NSF GRFP: Interpretable Reading Comprehension Models

Abstract

This research proposal focuses on developing interpretable AI models for reading comprehension that provide insights into both machine and human understanding. By combining hierarchical graph neural networks with cognitive science principles, the proposed approach aims to create more transparent and effective reading comprehension systems that can explain their reasoning processes.

Figure 1: Hierarchical Graph Neural Network Architecture for Reading Comprehension

Research Objectives

1. Develop a hierarchical graph neural network architecture that represents text at multiple levels of abstraction
2. Design interpretable attention mechanisms that highlight the reasoning processes used in comprehension tasks
3. Evaluate the model's ability to predict human reading patterns and comprehension challenges
4. Create visualization tools that make the model's reasoning processes accessible to educators and researchers

Intellectual Merit

This project advances the field of AI in several ways. First, it addresses the critical need for interpretability in deep learning models, particularly in educational applications. Second, it develops novel hierarchical graph neural network architectures that can represent the multi-level structure of text comprehension. Third, it bridges cognitive science and machine learning by creating models that not only perform well but also provide insights into human reading processes.

Broader Impacts

The research has significant potential to improve educational technologies, particularly for students with reading difficulties. Interpretable models can help educators identify specific comprehension challenges and develop targeted interventions. Additionally, the visualization tools developed in this project will make AI systems more accessible to educational researchers who lack technical backgrounds in machine learning.