| Summary: | Questioning is a fundamental part of the learning process. As new content arises and learning it becomes vital to the modern society, question generation becomes a necessary job that requires time and resources to be performed effectively. In this document, we propose a Seq2Seq approach that generates a variety of questions that are relevant to the contexts where they are asked. In order to ensure that the generated questions are diverse, relevant, and valuable to learning situations and environments, we use the Revised Bloom’s Taxonomy (RBT), a learning taxonomy that is oriented to learning objectives and can be used to separate questions based on their required cognitive level. However, neural network models require large collections of data to be trained, and datasets addressing RBT are small and scarce. To address this gap, we designed a question classifier that can be used to label current and future datasets using the guidelines provided by RBT. We employed this classifier to create a labeled dataset, which was then used as training data for our proposed Seq2Seq model. In addition, to cover the different taxonomy levels, we create six different fine-tuned models aimed specifically to each one of RBT cognitive levels. Results show that our approach is promising, guaranteeing a variety of questions for all levels of the taxonomy, surpassing the baseline when measured by BLEU-1, and deemed overall well-written, relevant and understandable, by human evaluators.
|
|---|