Dependency Parsing by Inference over High-recall Dependency Predictions

1
Dependency Parsing by Inference over High-recall
Dependency Predictions

• Sander Canisius S.V.M.Canisius_at_uvt.nl
• Toine Bogers A.M.Bogers_at_uvt.nl
• Antal van den Bosch Antal.vdnBosch_at_uvt.nl
• Jeroen Geertzen J.Geertzen_at_uvt.nl
• ILK / Language and Information Science
• Tilburg University
• Erik Tjong Kim Sang erikt_at_science.uva.nl
• Informatics Institute
• University of Amsterdam

2
Highlights

• No modelling, just classification
• Simultaneously predicting and labelling
  dependency relations
• Resolving inconsistencies on the basis of
  classifier confidence

3
vc
obj1
su
ik
hoor
haar
zingen
Dependent Head Relation
ik hoor SU
ik haar -
ik zingen -
4
vc
obj1
su
ik
hoor
haar
zingen
Dependent Head Relation
hoor ik -
hoor haar -
hoor zingen -
5
vc
obj1
su
ik
hoor
haar
zingen
Dependent Head Relation
haar ik -
haar hoor OBJ1
haar zingen -
6
vc
obj1
su
ik
hoor
haar
zingen
Dependent Head Relation
zingen ik -
zingen hoor VC
zingen haar -
7
ik hoor haar zingen
ik hoor haar zingen
SU-hoor - OBJ1-hoor / DET-zingen VC-hoor
8
ik hoor haar zingen
ik hoor haar zingen
SU-hoor - OBJ1-hoor0.8 / DET-zingen0.5 VC-hoor
9
ik hoor haar zingen
ik hoor haar zingen
SU-hoor - OBJ1-hoor VC-hoor
10
Features

• Head features
• 2-1-2 word part-of-speech windows
• Dependent features
• 2-1-2 word part-of-speech windows
• Relative position (LEFT / RIGHT)
• Distance

11

• The relation prediction/classification has a
  highly skewed class distribution
• Tends to result in high-precision, low-recall
  relation predictions
• Down-sampling the negative class increases recall
• (At the cost of precision)

12
(No Transcript)
13
Language UAS LAS Arabic 74.59 57.64 Bulgarian
82.51 78.74 Chinese 82.86 78.37 Czech 72.88
60.92 Danish 82.93 77.90 Dutch
77.79 74.59 German 80.01 77.56 Japanese
89.67 87.41 Portuguese 85.61 77.42 Slovene
74.02 59.19 Spanish 71.33 68.32 Swedish
85.08 79.15 Turkish 64.19 51.07 Average 78.41
70.80