Primerjalna analiza napak: Trankit in CLASSLA-Stanza
Testna množica SSJ-UD (1.282 povedi, 25.442 pojavnic) · Stolpci 7–8 (head + deprel) · Metriki LAS in UAS
Verzije modelov:
Trankit CLARIN 11356/1997
(trankit-sl-ssj+sst.zip, MD5 0ddfac8d7445f8fa300f59dde1a00352) s
trankit==1.1.2 ·
CLASSLA classla==2.2.1
(classla.Pipeline('sl', pos_use_lexicon=True)).
Splošna natančnost
Trankit LAS
94,59
+4,11 pp vs. CLASSLA
CLASSLA LAS
90,48
labeled attachment score
Trankit UAS
95,79
+3,67 pp vs. CLASSLA
CLASSLA UAS
92,12
unlabeled attachment score
Tabela 1 · Distribucija napak glede na tip (po vzoru Tabele 3 v Dobrovoljc idr. 2023)
Privzeto je vsak blok napak razvrščen po stolpcu Trankit, od največ do najmanj napak. Klik na glavo stolpca spremeni razvrščanje.
Tip napake
Trankit
CLASSLA
Razlika
Napačno jedro
666
1.202
-536
punct-punct
198
355
-157
advmod-advmod
97
161
-64
nmod-nmod
76
116
-40
conj-conj
80
108
-28
obl-obl
34
76
-42
acl-acl
21
52
-31
cc-cc
15
39
-24
aux-aux
12
37
-25
parataxis-parataxis
27
37
-10
mark-mark
10
34
-24
nsubj-nsubj
13
30
-17
cop-cop
11
29
-18
advcl-advcl
21
27
-6
appos-appos
15
19
-4
csubj-csubj
1
17
-16
case-case
9
17
-8
obj-obj
4
11
-7
amod-amod
4
10
-6
expl-expl
2
6
-4
nummod-nummod
5
4
+1
orphan-orphan
1
4
-3
iobj-iobj
1
4
-3
ccomp-ccomp
4
1
+3
xcomp-xcomp
1
2
-1
det-det
1
2
-1
vocative-vocative
1
0
+1
flat-flat
0
1
-1
discourse-discourse
0
1
-1
list-list
1
1
0
dep-dep
1
1
0
Napačno jedro in oznaka
404
803
-399
obl→nmod
44
89
-45
nmod→obl
39
62
-23
root→cop
7
28
-21
nsubj→root
10
24
-14
parataxis→root
21
23
-2
root→parataxis
12
18
-6
cop→root
3
15
-12
root→nsubj
6
13
-7
advcl→acl
9
12
-3
obj→nmod
4
11
-7
nsubj→conj
0
10
-10
nmod→nsubj
2
10
-8
nsubj→parataxis
1
9
-8
conj→cop
1
9
-8
parataxis→appos
8
9
-1
cc→advmod
0
8
-8
nsubj→nmod
3
8
-5
parataxis→acl
3
7
-4
nsubj→acl
4
7
-3
acl→advcl
5
7
-2
root→ccomp
7
7
0
root→csubj
0
6
-6
obl→root
1
6
-5
nsubj→ccomp
1
6
-5
ccomp→cop
1
6
-5
obl→appos
2
6
-4
nmod→flat
6
2
+4
conj→nmod
2
6
-4
parataxis→nummod
3
6
-3
parataxis→nsubj
3
6
-3
nmod→appos
6
4
+2
nmod→parataxis
0
5
-5
cop→parataxis
0
5
-5
conj→parataxis
1
5
-4
advcl→root
1
5
-4
conj→nsubj
2
5
-3
advmod→orphan
2
5
-3
cop→acl
3
5
-2
conj→amod
5
4
+1
conj→obl
0
4
-4
case→obl
0
4
-4
nmod→orphan
1
4
-3
det→advmod
1
4
-3
appos→obj
1
