Skip to main content Skip to navigation

Robust Educational Dialogue Act Classifiers with Low-Resource and Imbalanced Datasets

Project Overview

The document explores the integration of generative AI in education, specifically through the enhancement of educational dialogue act classifiers aimed at improving intelligent tutoring systems. It emphasizes the development of these classifiers to effectively operate in low-resource and imbalanced data environments, utilizing AUC maximization techniques for better performance. The research demonstrates the importance of automating the identification of dialogue acts within tutoring dialogues, which is crucial for optimizing educational interactions and outcomes. Findings indicate that AUC maximization methods significantly surpass traditional cross-entropy approaches in their ability to manage imbalanced datasets, thus reinforcing the potential of generative AI to transform educational practices by facilitating more effective and responsive tutoring systems.

Key Applications

Dialogue Act Classification

Context: Educational context focusing on tutoring dialogues, targeted at both tutors and students

Implementation: Utilized machine learning models to classify dialogue acts based on labeled sentences from tutoring dialogues, optimizing with AUC maximization instead of traditional cross-entropy loss

Outcomes: Improved classification performance and robustness of dialogue act classifiers under low-resource and imbalanced data conditions

Challenges: Limited labeled data leading to imbalanced class distribution, which impacts the performance of classifiers, especially for minority classes

Implementation Barriers

Data-related and Resource-related barrier

Imbalanced class distribution in training datasets due to limited labeled data, leading to poor performance on minority classes and challenges in low-resource scenarios.

Proposed Solutions: Adopt AUC maximization approaches to enhance classifier robustness to imbalanced data distribution and improve performance under low-resource conditions.

Project Team

Jionghao Lin

Researcher

Wei Tan

Researcher

Ngoc Dang Nguyen

Researcher

David Lang

Researcher

Lan Du

Researcher

Wray Buntine

Researcher

Richard Beare

Researcher

Guanliang Chen

Researcher

Dragan Gasevic

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Jionghao Lin, Wei Tan, Ngoc Dang Nguyen, David Lang, Lan Du, Wray Buntine, Richard Beare, Guanliang Chen, Dragan Gasevic

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies