Show HN: DeepQuiz – free NLP question generator | Hacker News Reader

Show HN: DeepQuiz – free NLP question generator

deepquiz.com.s3-website-us-east-1.amazonaws.com

Hi,

Here is an extremely early stage project that a friend and I have been working on this summer. DeepQuiz generates quiz questions from user-submitted text. Currently we're using a combination of statistical analysis and heuristic rules to do this - no machine learning. The main point in its favor right now is that it's free to use and pretty fast.

We would really appreciate your feedback, as well as ideas for how it could be better, and types of generated questions you would like to see. The system does best with English text.

Nice work! My suggestion at the moment is to avoid excessively long true/false questions (unless you actually plan to modify them, but that can be incredibly hard), and to use Named Entity Recognition/detection and semantic similarity (Word2Vec/Glove) for multiple-choice questions. I guess you are already using the former, but with the latter you should be able to really constrain the options to something more "credible".

This is a pretty cool project! You should take a look at this paper : https://arxiv.org/abs/1705.02012 . Using your heuristics to select the answers, you maybe able to generate some non fill in the blank questions. (All I saw was fill in the blank, not sure if it does other types of questions)

It also does true/false and multiple choice. Right now the text sample has a lot to do with what kind of questions it can find.

This looks cool! As you probably already know, it has some ways to go. Simple English Wikipedia would have plenty of content with sentences that are easier to parse. Also, it'd be pretty cool to see general trivia quizzes autogenerated from Wikipedia using your algo.

The problem of trivia quizzes is selecting what makes a good trivia quiz.

You can't simply "quizzify" any part of any Wikipedia page, or the quiz has a very high chance of being too hard or not interesting at all (trivia: how many people live in the German city of Kraiburg am Inn? 600, 4000, 8000 or 16000?. Even someone interested in German geography is probably not going to enjoy this)

This is so cool! What kind of tools are you using for the heuristic rules?

This is exactly the sort of project I wanted to implement, but linked to more than one data source. I was thinking wikipedia + stack overflow or something like that. My dream is to create a system like this one which dynamically creates questions in a media-wiki article. I think we need to think about categories of question. True / False and multiple choice are only one part of "knowing something".

To what extent does someone know something? This is probably a better way to approach asking questions. We also think about "blooms taxonomy of knowing" which is helpful for thinking about different types of assessment tasks.

I would imagine generating the questions would be first, and then ranking them according various tiers of knowledge.

Guess, FB Research's DrQA is very close to what your idea is. https://github.com/facebookresearch/DrQA

I have lots of text about a particular topic (50K Articles crawled from internet about travel experiences). Now I want to create a Q&A from this given text.

Is there any service that can do it? Basically I want to feed my text to this service and it creates a Q&A pairs from this given text. Not sure If I made it clear..

Nice project, a group I worked on did something very similar with retrieving questions from a syllabus. Is this open source? I'd be fascinated to see similarities/differences in our approaches.

Is you project open source?

This could be useful to avoid comment spammers, trolls in blogs.. are you looking into this? A plugin could be a good start

Do you by any chance have an API one can use ?

Yeah, just look at the Web page source to get the URL. Content-Type must be text/plain.

Reply on news.ycombinator.com