Automatic segmentation of texts into units of meaning for reading assistance
Project Overview
The document explores the application of generative AI in education, particularly focusing on enhancing reading accessibility for individuals with dyslexia through the development of enriched digital books. By employing Transfer Learning with Google's BERT model, the technology automates the segmentation of texts into meaningful units, known as rhesis, thereby aiding comprehension for dyslexic readers. This innovative approach not only improves educational outcomes by making reading materials more accessible but also aims to combat illiteracy in this population. The findings suggest that integrating generative AI into educational resources can significantly support diverse learning needs, demonstrating the potential of AI to create inclusive learning environments and foster better educational achievements for students with dyslexia.
Key Applications
Rhezor tool for automatic text segmentation
Context: Educational context for individuals with dyslexia, particularly children and young adults
Implementation: Developed through collaboration between Mobidys and Pacte Novation, utilizing AI techniques for automated segmentation based on linguistic models.
Outcomes: Improved accessibility to reading materials, enhanced comprehension for dyslexic readers, and positive feedback from users regarding the effectiveness of segmentation into rhesis.
Challenges: Initial versions of the tool had limitations in segmentation quality, and varying preferences among users for different segmentation methods.
Implementation Barriers
Technical barrier
The initial version of the Rhezor tool tended to cut texts too finely, affecting the quality of segmentation.
Proposed Solutions: Post-processing techniques were added to group segments more effectively while respecting the defined span.
Data barrier
Insufficient volume of manually segmented texts for effective training of the AI model.
Proposed Solutions: Utilizing Transfer Learning techniques to improve model performance despite limited training data.
Project Team
Jean-Claude Houbart
Researcher
Solen Quiniou
Researcher
Marion Berthaut
Researcher
Béatrice Daille
Researcher
Claire Salomé
Researcher
Contact Information
For information about the paper, please contact the authors.
Authors: Jean-Claude Houbart, Solen Quiniou, Marion Berthaut, Béatrice Daille, Claire Salomé
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