Results

1
Results see proceedings for full table
Ar Ch Cz Da Du Ge Ja Po Sl Sp Sw Tu Tot SD Bu
McD 66.9 85.9 80.2 84.8 79.2 87.3 90.7 86.8 73.4 82.3 82.6 63.2 80.3 8.4 87.6
Niv 66.7 86.9 78.4 84.8 78.6 85.8 91.7 87.6 70.3 81.3 84.6 65.7 80.2 8.5 87.4
ON 66.7 86.7 76.6 82.8 77.5 85.4 90.6 84.7 71.1 79.8 81.8 57.5 78.4 9.4 85.2
Rie 66.7 90.0 67.4 83.6 78.6 86.2 90.5 84.4 71.2 77.4 80.7 58.6 77.9 10.1 0.0
Sag 62.7 84.7 75.2 81.6 76.6 84.9 90.4 86.0 69.1 77.7 82.0 63.2 77.8 9.0 0.0
Che 65.2 84.3 76.2 81.7 71.8 84.1 89.9 85.1 71.4 80.5 81.1 61.2 77.7 8.7 86.3
Cor 63.5 79.9 74.5 81.7 71.4 83.5 90.0 84.6 72.4 80.4 79.7 61.7 76.9 8.5 83.4

Av 59.9 78.3 67.2 78.3 70.7 78.6 85.9 80.6 65.2 73.5 76.4 56.0 80.0
SD 6.5 8.8 8.9 5.5 6.7 7.5 7.1 5.8 6.8 8.4 6.5 7.7 6.3
2
Results continued

• Good parsers good on all languages
• Best overall scores achieved by two very
  different approaches
• Little difference in ranking (mostly just /-1)
  when using UAS or label accuracy metric instead
  of LAS
• except Two groups with special emphasis on
  DEPREL values score 2/3 ranks for label
  accuracy one group with bug in HEAD assignment
  scores 4 ranks for UAS
• Very little difference in scores as well as
  rankings when scoring all tokens (i.e. including
  punctuation)
• But some outliers in ranking for individual
  languages
• Turkish Johansson and Nugues 7, Yuret 7,
  Riedel et al. -5
• Dutch Canisius et al. 6, Schiehlen and Spranger
  8

3
Analysis easy data sets
Ar Ch Cz Da Du Ge Ja Po Sl Sp Sw Tu Bu
Top score 66.9 90.0 80.2 84.8 79.2 87.3 91.7 87.6 73.4 82.3 84.6 65.7 87.6
Av. score 59.9 78.3 67.2 78.3 70.7 78.6 85.9 80.6 65.2 73.5 76.4 56.0 80.0
Tokens (k) 54 337 1249 94 195 700 151 207 29 89 191 58 190
Tok./tree 37.2 5.9 17.2 18.2 14.6 17.8 8.9 22.8 18.7 27.0 17.3 11.5 14.8
DEP./CPOS 1.9 6.3 6.5 5.2 2 .88 .35 3.7 2.3 1.4 1.5 1.8 1.6
DEP./POS 1.4 .28 1.2 2.2 .09 .88 .09 2.6 .89 .55 1.5 .83 .34
H. prec. 82.9 24.8 50.9 75.0 46.5 50.9 8.9 60.3 47.2 60.8 52.8 6.2 62.9
H. foll. 11.6 58.2 42.4 18.6 44.6 42.7 72.5 34.6 46.9 35.1 40.7 80.4 29.2
np trees 11.2 0.0 23.2 15.6 36.4 27.8 5.3 18.9 22.2 1.7 9.8 11.6 5.4
new FOR. 17.3 9.3 5.2 18.1 20.7 6.5 0.96 11.6 22.0 14.7 18.0 41.4 14.5
new LEM. 4.3 n/a 1.8 n/a 15.9 n/a n/a 7.8 9.9 9.7 n/a 13.2 n/a
4
Analysis difficult data sets
Ar Ch Cz Da Du Ge Ja Po Sl Sp Sw Tu Bu
Top score 66.9 90.0 80.2 84.8 79.2 87.3 91.7 87.6 73.4 82.3 84.6 65.7 87.6
Av. score 59.9 78.3 67.2 78.3 70.7 78.6 85.9 80.6 65.2 73.5 76.4 56.0 80.0
Tokens(k) 54 337 1249 94 195 700 151 207 29 89 191 58 190
Tok./tree 37.2 5.9 17.2 18.2 14.6 17.8 8.9 22.8 18.7 27.0 17.3 11.5 14.8
DEP./CPOS 1.9 6.3 6.5 5.2 2 .88 .35 3.7 2.3 1.4 1.5 1.8 1.6
DEP./POS 1.4 .28 1.2 2.2 .09 .88 .09 2.6 .89 .55 1.5 .83 .34
H. prec. 82.9 24.8 50.9 75.0 46.5 50.9 8.9 60.3 47.2 60.8 52.8 6.2 62.9
H. foll. 11.6 58.2 42.4 18.6 44.6 42.7 72.5 34.6 46.9 35.1 40.7 80.4 29.2
np trees 11.2 0.0 23.2 15.6 36.4 27.8 5.3 18.9 22.2 1.7 9.8 11.6 5.4
new FOR. 17.3 9.3 5.2 18.1 20.7 6.5 0.96 11.6 22.0 14.7 18.0 41.4 14.5
new LEM. 4.3 n/a 1.8 n/a 15.9 n/a n/a 7.8 9.9 9.7 n/a 13.2 n/a
5
The future Using the resources Parsed test
sets

