search

udapi / udapi-python

Python framework for processing Universal Dependencies data
GNU General Public License v3.0
57 stars 31 forks source link

Corefud.MergeSameSpan does not work after certain blocks #101

Closed dan-zeman closed 2 years ago

dan-zeman commented 2 years ago

The block corefud.MergeSameSpan should be relatively independent of what the other blocks do. It collects mentions in a sentence (asking for node.coref_mentions at every node), then extracts the list of words for each mention m (set(m.words)), and looks for pairs of mentions that span the same set of words. Nevertheless, there is a mysterious bug that triggers when this block is combined with the conversion from the old CorefUD format. Consider this sentence and especially line 26, nuestro:

# sent_id = CESS-CAST-P-20020501-120-s5
# text = Y Charly, cómo no, se mojó: "Estamos, más o menos, un 75% a favor de ellos y un 25% nuestro" .
1   Y   y   CCONJ   cc  _   8   advmod  8:advmod    _
2   Charly  Charly  PROPN   np00000 _   8   nsubj   8:nsubj ClusterId=CESS-CAST-P-20020501-120-c1|ClusterType=Spec.person|MentionSpan=2|MentionMisc=CorefType:ident|SpaceAfter=No
3   ,   ,   PUNCT   fc  PunctType=Comm  4   punct   4:punct _
4   cómo    cómo    ADV _   _   8   dep 8:dep   _
5   no  no  PART    _   _   4   dep 4:dep   SpaceAfter=No
6   ,   ,   PUNCT   fc  PunctType=Comm  4   punct   4:punct _
7   se  él  PRON    p0300000    Case=Acc|Person=3|PrepCase=Npr|PronType=Prs|Reflex=Yes  8   expl:pv 8:expl:pv   _
8   mojó    mojar   VERB    vmis3s0 Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin   0   root    0:root  SpaceAfter=No
9   :   :   PUNCT   fd  PunctType=Colo  18  punct   18:punct    _
10  "   "   PUNCT   _   PunctType=Quot  18  punct   18:punct    SpaceAfter=No
10.1    _   _   PRON    p   _   _   _   11:nsubj    ClusterId=CESS-CAST-P-20020501-120-c3|ClusterType=Spec.organization|MentionSpan=10.1|MentionMisc=ClusterTypeMismatch:Spec:Spec.organization,CorefType:ident
11  Estamos estar   AUX vmip1p0 Mood=Ind|Number=Plur|Person=1|Tense=Pres|VerbForm=Fin   18  cop 18:cop  SpaceAfter=No
12  ,   ,   PUNCT   fc  PunctType=Comm  13  punct   13:punct    _
13  más más ADV rg  Degree=Cmp  18  advmod  18:advmod   _
14  o   o   CCONJ   cc  _   15  cc  15:cc   _
15  menos   menos   ADV rg  Degree=Cmp  13  conj    13:conj SpaceAfter=No
16  ,   ,   PUNCT   fc  PunctType=Comm  13  punct   13:punct    _
17  un  uno DET di0ms0  Definite=Ind|Gender=Masc|Number=Sing|PronType=Art   18  det 18:det  _
18  75% 75  SYM _   NumForm=Digit   8   advcl   8:advcl ClusterId=CESS-CAST-P-20020501-120-s5.sn.34|ClusterType=Spec.number|MentionSpan=17-22
19  a   a   ADP sps00   _   22  case    22:case _
20  favor   favor   NOUN    _   _   19  fixed   19:fixed    _
21  de  de  ADP _   _   19  fixed   19:fixed    _
22  ellos   él  PRON    pp3mp000    Case=Acc,Nom|Gender=Masc|Number=Plur|Person=3|PronType=Prs  18  nmod    18:nmod ClusterId=CESS-CAST-P-20020501-120-c4|ClusterType=Gen|MentionSpan=22|MentionMisc=CorefType:ident
23  y   y   CCONJ   cc  _   25  cc  25:cc   _
24  un  uno DET di0ms0  Definite=Ind|Gender=Masc|Number=Sing|PronType=Art   25  det 25:det  _
25  25% 25  SYM _   NumForm=Digit   18  conj    18:conj ClusterId=CESS-CAST-P-20020501-120-s5.sn.47|ClusterType=Spec.number|MentionSpan=24-26
26  nuestro nuestro PRON    px1ms0p0    Gender=Masc|Number=Sing|Number[psor]=Plur|Person=1|Poss=Yes|PronType=Prs    25  appos   25:appos    ClusterId[1]=CESS-CAST-P-20020501-120-c3|ClusterType[1]=Spec.organization|MentionSpan[1]=26|MentionMisc[1]=ClusterTypeMismatch:Spec:Spec.organization,CorefType:ident|ClusterId[2]=CESS-CAST-P-20020501-120-s5.p.54|ClusterType[2]=Gen|MentionSpan[2]=26|SpaceAfter=No
27  "   "   PUNCT   fe  PunctType=Quot  18  punct   18:punct    _
28  .   .   PUNCT   fp  PunctType=Peri  8   punct   8:punct _

There are two single-word mention annotations coming from the original data. One of them (c3) is coreferential with the empty subject 10.1. It is also coreferential with 6 mentions in other sentences. The other (CESS-CAST-P-20020501-120-s5.p.54) is a singleton and it probably appeared there because of some named entity annotation at another (higher or lower) constituent.

