| |
| |
| |
|
Page: 1 2 3
Comments:
<0> cwenner: So on a different topic, what kind of AI topics do you study normally?
<1> it might be easier to ask him what AI topics he doesn't study :)
<2> right now NLP but i try to study the E-E problem as well
<1> what's E-E?
<2> sorry, the exploration-exploitation dilemma/problem
<2> this might interest you two: http://www.lri.fr/~gelly/paper/nips_exploration_exploitation_mogo.pdf
<2> Exploration exploitation in Go: UCT for Monte-Carlo Go
<2> (haven't read it myself)
<1> by the way, i'm trying to get an interview at www.aetion.com
<2> hm, abductive inference
<0> cwenner: yeah, several people on the computer go mailing list are researching monte carlo engines
<0> It's apparently one of the more promising approaches.
<0> chessguy: cool. good luck
<2> i did not see it that much a year ago, is it new? i sort of like it, i've experimenting with ant-algorithms for node expansion, quite similar
<1> thanks. it's most definitely my dream job
<0> yeah, I can imagine
<2> the job looks interesting. hopefully you don't get stuck with GUI or testing :)
<0> hey, even that's a foot in the door
<1> Ceran, i agree
<1> i'd be thrilled with that
<1> sure beats what i'm doing now
<0> Interesting, they use Haskell.
<0> Although they also use a lot of open-source technologies.
<0> chessguy: Can you program in Haskell?
<1> i've been known to try :)
<0> heh
<0> They recommend you submit Haskell code submissions with your application.
<1> yes, i'm working on that now
<1> well, not right this minute
<2> two interesting ***ignments from our haskell course: Eliza clone http://www.cs.lth.se/DAT116/***ignment1 and melody accompaniment http://www.cs.lth.se/DAT116/***ignment2/
<1> i've heard good things about haskore
<2> the first one is of course just pattern matching but still rather fun
<0> cwenner: Are you working on any particular projects as part of your general study?
<2> right now the RTE system and i also toy with a small statistical texas hold'em bot
<0> RTE?
<2> Recognising Textual Entailment - http://www.pascal-network.org/Challenges/RTE3
<0> Is there a page on the wiki for listing what topics people are interested and/or currently researching?
<0> cool
<2> used to be one for #machinelearning but not for #ai. the #machinelearning wiki is gone though, was administrated by yaroslav
<2> create one though, i would surely join
<2> my humble (not very good) paper: ainerd.com/elfrte.pdf
<0> will do
<2> have not uploaded the new version yet but comments, even harsh such, are most welcome
<0> To be honest, I don't think these give-computers-common-sense projects are worthwhile.
<0> Unless they involve making computers figure out common-sense on their own.
<0> It would be interesting if someone were to create something like Cyc that could crawl the net and create a corpus without human involvment.
<2> that's pretty much what you work towards but as we know, we must make some ***umptions and we want to boost it as much as possible. if i am to actually continue on this topic long enough, my goals are to add probabilistic inference and to collect information like the TREC Q/A systems do (although hopefully deeper)
<0> How do you add pages to the wiki?
<2> go to ai-freenode.wikidot.com/somepage
<2> although we might have the constraint that you must register to add pages
<0> yeah, I'm registered
<0> as cerin
<0> wikidot is pretty slick
<0> kicks the crap out of wikipedia
<0> I'm going to create a page called "interests", where we can post our topics of interests
<2> RTE is not so much about adding common knowledge as to recognize entailment :), i.e. if some text implies another (or in my case, logical expressions, such as entails?(message, fol-exp("verb_attack3532...('x1, 'x2, 'x3), me('x2)")) )
<2> common sense*
<2> common sense is rather important for specific such but only a few rules have huge impacts, such as in(x, y) & in(y, z) -> in(x, z)
<0> ok, it's up: http://ai-freenode.wikidot.com/interests
<0> feel free to add
<2> do you frequent the Computer Go room at KGS?
<0> cwenner: I'm unsure symbolic rules can sufficiently model common sense.
<0> cwenner: No, I don't have a functioning engine yet.
<2> did i mention that i have 22000 dan go games by the way?
