[02:39:34] *** Joins: drbean (~drbean@124.219.83.228)
[04:14:24] *** Quits: jmvanel (~jmvanel@79.88.0.66) (Ping timeout: 245 seconds)
[08:37:56] *** Joins: jmvanel (~jmvanel@66.0.88.79.rev.sfr.net)
[09:02:06] *** Joins: crazydiamond (~crazydiam@178.141.72.158)
[09:51:16] *** Quits: vin-ivar (~vinit@122.170.53.73) (Quit: WeeChat 1.0.1)
[09:56:26] *** Joins: vin-ivar (~vinit@122.170.53.73)
[10:10:05] <vin-ivar> phew, fixed everything. had to go thermonuclear on ghc
[10:33:58] *** Quits: crazydiamond (~crazydiam@178.141.72.158) (Remote host closed the connection)
[12:08:02] *** Quits: spectre (~fran@115.114.202.84.customer.cdi.no) (Ping timeout: 252 seconds)
[17:21:49] *** Joins: spectre (~fran@115.114.202.84.customer.cdi.no)
[17:43:35] *** Joins: vinit-ivar (~vinit@122.169.12.162)
[17:45:56] *** Quits: vin-ivar (~vinit@122.170.53.73) (Ping timeout: 264 seconds)
[20:08:55] *** vinit-ivar is now known as vin-ivar
[20:09:04] <vin-ivar> folks
[20:09:08] <vin-ivar> ~/Dev/FOSS/GF-master/extmini/gf-apertium master $ echo "I love the sleep" | ./MorphAnalyser.hs ../Lang.pgf LangEng | ./MorphDisambiguator.hs ../Lang.pgf LangEng
[20:09:10] <vin-ivar> MorphAnalyser.hs: <stdin>: hGetLine: end of file
[20:09:12] <vin-ivar> ^I/i_NP<s><Nom>$ ^love/love_V2<s><VInf>$ ^the/the_Det<s>$ ^sleep/sleep_N<s><Sg>$
[20:09:14] <vin-ivar> MorphDisambiguator.hs: <stdin>: hGetLine: end of file
[20:09:26] <vin-ivar> \o/
[20:10:19] <vin-ivar> i split the thing into two scripts and tidied it up a bit, the disambiguator converts the apertium stream format to the original string
[20:34:47] <vin-ivar> spectre: what d'you reckon I should work on next? lexical selection?
[20:35:26] <spectre> can you try it with another language
[20:35:28] <spectre> to see what happens
[20:35:37] <vin-ivar> sure
[20:35:44] <spectre> e.g. do we get english (interlingua) in the output
[20:35:44] <vin-ivar> gimme a bit
[20:35:47] <spectre> or original lang
[20:46:44] <vin-ivar> you get the analyses in english
[20:47:10] <spectre> ok
[20:47:27] <vin-ivar> that's how the lexicon seems to be defined anyway
[20:47:35] <spectre> echo "I love the sleep" | ./MorphAnalyser.hs ../Lang.pgf LangEng
[20:47:39] <spectre> what do you get out of this ?
[20:47:41] <vin-ivar> like, sleep_N = mkN "sonno" ;
[20:47:42] <spectre> do you get multiple lemmas ?
[20:47:48] <vin-ivar>  yeah
[20:47:50] <vin-ivar> for sleep
[20:47:54] <vin-ivar> I added sleep as a noun
[20:48:19] <spectre> the lexical selection would be interesting
[20:48:22] <spectre> you can have e.g.
[20:48:28] <spectre> sleep_N = mkN "sonno" | mkN "hargle" ;
[20:48:41] <vin-ivar> ^I/i_NP<s><Nom>$ ^love/love_V2<s><VInf>$ ^the/the_Det<s>$ ^sleep/sleep_V<s><VInf>/sleep_N<s><Sg>$
[20:48:41] <spectre> and see if you can get it to output >1 translation
[20:48:48] <vin-ivar> yeah
[20:48:50] <vin-ivar> it sounds interesting
[20:49:00] <vin-ivar> isn't lexical selection handled statistically?