• Parser combination
• Check test cases where parser majority disagrees
  with treebank
• Possible reasons
• Challenge for current parser technology
• Treebank annotation wrong
• Conversion wrong
• (Non-sentence)
• for German test data checked cases where 17 out
  of 18 parsers said DEPRELX but gold standard had
  DEPRELY (11 cases)
• 4 challenge distinguishing PP complements from
  adjuncts
• 1 treebank PoS tag wrong
• 1 non-sentence
• 5 either treebank or conversion wrong (to be
  investigated)

6
The future Using the resources Parsers

• Collaborate with treebank providers or other
  treebank experts
• To semi-automatically enlarge or improve the
  treebank
• To use automatically parsed texts as input for
  other NLP projects
• Please let us know if you are willing to help
• Arabic Otakar Smrž, Jan Hajic also general
  feedback
• Bulgarian Petya Osenova, Kiril Simov
• Czech Jan Hajic
• German Martin Forst, Michael Schiehlen, Kristina
  Spranger
• Portuguese Eckhard Bick
• Slovene Tomaž Erjavec
• Spanish Toni Martí, Roser Morante
• Swedish Joakim Nivre
• Turkish Gülsen Eryigit
• Gertjan van Noord (Dutch Alpino treebank)
  interested in general feedback

7
The future Improving the resources

• http//nextens.uvt.nl/conll/ is a static web
  page
• But hopefully many other people will continue
  this line of research
• They need a platform to exchange (information
  about)
• Experience with/bug reports about/patches for
  treebanks
• Treebank conversion and validation scripts (esp.
  head tables!)
• Other new/improved tools (e.g. visualization,
  analysis)
• Details of experiments on new treebanks (e.g.
  training/test split)
• Predictions on test sets by new/improved parsers
• ...
• SIGNLL agreed to provide such a platform (hosted
  at Tilburg University)
• http//depparse.uvt.nl/ a Wiki site where
  everybody is welcome to contribute!