4
-3
appos→flat
4
1
+3
appos→conj
1
4
-3
root→advmod
2
4
-2
orphan→parataxis
2
4
-2
orphan→advmod
2
4
-2
nmod→conj
2
4
-2
iobj→nmod
2
4
-2
case→fixed
4
2
+2
amod→conj
2
4
-2
cop→conj
3
4
-1
nmod→obj
4
4
0
flat→nmod
4
4
0
appos→nummod
4
4
0
appos→nmod
4
4
0
root→amod
0
3
-3
parataxis→csubj
0
3
-3
obj→nsubj
0
3
-3
csubj→root
0
3
-3
conj→root
0
3
-3
conj→csubj
0
3
-3
advmod→nsubj
0
3
-3
root→obj
1
3
-2
parataxis→amod
1
3
-2
conj→flat
3
1
+2
ccomp→csubj
1
3
-2
appos→parataxis
1
3
-2
amod→advmod
1
3
-2
advmod→root
1
3
-2
acl→parataxis
3
1
+2
acl→csubj
1
3
-2
root→obl
2
3
-1
parataxis→conj
2
3
-1
appos→obl
2
3
-1
amod→root
2
3
-1
advmod→det
2
3
-1
advcl→cop
2
3
-1
acl→nsubj
2
3
-1
acl→cop
2
3
-1
mark→fixed
3
3
0
conj→advmod
3
3
0
xcomp→cop
0
2
-2
root→conj
0
2
-2
parataxis→cop
0
2
-2
obl→nsubj
2
0
+2
obj→conj
0
2
-2
nummod→obj
2
0
+2
nummod→flat
2
0
+2
nsubj→det
0
2
-2
nmod→root
0
2
-2
det→nsubj
0
2
-2
dep→nummod
0
2
-2
csubj→conj
0
2
-2
cop→advcl
0
2
-2
conj→ccomp
0
2
-2
conj→advcl
0
2
-2
ccomp→root
0
2
-2
cc→flat
2
0
+2
amod→obl
2
0
+2
amod→ccomp
0
2
-2
amod→acl
0
2
-2
advmod→mark
2
0
+2
advmod→cc
2
0
+2
advmod→amod
0
2
-2
advcl→nsubj
0
2
-2
advcl→conj
0
2
-2
advcl→ccomp
0
2
-2
advcl→aux
0
2
-2
acl→xcomp
0
2
-2
root→nmod
2
1
+1
root→aux
1
2
-1
parataxis→orphan
1
2
-1
obl→orphan
1
2
-1
obl→det
1
2
-1
obl→conj
1
2
-1
obj→root
1
2
-1
obj→iobj
2
1
+1
obj→appos
1
2
-1
nsubj→obj
2
1
+1
cop→ccomp
1
2
-1
conj→acl
1
2
-1
ccomp→nsubj
1
2
-1
amod→nsubj
1
2
-1
advmod→appos
1
2
-1
root→appos
2
2
0
parataxis→advmod
2
2
0
orphan→conj
2
2
0
obj→advmod
2
2
0
nummod→conj
2
2
0
nsubj→appos
2
2
0
nsubj→amod
2
2
0
nmod→iobj
2
2
0
amod→parataxis
2
2
0
xcomp→obj
0
1
-1
xcomp→nmod
0
1
-1
root→vocative
0
1
-1
root→reparandum
1
0
+1
root→expl
1
0
+1
root→discourse
0
1
-1
root→case
0
1
-1
root→advcl
0
1
-1
parataxis→aux
0
1
-1
orphan→obl
1
0
+1
orphan→nmod
0
1
-1
orphan→fixed
1
0
+1
orphan→ccomp
0
1
-1
orphan→cc
1
0
+1
obl→flat
1
0
+1
obl→ccomp
0
1
-1
obl→case
1
0
+1
obj→parataxis
0
1
-1
obj→flat
1
0
+1
obj→det
0
1
-1
obj→amod
1
0
+1
obj→acl
0
1
-1
nummod→obl
0
1
-1
nsubj→orphan
0
1
-1
nsubj→obl
0
1
-1
nsubj→nummod
0
1
-1
nsubj→iobj
0
1
-1
nmod→xcomp
1
0
+1
nmod→list
1
0
+1
nmod→case
0
1
-1
nmod→advmod
0
1
-1
list→flat
1
0
+1
iobj→obj
0
1
-1
iobj→det
0
1
-1
flat→obj
