Hacker News new | past | comments | ask | show | jobs | submit login
An AI app that generates quizzes from a photograph of a texbook (questo.ai)
142 points by taixhi on March 21, 2018 | hide | past | favorite | 44 comments

This looks wonderful, but it isn't a Show HN if people can't try it out. That's described at https://news.ycombinator.com/showhn.html.

Once it's possible for users to try out, please do post it here. I'm sure many will be very interested. And email us a link at hn@ycombinator.com so we can make sure it doesn't get flagged.

thank you for your response! We originally posted without "Show HN", don't know what happened there...

Ah sorry! Probably a moderator added it while trying to be helpful. In fact it may have been me :)

OP, I'm going to go out on a limb and assume you are Taichi Kato? If not, I'm assuming he, Arya, and Khush will read this discussion.

The three of you should be very proud of yourselves. This is a really cool, novel idea! The fact that you all are sophomores in high school is very impressive too. Continue to follow your passions and you will surely be very successful.

Haha, I am Taichi. Thank you so much!

I'm Arya, thank you!

This made me think of the Norwegian news site that required readers to take a quiz before commenting: https://www.independent.co.uk/life-style/gadgets-and-tech/ne...

Combining these two ideas could be really neat -- I wonder if it could even be sold as a service a la disqus.

Great idea! I am very guilty of commenting here without reading the article. Can we try this here?

Hah, I was going to post the same thing. Perhaps it is steam-engine-time for this idea.

The idea would combine nicely with voting in the elections, too - given a casted vote has more impact than leaving a comment under a news article ;)

The days of voting literacy tests are not exactly a bright point in American history.

Every thing you can think of has been misused at some point in (not necessarily American) history.

That's very true. I should clarify that in addition to its history designing a test to selectively disenfranchise voters seems not the best to me.

It’s all about context, though.

Do I have to sign up to see a demo? Where is the "show"?

Does the world need a large amount of highly automated but poorly designed quizzes?

It addresses a poorly designed form of education.

Until we address the stupidity of regurgitation of facts as a means of learning, if this helps someone generate flash cards to prep for exams, it's a positive thing.

How to build better instruction is already a known science.

This is fantastic. Congrats to the team! Education is an area that is still behind on using tech to dramatically improve the learning process so I'm very excited to see more work in this area.


I was looking into doing something similar for a side project and got discouraged by my lack of a strong ML background. This is both great work and inspirational. Time to go dust off that repo..

This looks like an amazing tool. As someone who often uses courses that have premade notes but highly limited exercise and quiz material, this looks like a serious blessing (assuming it works). Thanks!

Unfortunately, my courses seem to have the opposite problem, but at least there's always Wikipedia.

Wow! I love this idea. Congratulations! One question I have is what are you doing with the textbook pictures? Also, does the user categorize where the pictures came from so other people studying the same text can have additional questions?

Are you legally allowed to copy the contents of these textbooks? If I recall there is often on the first page of textbooks long legal texts which prohibits copying of any kind without written consent from the authors.

What our algorithm does is, we convolute the text from the textbooks to create questions such as "How did Hitler die?" from texts like "In August 1945, Hitler killed himself." from the textbook. So usually, that ends up being significantly different to the original texts in the textbook.

Copyright laws typically prohibit both identical copies, which appear to be done here by users during image processing / OCR even if only transitory, as well as derivatives of the copy. So the question then becomes, is this "fair use"?

If the use case for the app is personal use by the possessors of the textbook, very likely no problem. For all other uses you should consult a lawyer to make sure you're in the clear.

> as well as derivatives of the copy

Shouldn't this imply that it's illegal for me to make quizzes on my own from information I found in a textbook? If this app is implemented correctly, what it does should be indistinguishable from a human. It makes no sense to change the infringement status of a work based on how the work was produced.

Also, generally hard facts (like when Hitler died or how Hitler killed himself) are not copyrightable.

I don't know how you are doing this, but with NNs there's still be the chance that result contains significant parts of the text-books. I am not a lawyer, but because you are only interested in the syntactic equality (the sematic being the same) i think a simple algorithm like something based on edit-distance may be able to exclude such cases.

IIRC and IANAL, but, copyright laws are about reproducing text, e.g. taking this text and republishing it without consent of the copyright holder. In this case that does not apply; while the app does copy the text (and possibly sends it to their servers), this text is not republished.

Copyright law apples, not the first page legal looking text. They can make any claims they want to on the first page, but that has no legal standing in court. This quiz almost undoubtedly fall under fair use.

Of course if you need legal advice you need to consult a lawyer.

Would you consider this a "Derivative Work" in a legal sense though?

I would consider it an education use which puts it into a whole other area of copyright law.

I don't think it's copying verbatim, it might be caching it for a short while but not saving it (based on what I'm reading about the technology).

So the more interesting question here: is a computer legally allowed to generate an original interpretation of text on a page?

The user of the app is still copying the text verbatim and transferring it to the server, which is distributing.

Not exactly.

A. They probably cache only for a little while B. They aren't republishing it

The company isn't. The user is still distributing.

Great idea! Congrats!

I have bad experiences with google translate, but this seems like a simpler problem waiting to be solved.

Can you share any information of the percentage of question/answers that are generated correctly?

SUGGESTION: When I click on a link labelled "press," I expect to see a page with media mentions, not my e-mail client's Compose New Message window.

Would any of the authors be willing to comment on how questions are generated (i.e. what algorithm)? Asking as somebody with a casual interest in AI

Our algorithm roughly parallels NMT systems, and is mostly based on seq2seq.

I would suggest adding a picture of the generated quiz. The images there don’t really tell me much; perhaps that’s intentional.

Great suggestion, we didn't even realise that! Check back soon and they'll be there, thank you!

Does this work on any document that you pass into it? How does it handle questions for ideas that continue to the next page?

How does this work ? Any descriptions of it would be very informative. Also congrats

What kinds of questions would be generated for the Hitler Reichstag fire example?

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact