Slashdot Mirror
Creating a Better Chatbot Through Crowdsourcing
An anonymous reader writes "MIT Technology Review reports on a chatbot built at the University of Rochester that is capable of high quality, human-level conversation — thanks to software called Chorus that turns to Amazon's crowdsourcing service Mechanical Turk to generate and evaluate replies to a human's statements and questions. No one person is ever acting as the bot, instead multiple workers suggest responses that are then voted on to select the best. The crowd workers contributing change frequently, but Chorus also has them keep a running list of important contextual information to give the bot a kind of memory of a conversation's history. The researchers say Chorus-style chat bots could out-perform fully automated assistants such as Siri, while being considerably cheaper than a true concierge service."
Humans better at chatting than computers, and by replacing computers with humans, you can almost have a natural conversation.
Good try, but even if this passes the Turing test, it is not AI in any way. Responses are by humans and there is no intelligence in it. So it will be the collective human engine behind it that will pass the test. Not really any big achievement.
metageek
I can't imagine anyone fluent in the language being willing to work for the prices that could be paid. We're talking pennies per hour.
If video games influenced behavior the Pac Man generation would be eating pills and running away from their problems.
And if this ever takes off you know that it will become a game of trying to submit/vote the most inappropriate responses to the questions. A never ending battle between the censors forbidding words and phrases and the people finding new ones.
Read the subject line
What are you wearing?
Cleverbot isn't AI as it's claimed, and I think it's interesting how they try and pass it off as such. Maybe there's a bit of AI there, but from what I've seen it just connects two random users and then reshuffles every so often so you're getting responses from some other real person than the one before...
I haven't made mine live as I keep tweaking the code but I've made many generations of chatbots. The concept I wanted (and still want) to open up to all is ThinkMosaic. It doesn't have the vote for the best response.. More like, chat with it and then if you don't like the response then edit the logic by which it chose that response... Basically, there is a user statement that is best-matched to a "recognizer" (a text pattern with named wildcards optionally in it). This identifies a "meaning". Each meaning may have multiple "reactions". Each reaction has a "condition", some and, or, not, with parenthesis logic and an "action sequence". The action sequence is a stream of commands including: // this obviously responds with a textual (and/or html subset) message. // this stores a text message in "memory" which is a mainly just an array of strings // this deletes memories entered via the remember command // this acts as if the text (..) was entered as a user statement (can have multiple of these, of course) // this sets up so if the first ".." is matched in the next user statement then it is interpreted as if the user says the second ".."
* say ".."
* remember ".."
* forget ".."
* interpret as ".."
* expect ".." as ".."
Wildcards are named by putting them in brackets (.e.g. I love [food] for dinner). And those can be then used in the command strings listed above (including the forget one).
So it's easy to use and works well.. I've made many iterations of different command sets and this one seems very practical and powerful. But it can also be extended with plugins to add new commands, like one I made to grab sections from wikipedia pages.
Cool.. huh... I think so.. I'll have it up maybe in a month or two.. I development version is up but has quirks right now at http://thinkmosaic.com
User: Where's the nearest pizza place?
Chorus: Penis.
User: Excuse me?
Chorus: YOLO
"I feel like some Mexican food." - 10 minutes later: "40% of us say that there should be a Taco Bell somewhere around. Just look for the bell sign"
Of course news about a fake are Fake News.
Of course chatting with real people will feel like chatting with real people. How is this different from, for example tech support where you get connected to lots of random people?
A fake bot that can do maths!
Probably gets similar results.
I eat only the real part of complex carbohydrates.
Just think: If this technology lives up to its promise, Slashdotters fighting in the great OS Flamewar could create their own automated sentries to fight back with phrases like "reality distortion field' and "rounded corners". Maybe we can finally end all this bloodshed!
"I like to lick butts!" by MobileTatsu-NJG (#32700246) (Score:5, Informative)
Over in IRC, one of the denizens of the channel I frequent has neen running a chatbot that snarfs conversation from all over IRC.
It is like Siri, but drunk and insane, and knows all your secrets.
00:18 ? ascaris
00:18 Ascaris lumbricoides, or "roundworm", infections in humans occur when an ingested infective egg releases a riot on facebook
20:57 bmo, what is the survival length of cats in yogurt
20:57 I don't know, I never tried
--
BMO
Good try, but even if this passes the Turing test, it is not AI in any way. Responses are by humans and there is no intelligence in it. So it will be the collective human engine behind it that will pass the test. Not really any big achievement.
Totally agree.
What's the point of this? Is it to make a better chatbot? Do chatbots have any use beyond annoying chatroom participants and surreptitiously inserting ads?
Is it to study human language? Will useful insights come from this? Will there be results? A publishable algorithm? Will this inform future software packages?
Does this tell us something about AI? Can this bring us closer to machines which actually think?
I'm not even sure there is a purpose here. I don't see any academic value in this project. Perhaps someone more familiar with mainstream AI could shed some light.
This may be a good example of a hive mind implementation. The first being the church, and the second was the corporation. The difference is that a machine might be the parser, instead of another person. There was talk of this kind of machine logic being used by companies in the form of schedulers, automated personal assistants, and management analytic engines -- all based on business policies. Once interconnected, they form a collective intelligence that drives the workers, probably like DNA drives a colony of ants. One might think something sinister, but the results are likely to be unpredictable.
here is my 2 cents.
you are now paying for 7 people to do the job of 1 person, that should know your product or service better that 7 Indian strangers.
In other words your customer is in the hands of strangers paid in pennies, that took you dollars to acquire.
More dumass shit.
That which has both Cybernetic and Organic components... Their "Chatbot" is a Cyborg.
while being considerably cheaper than a true concierge service
A "true concierge" is an experienced individual who knows his clients and can perform all manner of complex tasks requiring intelligence.
Neither Siri nor choose-from-random-trollish-human-responses form a concierge service.
It's like pointing out that a vacuum cleaner and a washing machine are cheaper than a butler.
The government is creating logic for chatbots that are run by NGOs to influence the news cycle in places like egypt, syria, russia, to create a buzz that forshortened the news cycle and would influences geopolitical outcomes, Putin and Mubarak would certaily agree.
I have no proof but is no mystery that government capabilities are frequently 20 years ahead of civilian and I participated in a study to this effect (to help tune the math and logic) that anticipated by several days the news cycle. So this is just a logical extrapolation...
Something to consider when conversing on 4Chan, Pastebin, etc.
I guess it's old news, but this sounds exactly like what was being described in Neil Stephenson's Diamond Age. Actors there were paid to read/act short pieces of text/commands to reply to a young girl's questions. In the story, the girl was asking a book to explain a concept to her. Not much different from what happens with a chatbot.
I guess this might also relate to the earlier post on online math courses. Presumably grad students could be given micropayments to answer specific questions for an online course that the teacher doesn't have the time/inclination to answer.
My real concern in using this sort of thing for important information/decisions is how the answers get moderated quickly. Wikipedia has a pretty good, though certainly not perfect, way to deal with this, but it's not necessarily fast enough for real time issues.