Author |
Topic: the web as a killer of novel applications (Read 8096 times) |
|
amichail
Senior Riddler
Posts: 450
|
|
Re: the web as a killer of novel applications
« Reply #25 on: Aug 9th, 2006, 3:10am » |
Quote Modify
|
on Aug 9th, 2006, 2:46am, towr wrote: Unless you follow a conversation that was held before near perfectly, it would yield a very disjointed dialog. There is simply too little understanding in current systems. They don't understand context or references. Generally you can't even count on them remembering anything you said before. Just repeating sentences said before won't solve that. The single sentence they repeat may be a natural one, but often its context will fail. Aside from that there's the problem a computer can't a priori know every word a person might use. So it has to be able to learn, and deduce the meaning of words from their context. Not to mention it has to cope with people's typos and grammatical errors; and outside specific domains things like metaphors, sarcasm etc. The field is rife with problems that make natural language processing and generation intractable. At least for the current approach. (I suppose technically you could call people automatons, so a similar automatic system ought to be able to cope. But I doubt we could still understand a system complex enough to deal with it. The computational principles behind it would be as insightfull as the physical description of the brain on a subatomic level. i.e. not.) |
| But what happens if we use human computation? For example, to automate myself, I could build a system with everythng I have said publicly online in its database. During a conversation, a game would be played by human players in real-time. They would compete to pick the most appropriate reply to the conversation from the database. A scoring system could be used to reward those who picked good replies, etc.
|
|
IP Logged |
DropZap - a new kind of block elimination game
|
|
|
towr
wu::riddles Moderator Uberpuzzler
Some people are average, some are just mean.
Gender:
Posts: 13730
|
|
Re: the web as a killer of novel applications
« Reply #26 on: Aug 9th, 2006, 5:13am » |
Quote Modify
|
[comment]hmm.. finally found what gave the "error 500" problem [edit]but on second thought, perhaps I shouldn't mention it here, in case it's exploitable..[/edit][/comment] on Aug 9th, 2006, 3:10am, amichail wrote:During a conversation, a game would be played by human players in real-time. They would compete to pick the most appropriate reply to the conversation from the database. |
| Foregoing for a moment the problems a player would have to choose a response from such a large database, I very much doubt there will in general be a response that fits perfectly in the conversation. Interpolation may work reasonably up to a point (some response might be close enough to fit sufficiently), but extrapolation is always the biggest problem of computer learning. Even with a closed domain, you'd need to have said everything there is to say about it, and not only that, but in ways appropriate to any sort of questioning of it. (Alternatively you'll need a wide variety of ways to say "I don't know", without sounding like a robot. There isn't even a decent chatbot that can deny knowledge of everything in a natural way) And, the players would also have to find the appropriate respons to any sort of question applicable to the domain (keyword spotting simply won't cut it; it scarcely even deals with negation appropriately, let alone more complicated sentence structures) And that's just dealing with question-response, not even a conversation. In general people in a dialog build on earlier parts of the conversation. So to find the appropriate response can depend on the whole previous conversation.
|
« Last Edit: Aug 9th, 2006, 5:22am by towr » |
IP Logged |
Wikipedia, Google, Mathworld, Integer sequence DB
|
|
|
amichail
Senior Riddler
Posts: 450
|
|
Re: the web as a killer of novel applications
« Reply #27 on: Aug 9th, 2006, 5:38am » |
Quote Modify
|
on Aug 9th, 2006, 5:13am, towr wrote: Foregoing for a moment the problems a player would have to choose a response from such a large database, I very much doubt there will in general be a response that fits perfectly in the conversation. Interpolation may work reasonably up to a point (some response might be close enough to fit sufficiently), but extrapolation is always the biggest problem of computer learning. Even with a closed domain, you'd need to have said everything there is to say about it, and not only that, but in ways appropriate to any sort of questioning of it. (Alternatively you'll need a wide variety of ways to say "I don't know", without sounding like a robot. There isn't even a decent chatbot that can deny knowledge of everything in a natural way) And, the players would also have to find the appropriate respons to any sort of question applicable to the domain (keyword spotting simply won't cut it; it scarcely even deals with negation appropriately, let alone more complicated sentence structures) And that's just dealing with question-response, not even a conversation. In general people in a dialog build on earlier parts of the conversation. So to find the appropriate response can depend on the whole previous conversation. |
| It would be known to everyone that the conversation is being constructed on behalf of someone via a human computation game. To avoid the problems you mention, some guidance can be given as to where the conversation can go. This can take into account the database as well as the areas of expertise of the players.
|
« Last Edit: Aug 9th, 2006, 5:39am by amichail » |
IP Logged |
DropZap - a new kind of block elimination game
|
|
|
|