# Question Generation for Adaptive Education

**Authors:** Megha Srivastava, Noah Goodman  
**arXiv:** [2106.04262](https://arxiv.org/abs/2106.04262)  
**Date:** 2021-06-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.04262v1/

## Key points

- Intelligent and adaptive online education systems aim to provide high-quality education to a diverse range of students.
- Limited pool of hand-made questions restricts their ability to adapt to individual students in a fine-grained and open-ended manner.
- Researchers at Stanford University explore targeted question generation as a controllable sequence generation task.
- Pre-trained language models can be fine-tuned for deep knowledge tracing (LM-KT).
- LM-KT accurately predicts the probability of a student answering a question correctly and generalizes well to questions not seen during training.
- The team employs LM-KT to generate questions based on the student's abilities and target difficulty.
- Results show that they successfully generate novel, well-calibrated language translation questions for second language learners from an actual online education platform.
- The generated questions cover various levels of difficulty ranging from very easy (0.9) such as "spoon or tea" to more challenging ones like "translate 'the woman'".
- This approach has significant potential in providing personalized educational experiences for a broader population of students.

## Summary

The goal of intelligent and adaptive online education systems is to provide high-quality education to a diverse range of students. To address the limitation of relying on a limited pool of hand-made questions which restricts their ability to adapt to individual students in a fine-grained and open-ended manner, researchers at Stanford University explore targeted question generation as a controllable sequence generation task. They first demonstrate how pre-trained language models can be fine-tuned for deep knowledge tracing (LM-KT). This model accurately predicts the probability of a student answering a question correctly and generalizes well to questions not seen during training. The team then employs LM-KT to specify the objective and data for training a model that generates questions based on the student's abilities and target difficulty. Their results show that they successfully generate novel, well-calibrated language translation questions for second language learners from an actual online education platform. The generated questions cover various levels of difficulty ranging from very easy (0.9) such as "spoon or tea" to more challenging ones like "translate 'the woman'". This approach has significant potential in providing personalized educational experiences for a broader population of students.