0
1
-1
flat→fixed
0
1
-1
flat→conj
0
1
-1
flat→advmod
0
1
-1
fixed→parataxis
0
1
-1
fixed→mark
1
0
+1
fixed→cc
0
1
-1
discourse→root
0
1
-1
discourse→ccomp
0
1
-1
discourse→advmod
0
1
-1
det→obj
0
1
-1
dep→obl
0
1
-1
dep→discourse
1
0
+1
dep→cc
0
1
-1
dep→appos
0
1
-1
csubj→obl
0
1
-1
csubj→advcl
0
1
-1
cop→csubj
0
1
-1
cop→aux
0
1
-1
conj→mark
0
1
-1
conj→appos
0
1
-1
ccomp→parataxis
1
0
+1
ccomp→obl
0
1
-1
ccomp→nmod
0
1
-1
ccomp→conj
0
1
-1
ccomp→amod
0
1
-1
case→advmod
0
1
-1
aux→advcl
0
1
-1
appos→list
0
1
-1
appos→case
0
1
-1
appos→acl
1
0
+1
amod→xcomp
1
0
+1
amod→nmod
0
1
-1
amod→appos
0
1
-1
advmod→xcomp
0
1
-1
advmod→parataxis
0
1
-1
advmod→obl
0
1
-1
advmod→obj
0
1
-1
advmod→nmod
0
1
-1
advmod→iobj
0
1
-1
advmod→acl
1
0
+1
advcl→nmod
0
1
-1
advcl→csubj
0
1
-1
advcl→advmod
0
1
-1
acl→obl
1
0
+1
acl→obj
1
0
+1
acl→mark
0
1
-1
acl→conj
0
1
-1
acl→ccomp
0
1
-1
acl→advmod
0
1
-1
xcomp→conj
1
1
0
xcomp→advmod
1
1
0
orphan→xcomp
1
1
0
orphan→root
1
1
0
orphan→appos
1
1
0
obj→case
1
1
0
nummod→root
1
1
0
nummod→parataxis
1
1
0
nsubj→case
1
1
0
nmod→amod
1
1
0
mark→case
1
1
0
mark→advmod
1
1
0
flat→nummod
1
1
0
fixed→orphan
1
1
0
fixed→advmod
1
1
0
det→root
1
1
0
dep→conj
1
1
0
dep→case
1
1
0
csubj→ccomp
1
1
0
conj→obj
1
1
0
conj→nummod
1
1
0
appos→root
1
1
0
appos→iobj
1
1
0
appos→advmod
1
1
0
amod→advcl
1
1
0
advmod→conj
1
1
0
advcl→amod
1
1
0
acl→case
1
1
0
Napačna oznaka
306
416
-110
obj→nsubj
10
30
-20
obj→iobj
14
29
-15
nsubj→obj
12
28
-16
iobj→obj
19
10
+9
nmod→flat
17
12
+5
conj→parataxis
13
16
-3
obl→nsubj
3
11
-8
appos→conj
5
10
-5
flat→nmod
2
9
-7
cc→advmod
5
9
-4
conj→appos
2
8
-6
iobj→expl
5
8
-3
obl→obj
3
7
-4
nmod→obl
7
4
+3
nsubj→obl
0
6
-6
orphan→advmod
6
2
+4
advmod→amod
3
6
-3
appos→nmod
6
4
+2
advmod→nsubj
1
5
-4
advcl→advmod
5
1
+4
obl→nmod
3
5
-2
advcl→ccomp
5
3
+2
nmod→amod
4
5
-1
advmod→obl
5
4
+1
conj→flat
5
5
0
parataxis→obl
0
4
-4
cc→case
0
4
-4
nsubj→iobj
1
4
-3
csubj→advcl
1
4
-3
parataxis→conj
4
2
+2
nmod→conj
3
4
-1
advmod→cc
3
4
-1
orphan→nmod
4
4
0
obj→advmod
0
3
-3
obl→xcomp
1
3
-2
nmod→advmod
3
1
+2
ccomp→advcl
1
3
-2
appos→parataxis
3
1
+2
advcl→xcomp
1
3
-2
obl→advmod
2
3
-1
nmod→nummod
3
2
+1
mark→advmod
3
2
+1
expl→iobj
2
3
-1
amod→nmod
2
3
-1
amod→advmod
2
3
-1
advcl→csubj
2
3
-1
xcomp→obl
3
3
0
