search

own-pt / rte-sick

RTE Experiment
1 stars 3 forks source link

verbs and WSD #25

Open vcvpaiva opened 7 years ago

vcvpaiva commented 7 years ago

from #7
Update from https://github.com/own-pt/rte-sick/blob/master/expanded/lemmas/lemmas-VERB.txt have 562 513 "verb lemmas". 400 389 with 2 or more occurrences. 150 118 with 10 or more occurrences. 80 42 with 30 or more occurrences. 20 18 with more than 83 occurrences

if we could automatically discover the verbs with say less than 6 senses in PWN, it might be worth checking these, as WSD might work better for these.

verbs like "make, play, run, hold" with more than 30 senses are really difficult to get the right sense. (some verbs like "do, put, have" have fewer PWN senses, but are difficult too). and not clear how to deal with WSD not working.

if PWN->SUMO mappings don't work, we can correct them, but if WSD doesn't work, what to do?

top of the list:

895 play 35 senses 423 wear 9 409 stand 12 408 run 41 346 ride 14 345 sit 11 330 jump 16 287 walk 11 217 look 10 206 slice 4--2 189 hold 36 175 cut 42 141 dance 5 128 eat 6--4 110 put 9 106 do 13 97 make 52 84 have 20 84 climb 6 83 sing 6

vcvpaiva commented 7 years ago

@arademaker produced for the first verb above (900 occurrences) for the previous dependencies, using syntaxnet the following figures.

what is the new distribution of "play", please? is it the same?

763 play DramaticActing+ 78 play Game+ 6 play IntentionalProcess+ 2 play Contest+ 1 play Pretending+ 1 play MakingInstrumentalMusic+ 1 play InstrumentalMusic+

and a similar question for the 423 occurrences of "wear", they all seem to me "wearing clothes, or hats, or collars, etc". "wearing out" is hard to see in a picture, PWN has perhaps too many kinds of wear?

arademaker commented 7 years ago

One possible ideia is to investigate the performance of WSD with different algorithms, maybe using https://github.com/alvations/pywsd that provides some implementations packaged in a single lib.

vcvpaiva commented 7 years ago

new top of the list, only normalized sentences: 7176 be 688 play 336 stand 325 wear 304 run 270 sit 262 ride 253 jump 217 walk 169 look 156 slice 145 hold 132 cut 108 dance 98 eat 90 put 83 make 83 do 68 have 65 lie 62 sing 62 climb 57 talk 57 perform 56 take 55 watch 51 carry 45 pour 43 drive 43 dress 43 chase 41 kick 40 chop 39 fight 37 get 37 catch 35 cover 34 pose 34 drink 31 fall 30 swim 30 cook