PhD Thesis Defense: Mengxue Zhang, AI-Driven Scoring System for Open-Ended Mathematical Reasoning
Content
Speaker
Abstract
Mathematical reasoning assessment presents unique challenges for automated scoring systems due to the structured nature of mathematical language, the sensitivity of numerical values and operations, and the inherent variability of correct solution paths. While significant progress has been made in automated scoring for other subjects, mathematical responses remain difficult to evaluate comprehensively because they often intertwine natural language explanations with formal mathematical notation. Furthermore, effective assessment requires an understanding of the procedural steps underlying a student's reasoning, which can vary widely even among correct solutions. This thesis aims to address these challenges by developing a series of approaches for the automated assessment of open-ended mathematical responses. The core objectives are:
- Developing various scoring models and representation learning techniques to enhance performance and ensure generalization across different mathematical question types.
- Improving solution generation approaches to produce coherent, controllable, and interpretable reasoning steps, suitable for understanding solution strategies and potentially generating feedback.
- Establishing robust evaluation methodologies, including interpretable methods, to thoroughly assess the quality and reliability of the developed scoring and analysis models.
Finally, the thesis will discuss how the growing capabilities of large
language models (LLMs) can further benefit education, along with the considerations and potential concerns we must address to adapt responsibly in this rapidly evolving landscape.
Advisor
Andrew Lan