| |
| |
| |
|
Page: 1 2
Comments:
<0> Andares: the "meaning"? There are a number of clustering algorithms that can extract, I would say, information, based on the contents of the sentence
<0> That information being related to the clusters it had formed in whatever training corpus it was provided
<0> But I really don't know NLP
<1> Hmm, I keep wanting to ask Chessguy about his GP but he keeps not being here when I try to ask him!
<2> What's a GP?
<2> Ugh.. I lost my link to a backpropagation tutorial.
<1> man, I'm just going to post my backprop code
<1> It would answer 50% of this channels comment/questions
<3> the code is not exactly self explanatory
<3> and as far as I know chessguy doesn't have finished code for the GA part yet
<3> but he has ideas
<4> aha? whats this all about then?
<3> GA to make a chess computer
<2> What is it with chessguy and chess?
<2> er.. stupid question.
<3> :)
<5> hmm
<6> does anyone know anything about NLESB (Natural-Language Legal Expert System Builder)?
<7> you guys talking about me?
<6> or about using nlp (natural language processing) to feed an "AI"?
<8> nlp is very bad.
<8> well.
<6> why?
<8> it's pretty good at some things.
<8> but they're very limited domain. mostly verifiers, a few odd sorts of generators.
<8> why is it bad? because the state of the art is poor :p
<6> I don't think so
<8> why?
<6> "natural" language can contain lotx of information
<8> no ****. but the state of the art for extracting that information is laughable.
<6> not for all languages
<6> japenese is quite simple for nlp
<8> ... riiight.
<6> even german is easier to parse than english
<8> Parse, yes.
<8> But syntax is not the issue.
<8> The state of the art is very good at syntax, even English syntax.
<8> but all that actually gets you is stuff like grammar checkers.
<6> pos-tagging is just one step
<8> yes, but it's the only step we've had for quite some time now.
<6> no
<6> do you know nltk.sf.net?
<8> I didn't. I'm not seeing where it helps with the problem of semantics, though.
<8> in fact all it really seems to have is clustering and postagging. useful if you need that, but not, y'know, going to provide understanding.
<6> what does understanding mean for you?
<6> words by letters, sentence by words, text by sentence, need a textgrammar
<6> and the valency can help (http://en.wikipedia.org/wiki/Valency_%28linguistics%29)
<6> http://www.coli.uni-saarland.de/projects/sfb378/2002-2004/projects.php4?l=en
<8> the semantics. what does it "mean". what are referents, /why/ was the utterance made. etc.
<8> the same thing meant by understanding when talking about people.
<8> valency is all about parsing. syntax.
<8> syntax is /doable/.
<8> syntax is not the problem.
<8> the problem is that syntax only gets you so far.
<6> german syntax contains semantic information
<8> and? is it enough to get anywhere?
<8> I doubt it.
<8> I very strongly doubt it.
<8> I'm sure it makes parsing easier by reducing ambiguities ... but there's no way it can possibly eliminate them.
<8> actually that might be a decent definition of understanding: being able to resolve ambiguities in the way intended by the speaker.
<6> do you know the difference between N-, R-, O-Grammars?
<8> Don't recognize the terms, no.
<6> N-Grammars make "natural" language (empiric)
<6> O-Grammars are objectorientated
<8> so?
<6> R-Grammars are facts-of-case-orientated
<8> how does this help the problem that understanding language is beyond the state of the art?
<6> "rational"
<6> a complaint for example should contain: who did where when what, how, with what, why
<6> ever thought about what "law" PEOPLE DO?
<6> sry for caps
<6> they define what is in a word and what is not
<6> R-Language
<6> but that R contains a "real-world-model"
<6> since the whole state runs on R not on N
<6> In law text you always know the context
<6> context reduces ambiguities
<9> <perthmong@ef> hey
<10> hELLO
<10> Good to see u all
<11> I think true language understanding requires a complete cognitive architecture with knowledge representation
<8> I would agree.
<11> =)
<11> Syntax is just a tiny part of understanding
<11> Maybe 10%
<8> pretty sure it's not a linear combination, but.
<11> I've read that on average syntax contains about that much information
<11> Sotek what kind of AI applications are you interested in?
<8> I'm actually not sure anymore.
<11> Heh ok
<11> How about common sense reasoning
<11> I'm thinking about that
<8> hmm.
<8> common sense is really just generalization.
<11> There's a large knowledgebase I can use from Cyc
<2> Hey, any of you ever use WordNet?
<2> rrgh.
<2> http://www.ei.dtu.dk/teaching/04364/ex5/ex5.pdf#search=%22Backpropation%22
<2> Can anyone help me grok the math?
<3> Andares, that doesn't look like good teaching material. some writers write for themselves and oblivious to the notion that the reader might not know the stuff already
<3> didn't I send you the pages from mitchell's book?
<3> not that he's a lot better though
<10> Hey
<12> welcom
<2> DanF...
<13> Andares
<13> bbl
<2> Hi DrChandra_.
<2> DrChandra_ = DanF_DRC?
<2> *DrC
<14> who is here?
<2> NOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOOO
<2> Also, hi paros.
<2> I really hope this computer doesn't turn off.
<2> Because I have to back up everything.
<14> hi
<15> hi death
<15> pretty grim name
<16> hi
<15> any reason why youre called _death?
<16> DEATH is taken
<17> nice band, don't like the genre tho.
<16> DEATH is n=DEATH@unaffiliated/death * DEATH
<16> D
<16> oh hm i like industrial mostly
<15> do you guys do AI stuff?
<16> yes
<15> what kind?
<16> i'm working on kernel-weighted polynomial regression atm
<16> i'm planning on training the kernels because i am stupid
<16> also, i'm using missing data, but my approach to missing data is intelligent
<16> i forget what ai people call working with missing data
<16> "robust learning"
<15> uhm i dont know either
<2> kernel?
<16> for local regression
<16> (or local averages if local regression doesn't perform as well, but i expect it to perform better)
<16> i'd like to pull out the kernels analytically, but i'll probably just train them with some slow iterative technique
<15> sorry i dont know what you mean with kernels heh
<15> like kernels for SVM?
<16> i am not really the best person to learn about kernels from
<15> i never got how SVMs work
<16> svms use the same kind of kernels, but i think that mine will have some additional properties that svms will not (and i'm not sure they would be appropriate for svms)
<15> What sort of data or concepts are you trying to learn?
<16> svms draw a line (or a n-dimensional version of a line) between some points and the points can be in a non-euclidean space
<16> the space can be defined by a kernel
<16> i forget what kind of space they call it
<16> game theory, economics, that sort of thing
Return to
#ai
or
Go to some related
logs:
gpartd
#fedora
grub.conf (nd) ifconfig
set_session_handler $_SESSION
#perl
#lisp
cdrecord /dev/pg0 ubuntu
#web
alsamixer sound blaster live center channel volume
dh_testdir alien gentoo
|
|