8
Acknowledgements Many thanks to

• Jan Hajic for granting the PDT/PADT temporary
  licenses for CoNLL-X and talking to LDC about it
• Christopher Cieri for arranging distribution
  through LDC and Tony Castelletto for handling the
  distribution
• Otakar Smrž for valuable help during the
  conversion of PADT
• the SDT people for granting the special license
  for CoNLL-X and Tomaž Erjavec for converting the
  SDT for us
• Matthias Trautner Kromann and assistants for
  creating the DDT and releasing it under the GNU
  General Public License
• Joakim Nivre, Johan Hall and Jens Nilsson for the
  conversion of DDT to Malt-XML, for the conversion
  of the original Talbanken to Talbanken05 and for
  making it freely available for research purposes
• Joakim Nivre again for prompt and proper response
  to all our questions
• Bilge Say and Kemal Oflazer for granting the
  Metu-Sabanci license for CoNLL-X and answering
  questions
• Gülsen Eryigit for making many corrections to
  Metu-Sabanci and discussing some aspects of the
  conversion

9
Acknowledgements continued

• the TIGER team (esp. Martin Forst) for allowing
  us to use the treebank for the shared task
• Yasuhiro Kawata, Julia Bartels and colleagues
  from Tübingen University for the construction of
  the Japanese Verbmobil treebank
• Sandra Kübler for providing the Japanese
  Verbmobil data and granting the special license
  for CoNLL-X
• Diana Santos, Eckhard Bick and other Floresta
  sintá(c)tica project members for creating the
  treebank and making it publicly available, for
  answering many questions about the treebank
  (Diana and Eckhard), for correcting problems and
  making new releases (Diana), and for sharing
  scripts and explaining the head rules implemented
  in them (Eckhard)
• Jason Baldridge for useful discussions and to Ben
  Wing for independently reporting problems which
  Diana then fixed
• Gertjan van Noord and the other people at
  Groningen University for creating the Alpino
  Treebank and releasing it for free
• Gertjan van Noord for answering all our questions
  and for providing extra test material
• Antal van den Bosch for help with the
  memory-based tagger

10
Acknowledgements continued

• Academia Sinica for granting the Sinica treebank
  temporary license for CoNLL-X
• Keh-Jiann Chen for answering our questions about
  the Sinica treebank
• Montserrat Civit and Toni Martí for allowing us
  to use Cast3LB for CoNLL-X and supplying the head
  table and function mapping
• Kiril Simov and Petya Osenova for allowing us to
  use the BulTreeBank for CoNLL-X
• Svetoslav Marinov, Atanas Chanev, Kiril Simov and
  Petya Osenova for converting the BulTreeBank
• Dan Bikel for making the Randomized Parsing
  Evaluation Comparator
• SIGNLL for having a shared task
• Erik Tjong Kim Sang for posting the Call for
  Papers
• Lluís Màrquez and Xavier Carreras for sharing
  their experience from organizing the two previous
  shared tasks
• Lluís Màrquez for being a very helpful CoNLL
  organizer
• All participants, including those who could not
  submit results or cannot be here today
• It has been a pleasure working with you!

11
Future research

• More result analysis
• Baseline
• Correlation between parsing approach and types of
  errors
• Importance of individual features and algorithm
  details
• Repeat experiments on improved data
• POSTAG and FEATS for Talbanken05, LEMMA for
  Talbanken05 (DDT)
• LEMMA and FEATS for new version of TIGER, POSTAG
  for TIGER
• better DEPREL for Cast3LB
• larger treebanks
• having several good parsers facilitates
  annotation of more text!
• better quality
• check cases where parsers and treebank disagree!

12
Future research continued

• Repeat experiments with other parsers
• http//nextens.uvt.nl/conll/post_task_data.html
• Repeat experiments with additional external
  information
• large-scale distributional data harvested from
  the internet
• Similar experiments but including secondary
  relations
• Similar experiments on other data
• Hebrew treebank
• SUSANNE treebank (English)
• Kyoto University Corpus, ATR corpus (Japanese)
• Szeged Corpus (Hungarian)
• see list in Wikipedia article on treebank
  (please contribute!)
• Integrate automatic tokenization, morphological
  analysis and POS tagging