<2> i meant more to find people to discus with
<0> well, not a functioning engine that's any good ;)0
<0> cwenner: nice. I found a few thousand pro games myself, although I plan on focussing on RL through self-training rather than supervised learning.
<2> it does not need to be symbolic in this case. the best system of last year was a ML that works on tree-similarity (with the help of various common sense knowledge however), the second system used logical inference. both written by the Language Computer Corporation
<2> that will be very interesting to see :)
<2> it is quite common to extract patterns and to use them for evaluation
<0> Are you now talking about RTE or Go? :)
<2> but they usually only have specific patterns that are rather localized
<2> sorry, both
<0> heh
<2> the pattern part is where it turned to go
<0> RTE is interesting. I've never heard about it before. I'll have to read up on it.
<0> Am I correct in thinking that it's basically an approach to figuring out the "semantics" problem in NLP?
<2> first challenge was in 2005 but the TREC challenges are older. the QA task there is very interesting and i almost picked that one instead
<2> (although the systems there are even larger so this one was better for me)
<0> Is it required that code be in Pascal?
<2> they have huge amounts of data, so called TREC-discs where answers to various questions can be found explicitly or implicitly. the goal is to be able to answer as many questions as possibly correctly
<2> this involves going through the data and generalizing it (cause you cannot possibly keep the entire parse trees for all parts), the most advanced systems use probabilistic reasoning and usually a so called "web boost", where they also try to find the answers online
<2> (the simplest ones just do google searches and checks the number of hits :p )
<2> PASCAL is unfortunately an acronym for "Pattern Analysis , Statistical Modelling and Computational Learning"
<2> RTE and other AI/ML/NLP resources are supported by this "PASCAL network"
<0> heh, an unfortunate acronym indeed
<3> the problem with the stanford lexparser and the link grammar parser is that they don't learn dynamically, I want to tell it: 'you got the object wrong, this is the correct object' or 'the correct constituent tree is ____'
<3> the stanford lexparser gets "the horse raced past the barn fell"' wrong
<3> LexpAgent: please diagram "The horse raced past the barn fell."
<4> (ROOT
<4> (S (`` ``)
<4> (NP (DT The) (NN horse))
<4> (VP (VBD raced)
<4> (SBAR
<4> (S
<4> (NP (JJ past) (DT the) (NN barn))
<4> (VP (VBD fell)))))
<4> (. .) ('' '')))
<3> hmm maybe i should filter out the unnecessary lines.
<0> trane: What output are you expecting?
<0> What's a "barn fell"?
<0> remember, the principal of GIGO applies to NLP as much as anything ;)
<3> Ceran: the horse raced past the barn fell is a cl***ic linguistic example of a 'garden path' sentence
<3> the horse fell
<3> raced past the barn is a phrase modifying horse
<3> you'ld think the folks at stanford would've tested their parser on cl***ic linguistic example sentences :)
<0> I'm fairly talented in English, and that sentence sounds like gibberish to me. The correct grammar would be, "The horse racing past the barn, fell.
<3> The bird found in the room died is another example
<0> that's a better one
<3> it might be somewhat of a contrived example but it is legal english and i suspect sentences like that occur naturally as people add things to what they're saying that can be ambiguously parsed...
<0> "legal english" is a somewhat useless term. The number of "legal interpretations" of a sentence grow exponentially with sentence length. The trick is finding the one interpretation that is most common, and that's subject to experience.
<3> well that's precisely what i want to add to the parsers, a way to give them experience, to teach them when they make a mistake
<2> the legality could be a fuzzy term :)
<0> That said, if I heard a sentence phrased like "the horse raced past the barn fell', I would interpret the subject as "barn fell". However, if you rephrased it as "the horse racing past the barn fell" or "the horse that raced past the barn fell", then I would interpret it as you wish
<2> trane: they do learn, they are statistical. although you would need to retrain it with the counterexamples which would probably require you to get a hold of the original training data
<3> the point of garden path sentences is that you are led down a garden path of parsing until it doesn't make sense and you have to go back and reinterpret the parse. The particular example might be contrived but is very very common in linguistic texts, and just representative
<3> cwenner: sure, i want to just tell it as i'm interacting with it though and have it retrain or whatever in the background...