[20:49:28] <spectre> so
[20:49:30] <spectre> you can have e.g.
[20:49:39] <spectre> sleep_N = mkN "sonno" | mkN "hargle" ; (ambiguous in SL generation)
[20:49:40] <spectre> or
[20:49:48] <spectre> sleep1_N = mkN "sonno" ;
[20:49:52] <spectre> sleep2_N = mkN "sonno" ;
[20:49:52] <spectre>  
[20:49:59] <spectre> (ambiguous in SL analysis)
[20:50:03] <spectre> i don't really know how it works
[20:50:13] <spectre> lexical selection could be to do with both
[20:50:18] <vin-ivar> i think i had used the latter at one point
[20:50:27] <spectre> although the ideal is obviously that the interlingua contains all the concepts
[20:50:34] <spectre> to be able to distinguish the sentences
[20:50:47] <spectre> and only pure synonyms (lol kilgarrif there are no synonyms) are with the | way
[20:51:31] <spectre> brb
[20:52:01] <vin-ivar> hm, how would the interlingua store concepts related to lexical selection, though
[20:52:17] <spectre> it wouldn't
[20:52:23] <spectre> lexical selection would then be part of parsing
[20:52:26] <spectre> i guess
[20:52:49] <spectre> you would have to e.g. alter the parse tree probs (or have some other disambiguation module)
[20:52:55] <spectre> to give
[20:53:01] <vin-ivar> right, so there'd have to be some sort of information about what word sense to use. how do you do that without probabilities or rules?
[20:53:01] <spectre> a score for
[20:53:10] <spectre> p(estación|station_N)
[20:53:13] <spectre> p(estación|season_N)
[20:53:15] <vin-ivar> yeah, that's what i was thinking
[20:53:24] <spectre> because if you have
[20:53:37] <spectre> station_N = mkN "estació" ;
[20:53:44] <spectre> season_N = mkN "estació" ;
[20:53:47] <spectre> i'm not sure how it works at the moment
[20:54:00] <spectre> probably it is just conditioned on the probabilities of station_N and season_N in the penntreebank
[20:54:02] <vin-ivar> it could be contextual, I suppose
[20:54:10] <spectre> yes, that's the idea of lexical selection
[20:54:11] <vin-ivar> more helpful than just unigram probabilities
[20:54:15] <vin-ivar> right
[20:54:22] <spectre> that you use contexts to modify the parse tree probabilities
[20:54:23] <spectre> or something
[20:54:30] <vin-ivar> aye
[20:54:37] <spectre> e.g. if you have estació seca
[20:54:52] <spectre> then you bump estación->season_N
[20:54:58] <spectre> but if you have estació de tren
[20:55:06] <spectre> then you bump estació->station_N
[20:55:08] <vin-ivar> but then you also need corpora to get probabilities
[20:55:16] <vin-ivar> though I suppose I could just ghetto my own for starters
[20:55:19] <spectre> vin-ivar, you can do it either supervised (if you have parallel corpora)
[20:55:23] <spectre> or unsupervised (if you don't)
[20:55:34] <spectre> msg me your email addr
[20:55:39] <vin-ivar> okay
[20:56:43] <spectre> will send you a paper
[20:57:03] <spectre> sent
[20:57:28] <vin-ivar> got it
[20:57:42] <vin-ivar> looks nice
[20:58:03] <spectre> bbiab :)
[20:58:11] <vin-ivar> cool
[21:20:20] <vin-ivar> spectre: i'll take a look at the the paper tomorrow, about to crash now
[21:20:29] <vin-ivar> thanks, it looks really cool :D
[22:03:20] *** Quits: vin-ivar (~vinit@122.169.12.162) (Ping timeout: 264 seconds)
[23:47:59] *** Joins: crazydiamond (~crazydiam@178.141.71.146)