Hacker News Reader

Comment by tspike | Hacker News Reader

This could be amazing as input to a question answering engine.

I actually built an IRC bot that did this once. It searched reddit for your question, took the first result, and posted the top comment. It worked very well for certain kinds of questions. Especially if it came from the better subreddits like askscience, but even just reddit in general.

I improved on it a lot with a whitelist of subreddits and some machine learning to select the best thread. But I was only touching on what is possible with that data.

The scope of the discussions on reddit is huge. Despite a lot of jokes in the comments here, the quality of the comments on average, is pretty decent. And the metadata like subreddit and score are extremely useful for filtering it down more.

Sample of conversation with it. Not cherry picked, just showing some of the good and the bad:

https://i.imgur.com/LDD9isL.png?1

awesome! how does one get started programming 'intelligent' things? I know how to program, haven't coded anything that is 'intelligent'. I would really love to know what all the things/concepts I should know to build a bot like yours. Thanks!

This is awesome, I love the "current drama on reddit" answer..

I'd be very interested to learn more about your project and findings. Is that bot still alive?

I turned it back on for the time being. It can be found at irc.snoonet.org at #mybots

https://kiwiirc.com/client/irc.snoonet.org/mybots

EDIT: Reddit changed their api, give me a moment to fix it.

EDIT2: It works now!

AMAbot - is that the one?

Yes. It automatically replies to anything you say in the chat on that channel.

If it's open source I'll love to take a look

I pastebined the code. It's terrible. I should rewrite it. But here it is: http://pastebin.com/CM9u17jq

Nice, I see you use neural networks, can you explain a bit how you are training them?

It searches reddit for whatever the user queries. Reddit returns up to 100 threads. I then choose a thread and take the top comment.

The neural network predicts which thread is most likely to produce a satisfying answer. The main features are number of n-gram matches with the question, the score, the number of comments, and some other metadata.

It's far from optimal but it does improve it a bit.

I think Reddit was one of the places they had to wall off Watson from data-mining, because it devolved into foul-mouthed memes. I wish they hadn't, nothing would be better than a billion dollar piece of technology deciding to reenact Sean Connery Jeopardy skits, on Jeopardy.

[2001 HAL voice]: I have a question about the penis mightier. Does it work?

It's not the hard to filter put the foul-mouthed content

But what about phrases with double meanings (aka: Fun With English):

Children make nutritious snacks

Not necessarily an issue in the grand scheme of things.

I would love to feed my megahal with that.

Reply on news.ycombinator.com