial purposes. Popularly, ATC is a Natural Language Processing (NLP) task performed by Machine Learning (ML), using deep language models that require large amounts of annotated data to achieve satisfactory performance. Large Language Models (LLMs) have been gaining increasing popularity due to their ability to perform tasks they were not explicitly trained for, achieving good results even with few or no annotated data for the task of interest, including for ATC. Currently, LLMs are published and made available very frequently, and differences in architecture, training, and hyperparameters mean they have relative advantages and disadvantages when compared on different NLP tasks. In this context, some works explore the use of these LLMs in a ensemble format to avoid the limitations of individual models. However, research in this area is still scarce, with only simple forms of ensemble aggregation, such as weighted voting, having been considered, partly due to the natural language output nature of LLMs. In contrast to the exclusive use of local vector representations like embeddings, modeling ATC problems as graphs shows that there are advantages to using a structure that considers the global context of the task. This modeling allows the use of Graph Neural Networks (GNNs), which can incorporate diverse information from a node and its neighbors to generate a final representation. The synergy between LLMs and GNNs is currently a popular area of study, especially considering LLMs as weak annotators in graph contexts with few labeled data. On the other hand, a gap exists in the use of GNNs as aggregators for ensembles of LLMs. Thus, this doctoral project proposes to investigate the use of GNNs as aggregators in a scheme similar to stacking for ensemble learning, being able to aggregate heterogeneous information from different LLMs to train a graph-based classifier in scenarios with data scarcity. (AU)