Open jbrry opened 4 years ago
Another example in UD_Lithuanian-ALKSNIS/lt_alksnis-ud-train.conllu
.
Token 8 has two edges to token 5 with different deprels: 5:acl:relcl|5:ref
:
# sent_id = kalbeti_ar_tyleti-s18
# text = Pasistenkite paaiškinti, kodėl įvyko neištikimybė: kokie tuo metu buvo jūsų tarpusavio santykiai, ko ieškojote.
# orig_file_sentence kalbeti_ar_tyleti#18
1 Pasistenkite pasistengti VERB vksm.asm.sngr.liep.dgs.2. Mood=Imp|Number=Plur|Person=2|Polarity=Pos|Reflex=Yes|VerbForm=Fin 0 root 0:root _
2 paaiškinti paaiškinti VERB vksm.bndr. Polarity=Pos|VerbForm=Inf 1 xcomp 1:xcomp SpaceAfter=No
3 , , PUNCT skyr. _ 5 punct 5:punct _
4 kodėl kodėl ADV prv.nelygin. Degree=Pos|PronType=Int,Rel 5 advmod 5:advmod _
5 įvyko įvykti VERB vksm.asm.tiesiog.būt-k.vns.3. Aspect=Perf|Mood=Ind|Number=Sing|Person=3|Polarity=Pos|Tense=Past|VerbForm=Fin 2 acl 2:acl|17:obl:arg:gen _
6 neištikimybė neištikimybė NOUN dkt.mot.vns.V. Case=Nom|Gender=Fem|Number=Sing 5 nsubj 5:nsubj SpaceAfter=No
7 : : PUNCT skyr. _ 5 punct 5:punct _
8 kokie koks DET įv.vyr.dgs.V. Case=Nom|Definite=Ind|Gender=Masc|Number=Plur|PronType=Int,Rel 5 acl:relcl 5:acl:relcl|5:ref _
9 tuo tas DET įv.vyr.vns.Įn. Case=Ins|Definite=Ind|Gender=Masc|Number=Sing|PronType=Dem 10 det 10:det _
10 metu metas NOUN dkt.vyr.vns.Įn. Case=Ins|Gender=Masc|Number=Sing 8 obl 8:obl:ins _
11 buvo būti AUX vksm.asm.tiesiog.būt-k.dgs.3. Aspect=Perf|Mood=Ind|Number=Plur|Person=3|Polarity=Pos|Tense=Past|VerbForm=Fin 8 cop 8:cop _
12 jūsų tu PRON įv.dgs.K. Case=Gen|Definite=Ind|Number=Plur|Person=2|PronType=Prs 14 nmod 14:nmod:gen _
13 tarpusavio tarpusavis NOUN dkt.vyr.vns.K. Case=Gen|Gender=Masc|Number=Sing 14 nmod 14:nmod:gen _
14 santykiai santykis NOUN dkt.vyr.dgs.V. Case=Nom|Gender=Masc|Number=Plur 8 nsubj 5:nsubj|8:nsubj SpaceAfter=No
15 , , PUNCT skyr. _ 17 punct 17:punct _
16 ko kas PRON įv.K. Case=Gen|Definite=Ind|PronType=Int,Rel 17 obl:arg 5:ref _
17 ieškojote ieškoti VERB vksm.asm.tiesiog.būt-k.dgs.2. Aspect=Perf|Mood=Ind|Number=Plur|Person=2|Polarity=Pos|Tense=Past|VerbForm=Fin 8 conj 5:acl:relcl|8:conj SpaceAfter=No
18 . . PUNCT skyr. _ 1 punct 1:punct _
Some tokens, e.g. token 39 from
UD_Dutch-LassySmall/nl_lassysmall-ud-train.conllu
have multiple heads but two of the heads are the same but have a different dependency label:36:obj|46:nsubj:relsubj|46:nsubj:xsubj
.At the moment the parser considers edges between all pairs of words above a threshold. When a score is above the threshold it takes the edge and the label for that edge with the highest score. In such a case it will choose the deprel with the highest probability and not and not consider the same head twice and choosing a different deprel each time.
I'm not sure how best to deal with this at the moment but it warrants further investigation so am noting it here for now.