Hacker News new | comments | show | ask | jobs | submit login
Deep Language Modeling for Question Answering Using Keras (benjaminbolte.com)
106 points by fchollet 416 days ago | hide | past | web | 6 comments | favorite



I personally find Google's natural language abilities impressive sometimes but often less useful than other features that Google removed from it's search (like the ability to do exact phrase search - if not removed, it's so disfunctional now it might as well be removed).


Regarding Google's natural language abilities, I think that if they don't develop a decent chat bot soon, much more natural and conversational than Google Now, they will lose to Facebook.

On the other part if they deliver, they will have a new medium that is amenable to advertising and is fit for mobile use, unlike the text search box. That means they have to be able to answer in human language to any query and remember the whole conversation context.

Right now I am trying hard to learn how to command Google Now to do useful things. It's hard and I discover new tricks every day. I think it is extremely limited at the moment, and poorly integrated in the system. I want to control the apps, to add my own words/phrases/scripts/external apis to it, I want it to fake a believable casual conversation and to have a little more personality.

At some point in the future audio recognition will be good enough (if it isn't already) and understanding of meaning will have become useful enough to make it easier than any other way of interfacing to a computer when giving various commands or searching. Then everyone will adopt it en masse.


Regarding Google's natural language abilities, I think that if they don't develop a decent chat bot soon, much more natural and conversational than Google Now, they will lose to Facebook.

Lose what to Facebook? I suppose there's a "chatbot" war now but so-far Facebook's chatbots have opened to terrible reviews. The chatbot market has a long way to go and I suspect it's one place where best will quickly outpace first if things do get going.

Google controlled the search engine market before they AI-ified their search engine and find myself usually using it the same way though occasionally jumping to full questions - which it does OK at. So here it's hard to tell whether AI actually help Google compete.

At some point in the future audio recognition will be good enough (if it isn't already) and understanding of meaning will have become useful enough to make it easier than any other way of interfacing to a computer when giving various commands or searching. Then everyone will adopt it en masse.

The rise of text-messaging shows that text is often the superior medium even for people who fully understand natural language. I'd rather send or receive text than deal with voice mail - phone calls for people I care about or things that are hard to understand.


I try to avoid using chatbots from major companies, because they gain way too much information about you via those means. The voice-based chatbots, as commonly implemented, have the ability to listen to everything you say.

I hope Google doesn't end up in the same boat Microsoft is in with Cortana.


Well, OS X already has a form of offline voice-control, so obviously an internet connection is not absolutely mandatory for these things.

It's a bit buried though; you have to go to System Preferences -> Dictation & Speech and turn on Use Enhanced Dictation, then go to System Preferences -> Accessibility -> Dictation and Enable the dictation keyword phrase.

By default it's "Computer" but it can be "Hey Siri" or anything you want, so you can say things like "Computer, switch to Safari" and "scroll to the bottom" etc.

I wonder why Apple hasn't advertised this more prominently. It seems to be an edge over whatever's available in Windows.


Nice tutorial-esque writeup as suggested by title. Keras makes neural net ML so easy to play with... Would have been nice to see a bit more exploration of the results though (examples of input and output, how significant embedding and attention are to this model, etc.).




Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | DMCA | Apply to YC | Contact

Search: