Deep learning based Answering Questions using T5 and Structured Question Generation System’

Since there is an agile change in all the domains, it is imperative for students to keep themselves updated to gain expertise in their field of study. Although there is easy access to resources from the web, learning only becomes complete after a thorough system of assessment to identify gaps in the learning process if any. Automatic question generation techniques have been introduced to reduce the time associated with the manual construction of questions. Researchers have come up with stateof-art models to generate these questions. Models like BERT and GPT gained popularity among the masses due to their deduction of complex sentences. These have however been replaced with advanced models such as Transformers which are much more versatile. This paper has fine-tuned one such transformer and proposed a custom model to generate questions. Our system mainly focuses on two types of questions, namely multiple-choice, and long answer types. The generator model is primarily concerned with producing questions from selected phrases and generating semantically motivated phrase-level distractors as MCQ response choices. In contrast, the evaluator model is concerned with identifying the most relevant question-answer combinations. This research work has gained the ability to acquire 76% syntactic accuracy and 84% fluency from the question-answer pairs evaluated using our approach. Along with this, the proposed result has achieved a final accuracy of 82% from the evaluator's bert-base-cased model.