Initial Explorations of CCG Supertagging for Universal Dependency Parsing

Summary

The METU team from Middle East Technical University describes its system for the CoNLL 2017 Shared Task. The task of the competition is multilingual parsing in the UD scheme, starting from raw text. The authors test whether CCG supertags help as additional features for a dependency parser in languages that have no CCG-annotated resources. First, the English CCGbank was aligned with the Penn Treebank converted to the CoNLL-U format. On these data, they trained a classifier that predicts a word's CCG category solely from the universal POS tags and dependency labels in its neighborhood. The categories were deliberately simplified: slash directions and features were removed so as not to carry English word order over to other languages. This classifier was used to annotate the train and dev portions of all treebanks. Then, for each language separately, they trained a CRF supertagger and a neural greedy transition-based parser in the style of Chen and Manning (2014). The system's average LAS F1 is 61.98. Results on large corpora are considerably better than on small and "surprise" languages. The paper is of interest as a bold experiment in transferring a lexicalized grammar from English to other languages.

Key points

An example of the simplification: (S[dcl]\NP)/(S[b]\NP) → (S-NP)-(S-NP). A logistic regression classifier (Scikit-Learn) recovers English CCG categories with an accuracy of 91.78%. With XPOS features and argument positions it would reach 93.55%, but these features are not universal, so they were dropped.
The CRF supertagger is built on CRFSuite. Its features are: the word, POS, FEATS, prefixes and suffixes of up to 5 characters, and a ±2-token window. On CCGbank (training on sections 02–05, testing on section 00) it achieves an accuracy of 93.74%.
Macro LAS F1 by category: big treebanks — 68.77, PUD — 65.30, small — 30.84, surprise languages — 19.39. The authors attribute the weak results on small corpora to the fact that each model was trained only on data from a single language.
For languages for which no dedicated model could be trained, the model of a related language was used: bxr→ru, kmr→fa, sme→fi, hsb→cs, kk/ug→tr. Ukrainian (uk) was parsed by the Russian model, and LAS was only 31.68. For comparison: ru — 71.40, ru_syntagrus — 80.88.
The best LAS scores were obtained for hi (84.54), pt_br (83.17) and it (83.01). The worst were for ga (13.26), ug (15.66) and kmr (15.86).

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Linguistics & languages

Topics: Dependency parsing · Treebanks

Cited by

1 paper of our library cites it (of the 140673 whose bibliographies we read); each has a page here.

CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/k17-3023

@inproceedings{akkus2017initial,
  title = {{Initial Explorations of CCG Supertagging for Universal Dependency Parsing}},
  author = {Akkus, Burak Kerim and Azizoglu, Heval and Cakici, Ruket},
  booktitle = {Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies},
  year = {2017},
  doi = {10.18653/v1/k17-3023},
  url = {https://doi.org/10.18653/v1/k17-3023}
}

References

4 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Bharat Ram Ambati, Tejaswini Deoskar, and Mark Steedman. 2013a. Using CCG categories to improve hindi dependency parsing. In Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, ACL 2013, 4-9 August 2013, Sofia, Bulgaria, Volume 2: Short Papers. pages 604– 609. find
Bharat Ram Ambati, Tejaswini Deoskar, and Mark Steedman. 2013b. Using ccg categories to improve hindi dependency parsing. In ACL (2). pages 604– 609. find
Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev, Slav Petrov, and Michael Collins. 2016. Globally normalized transition-based neural networks. arXiv preprint arXiv:1603.06042 . find our card
Michael Auli and Adam Lopez. 2011. Efficient ccg parsing: A* versus adaptive supertagging. In Dekang Lin, Yuji Matsumoto, and Rada Mihalcea, editors, ACL ’11: Proceedings of the 49th Annual Meeting on Association for Computational Linguistics. The Association for Computer Linguistics, pages 1577–158… find

This page in Markdown → https://md.research.elamur.ai/free/en/papers/akkus-2017-initial-explorations-ccg-supertagging-universal/

📄 Initial Explorations of CCG Supertagging for Universal Dependency Parsing