Now when I run this scenario:

udapy -s read.OldCorefUD corefud.FixCorefUD02 corefud.MergeSameSpan < bug0.conllu > bug1.conllu

the two mentions on line 26 are not merged:

# newdoc
# global.Entity = eid-etype-head-other
# sent_id = CESS-CAST-P-20020501-120-s5
# text = Y Charly, cómo no, se mojó: "Estamos, más o menos, un 75% a favor de ellos y un 25% nuestro" .
1   Y   y   CCONJ   cc  _   8   advmod  8:advmod    _
2   Charly  Charly  PROPN   np00000 _   8   nsubj   8:nsubj Entity=(CESSCASTP20020501120c1-person-1-CorefType:ident,gstype:spec)|SpaceAfter=No
3   ,   ,   PUNCT   fc  PunctType=Comm  4   punct   4:punct _
4   cómo    cómo    ADV _   _   8   dep 8:dep   _
5   no  no  PART    _   _   4   dep 4:dep   SpaceAfter=No
6   ,   ,   PUNCT   fc  PunctType=Comm  4   punct   4:punct _
7   se  él  PRON    p0300000    Case=Acc|Person=3|PrepCase=Npr|PronType=Prs|Reflex=Yes  8   expl:pv 8:expl:pv   _
8   mojó    mojar   VERB    vmis3s0 Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin   0   root    0:root  SpaceAfter=No
9   :   :   PUNCT   fd  PunctType=Colo  18  punct   18:punct    _
10  "   "   PUNCT   _   PunctType=Quot  18  punct   18:punct    SpaceAfter=No
10.1    _   _   PRON    p   _   _   _   11:nsubj    Entity=(CESSCASTP20020501120c3-organization-1-ClusterTypeMismatch:Spec:Spec.organization,CorefType:ident,gstype:spec)
11  Estamos estar   AUX vmip1p0 Mood=Ind|Number=Plur|Person=1|Tense=Pres|VerbForm=Fin   18  cop 18:cop  SpaceAfter=No
12  ,   ,   PUNCT   fc  PunctType=Comm  13  punct   13:punct    _
13  más más ADV rg  Degree=Cmp  18  advmod  18:advmod   _
14  o   o   CCONJ   cc  _   15  cc  15:cc   _
15  menos   menos   ADV rg  Degree=Cmp  13  conj    13:conj SpaceAfter=No
16  ,   ,   PUNCT   fc  PunctType=Comm  13  punct   13:punct    _
17  un  uno DET di0ms0  Definite=Ind|Gender=Masc|Number=Sing|PronType=Art   18  det 18:det  Entity=(CESSCASTP20020501120s5.sn.34-number-2-gstype:spec
18  75% 75  SYM _   NumForm=Digit   8   advcl   8:advcl _
19  a   a   ADP sps00   _   22  case    22:case _
20  favor   favor   NOUN    _   _   19  fixed   19:fixed    _
21  de  de  ADP _   _   19  fixed   19:fixed    _
22  ellos   él  PRON    pp3mp000    Case=Acc,Nom|Gender=Masc|Number=Plur|Person=3|PronType=Prs  18  nmod    18:nmod Entity=(CESSCASTP20020501120c4--1-CorefType:ident,gstype:gen)CESSCASTP20020501120s5.sn.34)
23  y   y   CCONJ   cc  _   25  cc  25:cc   _
24  un  uno DET di0ms0  Definite=Ind|Gender=Masc|Number=Sing|PronType=Art   25  det 25:det  Entity=(CESSCASTP20020501120s5.sn.47-number-2-gstype:spec
25  25% 25  SYM _   NumForm=Digit   18  conj    18:conj _
26  nuestro nuestro PRON    px1ms0p0    Gender=Masc|Number=Sing|Number[psor]=Plur|Person=1|Poss=Yes|PronType=Prs    25  appos   25:appos    Entity=(CESSCASTP20020501120s5.p.54--1-gstype:gen)(CESSCASTP20020501120c3-organization-1-ClusterTypeMismatch:Spec:Spec.organization,CorefType:ident,gstype:spec)CESSCASTP20020501120s5.sn.47)|SpaceAfter=No
27  "   "   PUNCT   fe  PunctType=Quot  18  punct   18:punct    _
28  .   .   PUNCT   fp  PunctType=Peri  8   punct   8:punct _

