| |
| |
| |
|
Page: 1 2 3 4
Comments:
<0> it's horrible when you parse 800 pairs of texts :p
<0> although you can concatenate it to a single file but MXPOST is slightly faster then still
<0> not sure how much though
<0> what do you plan on using them for, JohnnyL and trane?
<1> I am making a chatbot ai, i need parsers to chunk input and carry out transformations and such. i'm using link now to do that but it doesn't work all the time...
<2> cwenner, i'd like to make a MOO over and beyond what zork engines there are out thre.
<0> will be interesting to see
<0> you may want to check interactivestory.net . It's a small game and somewhat limited but it contains real NLP
<2> cwenner, is there any more documentation for Stanford than the javadoc?
<0> there ought to be numerous articles describing the theory behind it but i'm not sure about the actual implementation
<0> do note that if you use these ready modules, you need to give the proper credits, you may not make it commercial and people taht download needs to accept a number of licenses. although it is probably better to use this for njow and implement or finding something almost as good later
<2> cwenner, i am positioning it to be GPL. is that ok?
<0> okay, no licenses for using the stanford parser, "Note that this is the full GPL, which allows its use for research purposes or other free software projects"
<0> no problem as long as you keep it free
<2> oh ok
<0> although you will probably want to add other NLP modules
<0> may need to check the license for those
<0> although in this development phase it's not that important
<2> however, if i through it on a server, and sell the server to a company. i can still charge $7000 for the server.
<2> still they'll be able to DL it (if it gets put on the net).
<0> if they pay for the hardware rather tahn the software that uses these modules
<2> right
<0> i'm not a lawyer though
<2> commerical MOOS/MMORPGs are so static.. when it's the MUD/MOOS that hold special interest. I want to make something everyone can use.
<0> i still recommend dependency graphs for language understanding though
<2> i have no idea what a dependecy graph is, but i'll look it up.
<0> make it in small steps then or you're unlikely to finish it or even get a base line
<2> a directed graph of dependencies?
<2> ok
<0> Linguistics.depgraph_sentence('the angry man threw the can').to_s
<0> "<throw/VBD SUB:<man/NN NMOD:<the/DT >, NMOD:<angry/JJ >>, VMOD:<can/MD SUB:<the/DT >>><the/DT ><angry/JJ ><man/NN NMOD:<the/DT >, NMOD:<angry/JJ >><throw/VBD SUB:<man/NN NMOD:<the/DT >, NMOD:<angry/JJ >>, VMOD:<can/MD SUB:<the/DT >>><the/DT ><can/MD SUB:<the/DT >>
<0> a bit difficult to see it like that perhaps, you need to unwrap it
<3> cwenner: what library is that?
<3> what language?
<0> 'the red cat ran away' ~> <run/VBD SUB:<cat/NN NMOD:<the/DT >, NMOD:<red/JJ >>, VMOD:<away/RP >>
<0> here you have 'run' as the root of the graph
<0> the run node is connected to the cat node with a 'sub' relationship (i.e. the subject of the event)
<0> the cat node is connected to the 'the' node and the 'red' node
<0> the dterminer and the adjective here act as modifiers of the cat noun, which is the subject of the run event
<0> the hierarchical structure makes the interpretation somewhat more natural
<0> psaronius: merely my refactoring of my NLP project, ruby
<0> there is a Linguistics module for Ruby but it was rather restrictive
<4> what Bayesian prior do you use to derive the Laplace estimator (add one) for a probability?
<4> I read somewhere that you use a uniform prior, but I don't see how a uniform prior is going to change the MLE estimate
<0> you can derive the laplace estimation by ***uming a uniform distribution over all possible ***ignments of probabilities to all tokens
<0> although, i don
<0> 't
<0> i just thought it was an interesting relationship
<4> I was always under the impression that if you use a uniform prior then nothing changes
<0> ML and a bayesian prior approach are equivalent
<0> you can derive one fromt he other
<4> not really, in MLE you don't use a prior
<0> MLE as in?
<4> maximum likelihood estimation
<4> (using t for theta), t_MLE = argmax_t P(X|t), but t_MAP = argmax_t P(t|X) = argmax_t P(X|t)P(t)
<0> in ML, you can ***ume a prior and account it in your calculations, i.e. you can motivate the bayesian approach from ML
<0> you can also derive ML from bayesian with a probability that totals on the most likely
<4> if you use a prior it's called MAP
<0> it doesn
<0> t matter what they are called
<4> it kind of does, since you're maximizing two different things.
<0> what i'm saying is that both approaches are equivalent, wioth the right ***umptions you can entail identical results
<4> what do you mean by "equivalent"?
<0> if you have a model M and what to solve/estimate some variables, you may either make the Bayesian or the ML ***umption, and for each, there is a set of non-contradicting ***umptions/parameters to the B/ML ***umption that allows you to deduce the same estimates from the model
<4> what is the "ML ***umption"?
<0> ah sorry, for each non-contradictous ***umption/parameters for either the B or ML ***umption there is one for the other
<0> most likelihood
<4> but one of them isn't using a prior...
