Skip to main content Skip to navigation

CIKMar: A Dual-Encoder Approach to Prompt-Based Reranking in Educational Dialogue Systems

Project Overview

The document explores the application of generative AI in education through the development and assessment of CIKMar, an educational dialogue system powered by the Gemma Language model and a Dual-Encoder ranking system. CIKMar is designed to produce relevant and accurate responses tailored to educational needs, demonstrating the promise of smaller language models in making advanced educational AI more accessible. However, the study also notes specific challenges, particularly the Dual-Encoder's inclination to favor theoretical responses over practical ones, which could impact the effectiveness of the system in real-world educational scenarios. Overall, the findings underscore both the potential benefits and the limitations of integrating generative AI in educational settings, highlighting the importance of continuous refinement to enhance the practical applicability of AI-generated content in learning environments.

Key Applications

CIKMar - an educational dialogue system powered by the Gemma Language model

Context: This system is designed for educational dialogue, particularly in teacher-student interactions to provide personalized learning experiences.

Implementation: CIKMar uses a Dual-Encoder ranking system with BERT and SBERT to rerank responses generated by the Gemma model based on context relevance.

Outcomes: Achieved a recall and F1-score of 0.70, indicating effective response generation and context relevance.

Challenges: The Dual-Encoder may prioritize theoretical responses over practical application, leading to less relevant rankings.

Implementation Barriers

Technical Barrier

Large language models are often impractical due to their size and memory requirements, making them inaccessible for many users.

Proposed Solutions: Utilizing smaller models like Gemma that require less computational power and RAM.

Response Quality Barrier

Inconsistently accurate and contextually appropriate responses from the model.

Proposed Solutions: Implementing a prompt-and-rerank approach to optimize response generation based on contextual understanding.

Project Team

Joanito Agili Lopo

Researcher

Marina Indah Prasasti

Researcher

Alma Permatasari

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Joanito Agili Lopo, Marina Indah Prasasti, Alma Permatasari

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