There are lots of times that we don't know if we truly remember the key points from the textbook. A very effective way to do check that is by taking a quiz about the materials. However, it is too bad that if we realize that we didn't study well during the actual quiz. So InstantQuiz can help students to summarize an article and automatically generate review questions.
What it does
Given an input article, InstantQuiz can capture the important sentences within the text. And turn them into factual questions so the users can test themselves.
How I built it
First, we use TextRank algorithm to summarize the entire article to a couple of sentences. We as a seq2seq model from deep learning to generate the questions base on summarized sentences. We trained our model on The Stanford Question Answering Dataset: https://rajpurkar.github.io/SQuAD-explorer/ which contains human generated questions and solutions on the Wikipedia pages.
Challenges I ran into
This model is quite new in the research area so it is very complicated to implement and train. And we are a little bit underestimated the difficulty of the problem at the very beginning.
Accomplishments that I'm proud of
The model actually works and the questions look good and reasonable.
What I learned
To find the model, we read dozens of papers to find the best one. We also use lots of NLP algorithm in our work. We also learned a lot about making and running a website.
What's next for InstantQuiz
InstantQuiz now cannot do automatically information retrieval. Which means the users can't copy a Wikipedia page link and make it generate questions. All the demo cases we have so far are carefully formatted by a human. Also, we don't have a chance to check the result on actual textbooks. There are further experiments we need to do on those data.