<0> right but you may construct a different probability space which takes the model and a prior and deduce the bayesian appraoch from the ML appraoch
<4> there's no question of "deducing" anything
<4> you can choose what to maximize
<4> there are two commonly used things, and they're related but different
<0> and that is the B or ML ***umption
<4> it's not an ***umption, it's just a choice of optimization objective
<0> and that is the ***umption i speak of
<4> I think the language you're using is highly confusing. in any case, my real question is how a uniform prior leads to the add-one penalty
<0> ***ume you have some disjoint events p_k, sum p_k = 1
<4> by events you mean probabilities? ok.
<2> ML, B what on earth are you two talking about?
<0> that was implicit
<0> by events i mean events but of course p_k refers to teh probabilities
<4> I think you're being too casual about the use of technical words, no offense. it makes things confusing.
<4> JohnnyL: maximum likelihood, bayesian. we're talking about parameter estimation.
<0> i'll take it into account
<4> thanks.
<4> but ok, probabilities p_k.
<0> is it called 'tupel' in english?
<0> <1,2,3>
<4> tuple
<4> (and standard notation here is (1,2,3))
<0> let T be the uncountable(?) set of n-tuples {<q_1,...q_n>} for which sum q_k = 1. if we in a different probability space ***ume that these are uniformly distributed the expected probability for some event k given a number of observations {o_i} drawn from the probability space given by T is just (o_k + 1) / sum(o_i + 1)
<0> unless i did i mistake calculating it
<4> where'd you pull that formula from though?
<0> i was just pondering
<4> and the usual definition of the estimator is theta-hat = argmax_theta P(X|theta)P(theta)
<4> so if you're going to get this expression it should be derived through that optimization
<0> i didn't do an ML ***umption
<4> oh
<4> then I'm not sure what you're talking about.
<0> you asked about the derivation of the Laplace rule when ***uming an uniform prior
<4> yes, but via an MAP estimation
<4> the Laplace rule comes from MAP, not from whatever it is you did
<4> it's supposed to be the solution to an optimization problem
<0> i wasn't referring to the old meaning of Bayesian, uniform probability of outcomes
<0> by Bayesian, I above meant MAP
<4> fine, but where did you do MAP above?
<4> you just said ***ume a uniform distribution over all these tuples and magically pull out the Laplace rule without doing anything
<0> uniform distribution over the tuples and {o_i} observations from the probability space complemented by the probabilities in the tuple
<0> i.e. E[p_k | {o_i}]
<4> but why are you computing an expectation?
<0> sorry about the implicit notation there, O_i = o_i is perhaps better?
<4> that much is ok, I just don't understand how what you're saying has anything to do with the standard definition of an MAP estimator.
<0> cause the expected probability given those observations leads to the Laplace rule
<0> huh
<0> this has nothing to do with the MAP vs. ML ***umption
<4> ok, but every time I've read about the Laplace rule, it's been justified as being an MAP estimator.
<4> yes, it does
<0> how?
<4> that's my question! :)
<0> i'm not saying it has anything to do with it
<4> but it definitely does, it comes from choosing a certain prior distribution and solving theta_MAP = argmax_theta P(X|theta)P(theta) for specific choices of P(X|theta) and P(theta)
<4> yes, but I *am* saying it has something (everything) to do with it
<0> you ***ume an uniform prior distribution over all legal ***ignments of probabilities to the events and you end up with the Laplace rule for estimating the probability of an event
<4> the Laplace rule is just a regularized ML estimate
<4> yes, but when you're sayign "end up," you're glossing over the fact that you think it's an expectation and I think it's the MAP optimization
<4> I agree that the prior is supposed to be uniform
<0> it's an ***umption
<4> what's an ***umption?
<0> i thought it was an interesting relation that you end up with the Laplace rule solving the expected value
<0> that you ***ume an uniform prior over the legal ***ignments of probabilities
<4> maybe it is, ***uming that's true (which I still don't see), but it has nothing to do with the way everyone seems to justify the Laplace rule, which is via MAP.
<0> perhaps not, what is that justification?
<4> unfortunately I have to run, but I'll try to continue this later.
<4> the justification is that it's an MAP estimate when you choose a uniform prior
<4> that's all
<0> okay, i'll try to formalize it, if you wait til saturday i can have it prepared
<4> there's nothing complicated to it, I just don't see how the math works
<4> where a uniform prior gives the + 1
<4> sure, I can wait
<4> thanks a lot. have to run now. I'll let you know if I figure out something.
<0> ***uming that what i say is true, and that these ***umptions (and some other) are equivalent, there are plenty of ways to deduce it
<0> okay, sorry, i'll have it prepared. good luck
<3> Are there any commercial AI projects that use biologically-modelled synaptic plasticity for their products? I am trying to understand synaptic plasticity at a basic yet deep level and this would help thanks
<3> "unsupervised learning algorithms"
<3> I was thinking.. man, that would be great for a gigapet
<3> once the computing power becomes availible of course
<3> have you guys seen that episode of MacGuyver where the machine takes over the entire lab and macguyver looks at the screen and all this matrix algebra pops up? I understand that that is significant
<3> MacGuyver uses his expert skills to short out the computer
<3> Macgyver
<3> MacGyver
<5> gya much
Return to
#ai
or
Go to some related
logs:
#centos
#debian
debian start-stop-daemon is fake
#css
#math
ubuntu change workgroup
rpc.mountd: getfh failed: Function not implemented kernel
openmotif for puppy
#debian
cross-compile+icecc
|
|