<0> trane: naturally, when I think about the sentence semantically, I would realize that the sentenance is really "...barn, fell", but now I'm ignoring grammar and relying on semantics. I'm unaware of any lexical analyzers that can handle semantics.
<2> i think you are wrong about that. the parsers are not made by couch linguistics but is learned through examples. if it fails, the model is either inaccurate or it lacks that data. the data should be rather well distributed but could very well still not be enough.
<3> Ceran: interesting. the cl***ic linguistic interpretation is that it's a syntactic issue. the link grammar parser gets it right for example
<0> trane: until someone invents something called a "barn fell" and then you're screwed ;)
<3> hehe
<2> does not the stanford parser come with these huge data files? if you know the format of those, it should be possible to modify it rather quickly, even though it might not follow the model accurately
<0> "fell" can mean ghost, so maybe it's the ghost of a barn?/
<2> (i.e. perhaps an expected worsening but improvement for your needs)
<3> cwenner: i might try that, but the way i'm thinking right now is just to correct it online and store the corrections in a look-up table, kind of a hack, sure, but...
<3> then later i (or the program itself) could go back and add the new data to the data files and retrain
<3> the point is that language is not static, but the data files are, once the program is running
<2> you are frequently warned to fix specific samples yourself though, cause there are simple too many of them
<3> i would just prefer a way of fixing those new samples at runtime instead of editing a data file and restarting
<2> would need a different system then
<2> yikes, MXPOST cl***ified "racing" in "The horse racing past the barn fell." as a noun. for the race example, at least it identifies past as a preposition but the subject is the barn
<2> The logical form became: horse (x1)race (e1, x1)past (e1, e2)barn (x2)fall (e2, x2). while the correct would be horse(x1) race(e1, x2) past(e1, x2) barn(x2) fall(e2, x1)
<2> <race/VBD SUB:<horse/NN NMOD:<The/DT >>, VMOD:<past/IN SBAR:<fall/VBD SUB:<barn/NN NMOD:<the/DT >>>>, VMOD:<./SENT >><The/DT ><horse/NN NMOD:<The/DT >><race/VBD SUB:<horse/NN NMOD:<The/DT >>, VMOD:<past/IN SBAR:<fall/VBD SUB:<barn/NN NMOD:<the/DT >>>>, VMOD:<./SENT >><past/IN SBAR:<fall/VBD SUB:<barn/NN NMOD:<the/DT >>>><the/DT ><barn/NN NMOD:<the/DT >><fall/VBD SUB:<barn/NN NMOD:<the/DT >>><./SENT >
<3> that's an interesting way of representing the sentence, i don't quite see how i can use it to get subject verb object agent etc. yet though
<2> oups, was posted twice, take this one instead: <race/VBD SUB:<horse/NN NMOD:<The/DT >>, VMOD:<past/IN SBAR:<fall/VBD SUB:<barn/NN NMOD:<the/DT >>>>, VMOD:<./SENT >>
<3> ah it has a subject tag
<2> my method for finding the subject is described in the elfrte.pdf paper. if there is a subject, take that one (special rule for noun-groups)
<2> otherwise, do an "ordinary" searh which is described in 3.3.2, but since i'm worthless at describing...
<2> for finding the subject if there is no SUB relation (and this is just empirical and approximate, still it yields some % extra avg. over all arguments, not just subjects): check siblings before itself in the current node, if none is found, check children, if none is found still, check siblings after the node itself
<2> if that too fails, check the parent
<2> unless the parent is a noun in which case it's checked just before teh children
<2> sorry, other way around. after all if noun, before children if not a noun
<3> can you also get object, agent, verb?
<2> that actually adds 1.1%
<2> i only consider 2 arguments mandatory, all other modifiers use their own modifier-nodes
<2> i don't know how well they operate, we could try if you have an example but i have not run any benchmark
<2> this is just the best i have found by experimenting on a development set
<2> although for the logical forms extraction i got good results so it seems fine
<2> (my RTE results are poor still)
<2> (good results on the test set)
Return to
#ai
or
Go to some related
logs:
sourceomatic
nuller efnet
#perl
ubuntu gbindadmin
asimptotes
#debian
pedora core 5
vestige in php
netatalk ubuntu avahi
gentoo overlay knetworkconf ebuild
|
|