obj→obl
3
3
0
mark→cc
3
3
0
csubj→ccomp
3
3
0
cc→dep
3
3
0
advmod→mark
3
3
0
advcl→conj
3
3
0
xcomp→nsubj
0
2
-2
nummod→nmod
0
2
-2
nmod→appos
2
0
+2
expl→obj
2
0
+2
cop→aux
2
0
+2
conj→list
2
0
+2
ccomp→csubj
0
2
-2
cc→orphan
0
2
-2
amod→flat
2
0
+2
advmod→discourse
2
0
+2
advmod→aux
0
2
-2
parataxis→obj
2
1
+1
obj→expl
1
2
-1
nsubj→xcomp
1
2
-1
nsubj→advmod
2
1
+1
mark→obj
1
2
-1
discourse→advmod
2
1
+1
det→amod
2
1
+1
csubj→acl
1
2
-1
cc→mark
1
2
-1
appos→amod
1
2
-1
advcl→nsubj
1
2
-1
orphan→nummod
2
2
0
discourse→ccomp
2
2
0
ccomp→xcomp
2
2
0
advmod→xcomp
2
2
0
xcomp→obj
1
0
+1
xcomp→csubj
0
1
-1
xcomp→advcl
0
1
-1
vocative→ccomp
0
1
-1
parataxis→nsubj
0
1
-1
parataxis→cop
0
1
-1
parataxis→ccomp
1
0
+1
orphan→conj
1
0
+1
orphan→amod
0
1
-1
obl→orphan
1
0
+1
obl→det
0
1
-1
obl→conj
0
1
-1
obj→xcomp
1
0
+1
obj→conj
0
1
-1
obj→amod
0
1
-1
nummod→det
1
0
+1
nsubj→parataxis
0
1
-1
nsubj→nummod
0
1
-1
nsubj→conj
0
1
-1
nmod→orphan
0
1
-1
nmod→obj
1
0
+1
nmod→cc
0
1
-1
nmod→acl
1
0
+1
mark→nsubj
1
0
+1
list→flat
0
1
-1
list→conj
1
0
+1
flat→conj
1
0
+1
flat→amod
0
1
-1
fixed→orphan
1
0
+1
expl→nsubj
0
1
-1
expl→advmod
1
0
+1
discourse→obl
1
0
+1
discourse→mark
0
1
-1
det→orphan
1
0
+1
det→obl
1
0
+1
det→nsubj
0
1
-1
det→advmod
1
0
+1
dep→orphan
1
0
+1
dep→nmod
0
1
-1
dep→flat
1
0
+1
csubj→orphan
0
1
-1
csubj→obj
0
1
-1
conj→obj
1
0
+1
conj→nmod
0
1
-1
conj→ccomp
1
0
+1
ccomp→aux
0
1
-1
cc→discourse
0
1
-1
case→advmod
1
0
+1
aux→cop
0
1
-1
appos→xcomp
0
1
-1
appos→obl
1
0
+1
appos→nummod
0
1
-1
amod→parataxis
1
0
+1
amod→obl
0
1
-1
amod→nummod
0
1
-1
amod→nsubj
0
1
-1
amod→det
1
0
+1
amod→acl
1
0
+1
advmod→obj
0
1
-1
advmod→advcl
0
1
-1
advcl→obj
0
1
-1
acl→appos
0
1
-1
parataxis→appos
1
1
0
parataxis→advcl
1
1
0
orphan→det
1
1
0
nummod→flat
1
1
0
nsubj→nmod
1
1
0
nsubj→advcl
1
1
0
nmod→list
1
1
0
mark→case
1
1
0
discourse→orphan
1
1
0
csubj→xcomp
1
1
0
conj→obl
1
1
0
conj→advcl
1
1
0
aux→ccomp
1
1
0
appos→list
1
1
0
advmod→orphan
1
1
0
advmod→nmod
1
1
0
advmod→det
1
1
0
advcl→parataxis
1
1
0
advcl→obl
1
1
0
advcl→acl
1
1
0
acl→orphan
1
1
0
acl→csubj
1
1
0
acl→advcl
1
1
0
VSE NAPAKE
1.376
2.421
-1.045
Barvno ozadje:zelena = Trankit naredi manj napak (Trankit boljši) ·
modra = CLASSLA naredi manj napak (CLASSLA boljši) ·
brez ozadja = enako.
