Adaptive Question Generation and Answer Evaluation with Dynamic Text Complexity Adjustment

Abstract

The ability to read and comprehend a non-native language should be developed from the early childhood. The education system in Sri Lanka requires students to learn English language which is a dominant universal language as the secondary language. Still, the individual attention provided in the classroom is not enough for the proper grasp of language skills by the students. And it is time consuming for the teachers to generate multiple questions and to evaluate the students individually (PROBLEM). The system proposed as solution facilitates enhancing language comprehension of students within the age group of 8-11 years, in a personalized manner (SOLUTION). The student is provided with an interactive and eye-catching web application interface built using React and Tailwind. System allows the teachers to add comprehension passages related to the student’s age. The system can challenge the students with questions of different difficulty levels that are individually catered to the student’s performance level, which is an exceptional feature of this system. To achieve this, a question-generation model is created and trained using a custom-made dataset created from domain-specific textbooks, customizing the T5 pre-trained model, incorporating style embeddings, and custom methods. In parallel answer evaluation is done using a pre-trained model deepset/roberta-base-squad2. The difficulty level of the provided text passage is calculated based on Flesch Reading Ease score. The pain points in learning and teaching comprehension skills are reduced through this solution due to the personalized and automatic question generation, and answer evaluation features. The educators are given the privilege to add relevant text passages and keep in track with the student’s engagement and scoring with ease. The students are given personalized attention to their performance and relevant passages are provided.