However, when I save the file and re-read it with another Udapi process, the block succeeds in merging the spans (unfortunately it picks the second mention as the survivor and thus breaks coreference with 10.1 and with the antecedents in other sentences, but that's another issue):

udapy -s read.OldCorefUD corefud.FixCorefUD02 corefud.MergeSameSpan < bug0.conllu | udapy -s corefud.MergeSameSpan > bug1.conllu
# newdoc
# global.Entity = eid-etype-head-other
# sent_id = CESS-CAST-P-20020501-120-s5
# text = Y Charly, cómo no, se mojó: "Estamos, más o menos, un 75% a favor de ellos y un 25% nuestro" .
1   Y   y   CCONJ   cc  _   8   advmod  8:advmod    _
2   Charly  Charly  PROPN   np00000 _   8   nsubj   8:nsubj Entity=(CESSCASTP20020501120c1-person-1-CorefType:ident,gstype:spec)|SpaceAfter=No
3   ,   ,   PUNCT   fc  PunctType=Comm  4   punct   4:punct _
4   cómo    cómo    ADV _   _   8   dep 8:dep   _
5   no  no  PART    _   _   4   dep 4:dep   SpaceAfter=No
6   ,   ,   PUNCT   fc  PunctType=Comm  4   punct   4:punct _
7   se  él  PRON    p0300000    Case=Acc|Person=3|PrepCase=Npr|PronType=Prs|Reflex=Yes  8   expl:pv 8:expl:pv   _
8   mojó    mojar   VERB    vmis3s0 Mood=Ind|Number=Sing|Person=3|Tense=Past|VerbForm=Fin   0   root    0:root  SpaceAfter=No
9   :   :   PUNCT   fd  PunctType=Colo  18  punct   18:punct    _
10  "   "   PUNCT   _   PunctType=Quot  18  punct   18:punct    SpaceAfter=No
10.1    _   _   PRON    p   _   _   _   11:nsubj    Entity=(CESSCASTP20020501120c3-organization-1-ClusterTypeMismatch:Spec:Spec.organization,CorefType:ident,gstype:spec)
11  Estamos estar   AUX vmip1p0 Mood=Ind|Number=Plur|Person=1|Tense=Pres|VerbForm=Fin   18  cop 18:cop  SpaceAfter=No
12  ,   ,   PUNCT   fc  PunctType=Comm  13  punct   13:punct    _
13  más más ADV rg  Degree=Cmp  18  advmod  18:advmod   _
14  o   o   CCONJ   cc  _   15  cc  15:cc   _
15  menos   menos   ADV rg  Degree=Cmp  13  conj    13:conj SpaceAfter=No
16  ,   ,   PUNCT   fc  PunctType=Comm  13  punct   13:punct    _
17  un  uno DET di0ms0  Definite=Ind|Gender=Masc|Number=Sing|PronType=Art   18  det 18:det  Entity=(CESSCASTP20020501120s5.sn.34-number-2-gstype:spec
18  75% 75  SYM _   NumForm=Digit   8   advcl   8:advcl _
19  a   a   ADP sps00   _   22  case    22:case _
20  favor   favor   NOUN    _   _   19  fixed   19:fixed    _
21  de  de  ADP _   _   19  fixed   19:fixed    _
22  ellos   él  PRON    pp3mp000    Case=Acc,Nom|Gender=Masc|Number=Plur|Person=3|PronType=Prs  18  nmod    18:nmod Entity=(CESSCASTP20020501120c4--1-CorefType:ident,gstype:gen)CESSCASTP20020501120s5.sn.34)
23  y   y   CCONJ   cc  _   25  cc  25:cc   _
24  un  uno DET di0ms0  Definite=Ind|Gender=Masc|Number=Sing|PronType=Art   25  det 25:det  Entity=(CESSCASTP20020501120s5.sn.47-number-2-gstype:spec
25  25% 25  SYM _   NumForm=Digit   18  conj    18:conj _
26  nuestro nuestro PRON    px1ms0p0    Gender=Masc|Number=Sing|Number[psor]=Plur|Person=1|Poss=Yes|PronType=Prs    25  appos   25:appos    Entity=(CESSCASTP20020501120s5.p.54--1-gstype:gen)CESSCASTP20020501120s5.sn.47)|SpaceAfter=No
27  "   "   PUNCT   fe  PunctType=Quot  18  punct   18:punct    _
28  .   .   PUNCT   fp  PunctType=Peri  8   punct   8:punct _
dan-zeman commented 2 years ago

In the problematic run (with the conversion blocks), when iterating coreference mentions in corefud.MergeSameSpan, only two mentions are found. These are multi-word mentions; single word mentions are invisible.

2022-02-10 11:37:07,084 [   INFO] process_tree - mA: CESSCASTP20020501120s5.sn.47 ['24:un', '25:25%', '26:nuestro']
2022-02-10 11:37:07,084 [   INFO] process_tree - mB: CESSCASTP20020501120s5.sn.34 ['17:un', '18:75%', '19:a', '20:favor', '21:de', '22:ellos']
martinpopel commented 2 years ago

I'll investigate it. Thanks for reporting.