Pozitivna razlika v stolpcu Razlika pomeni, da Trankit naredi več napak; negativna, da manj.
Klikni relacijo v Tabeli 1 za prikaz primerov.
Prikažejo se vsi primeri za izbrani tip napake in relacijski vzorec.
Trankit
Ni izbrane relacije.
CLASSLA
Ni izbrane relacije.
Tabela 2 · Natančnost po relacijah (LAS, v %)
Privzeto je tabela razvrščena po stolpcu Trankit, od najvišje do najnižje vrednosti LAS. Klik na glavo stolpca spremeni razvrščanje.
Relacija
Opis
Testna
Trankit
CLASSLA
Razlika
Vizualizacija
csubj
clausal subject
78
90,9
50,3
+40,6
discourse
discourse element
15
66,7
50,0
+16,7
obj
object
1.085
95,2
91,3
+4,0
nsubj
nominal subject
1.312
96,0
89,0
+7,0
cop
copula
542
95,7
86,6
+9,1
acl
clausal modifier of noun
437
90,1
80,7
+9,4
obl
oblique nominal
1.722
94,0
88,3
+5,8
aux
auxiliary
1.162
98,8
96,3
+2,5
conj
conjunct
1.108
89,0
83,1
+5,9
amod
adjectival modifier
2.165
98,8
98,1
+0,8
case
case marking
2.415
99,4
98,9
+0,5
nmod
nominal modifier
1.887
90,9
86,1
+4,8
mark
marker
804
97,4
94,5
+3,0
root
root
1.282
96,6
92,5
+4,1
iobj
indirect object
145
82,5
76,0
+6,5
orphan
orphan
83
73,2
66,3
+6,9
list
list
16
74,3
82,3
−8,1
fixed
fixed multiword expression
95
93,8
94,8
−1,0
flat
flat multiword expression
156
84,6
86,1
−1,5
advmod
adverbial modifier
1.929
92,8
88,5
+4,3
parataxis
parataxis
325
76,7
64,9
+11,8
det
determiner
617
98,7
97,9
+0,8
VSE RELACIJE
25.442
94,59
90,48
+4,11
Barvno ozadje:zelena = Trankit boljši ·
modra = CLASSLA boljši ·
Miniaturne vrstice: ■ Trankit ■ CLASSLA. Vrednosti so preproste točnosti (gold count v imenovalcu), ne F1.
Povzetek: kje je kateri model boljši / slabši
Dimenzija
Trankit boljši
CLASSLA boljši
Jedrne arg. strukture
obj (+4,0), nsubj (+7,0), cop (+9,1), iobj (+6,6) — Trankit je boljši pri vseh glavnih argumentnih relacijah
—
Stavčni odvisniki
csubj (+40,6), parataxis (+11,8), advcl (+9,9), acl (+9,4) — izrazita prednost Trankita pri vseh štirih relacijah
—
Prislovna določila
advmod (+4,3) — Trankit je tudi tu stabilno boljši na ravni LAS
—
Večbesedne enote
—
fixed (−1,0), flat (−1,5) — CLASSLA je nekoliko boljši pri obeh relacijah
Elipsa / redke str.
orphan (+6,9) — Trankit je boljši pri relaciji orphan
list (−8,1) — CLASSLA ostaja boljši pri relaciji list
Oznake napak
Trankit: manj napak tipa "napačno jedro in oznaka" (803 → 404) in "napačna oznaka" (416 → 306), zlasti pri obl/nmod zamenjavah
Pri iobj→obj ima Trankit več specifičnih zamenjav (19 vs. 10), zato je tu boljša CLASSLA
Analiza temelji na pojavnicah, za katere se pokrivata gold in predicted segmentacija (25.442 pojavnic, 1.282 povedi). Metrika je token-level LAS (labeled attachment score), izračunana z enakim postopkom kot v Tabeli 2 članka Dobrovoljc idr. (2023). Razlike pri relacijah z majhnim številom pojavnic (discourse, csubj, list, orphan) so statistično manj zanesljive.