| Subject: | possible bug in depth finding of wup |
The following was reported via Hideki Shima of CMU via email. I wanted to get this recorded here as well.
-----------------------------------------------------
(1) WUP: Deepest LCS is not found
-----------------------------------------------------
When coffee#n#1 tea#n#1 is given, the following trace is obtained:
<trace>
HyperTree: *Root*#n#1 entity#n#1 physical_entity#n#1 matter#n#3
substance#n#7 food#n#1 beverage#n#1 coffee#n#1
HyperTree: *Root*#n#1 entity#n#1 physical_entity#n#1 matter#n#3
substance#n#1 fluid#n#1 liquid#n#1 beverage#n#1 coffee#n#1
HyperTree: *Root*#n#1 entity#n#1 abstraction#n#6 relation#n#1 part#n#1
substance#n#1 fluid#n#1 liquid#n#1 beverage#n#1 coffee#n#1
HyperTree: *Root*#n#1 entity#n#1 physical_entity#n#1 matter#n#3
substance#n#7 food#n#1 beverage#n#1 tea#n#1
HyperTree: *Root*#n#1 entity#n#1 physical_entity#n#1 matter#n#3
substance#n#1 fluid#n#1 liquid#n#1 beverage#n#1 tea#n#1
HyperTree: *Root*#n#1 entity#n#1 abstraction#n#6 relation#n#1 part#n#1
substance#n#1 fluid#n#1 liquid#n#1 beverage#n#1 tea#n#1
Lowest Common Subsumers: beverage#n#1 (Depth=7)
Depth(coffee#n#1) = 8
Depth(tea#n#1) = 8
coffee#n#1 tea#n#1 0.875
</trace>
The deepest subsumer seems to be beverage#n#1 (Depth=9) from
the 3rd and 6th hyper-tree. If so, the score would be
2 * 9 / (10 + 10) = 0.9
Below are some more of such cases.
wup( goat#n#3 , scapegoat#n#1 ) = 2 * 8 / (9 + 11) = 0.8000?
wup( boy#n#1 , sage#n#1 ) = 2 * 8 / (10 + 11) = 0.7619?
wup( successor#n#1 , precursor#n#2 ) = 2 * 8 / (10 + 9) = 0.842?
wup( h2o#n#1, co2#n#1 ) = 2 * 9 / (11 + 12) = 0.7826?
wup( tobacco#n#1, alcohol#n#1 ) = 2 * 8 / (9 + 9) = 0.8889?
This phenomenon has been observed in about 200 out of
10k (=2%) randomly generated noun-noun pairs of synsets.
Currently, wup score is estimated to be lower than the expected
score by 0.09 on average over these 2% of pairs.