Design of a Deep-Learning Model to Improve Learning Capabilities of LD Children via Statistical Modelling of Examination Behavioral Patterns

Students with learning disorders (LD) are unable to perform certain set of tasks due to their difficulty in understanding & interpreting them. These tasks include, but are not limited to, solving simple Mathematical identities, understanding English Grammar related questions, spelling certain words, arranging words in sequence, etc. A wide variety of system models are proposed by researchers to analyze such issues with LD students, and recommend various remedies for the same. But a very few of these models are designed for end-to-end continuous learning support, which limits their applicability. Moreover, even fewer system models are designed to improve capabilities of LD students, via modification of system’s internal parameters. To cater these issues, a novel deep-learning model (DL2CSMBP) is proposed in this text, which assists in incrementally improving learning capabilities of LD children via statistical modelling of examination behavioural patterns. The model initially proposes design of a novel examination system that generates question sets based on student’s temporal performance, and collects their responses via an LD-friendly approach. These responses are processed using a deep learning model that extracts statistical characteristics from student responses. These characteristics include question skipping probability, percentage of correct answers, question revisit probability, time spent on each question, un-attempted questions, & frequently skipped question types. They were extracted from 12 different question types which include Basic English Grammar, Medium English Grammar, Advanced English Grammar, direct comprehension, inference comprehension, vocabulary comprehension, sequencing, spelling, synonyms, Mathematics (addition & subtraction), and finding the odd Man out. The results of these questions were evaluated for 80+ LD students, and their responses were observed. Based on these responses a customized 1D convolutional Neural Network (CNN) layer was trained, which assisted in improving classification performance. It was observed that the proposed model was able to identify LD students with 95.6% efficiency. The LD students were able to incrementally improve the performance by attempting a series of exam sessions. Due to this incremental performance improvement, the LD students were able to cover 28% more questions, and answer almost 97% of these questions with precision & correctness. Due to such promising results, the system is capable of real-time deployment, and can act as an automated schooling tool for LD students to incrementally improve their examination performance without need of medical & psychological experts. This can also assist in reducing depression among LD students, because they don’t need to interact with a physical doctor while improving their LD condition in real-time, thus suggesting its use in non-intrusive medical treatments of these students.