A non-projective greedy dependency parser with bidirectional LSTMs

What it is about

BIST-COVINGTON is a greedy parser that implements Covington's algorithm for unrestricted non-projective dependency parsing as a transition system, with O(n^2) complexity. The scorer is a bidirectional LSTM in the style of Kiperwasser and Goldberg, trained with a dynamic oracle to reduce error propagation. The submission used the official UDPipe 1.1 segmentation and tagging, no ensemble and no custom tokenization. It ranked 7th of 33 teams on the CoNLL 2017 big-treebank category (55 treebanks), and 16th and 12th in LAS and UAS over all treebanks. The authors trace their weaker "all" results to four parallel PUD treebanks, where suffixed treebank models transfer poorly.

Ideas

• I1. en/src/parse.rs: add Covington's non-projective transition system as a second decoder over the same features; measure LAS and UAS on EWT test and on GUM test (test only), and the share of gold non-projective arcs that are recovered.
• I2. gparse/src/eisner.rs: measure how many gold EWT test arcs are non-projective and therefore cannot be produced by the Eisner decoder; this sets whether a non-projective decoder is worth the run in I1.
• I3. Runtime of the two decoders on EWT test per 1,000 sentences, to check the O(n^2) claim of the paper against the beam-8 arc-standard search in en.

For Mova

Non-projective arcs are the displacement phenomenon of 29-ellipsis-word-order-information-structure and of extraposed clauses in 25-relative-clauses, and the general syntax is in 20-syntax-overview. The code touched is en/src/parse.rs and gparse/src/eisner.rs. Measure first: the non-projective share and the Eisner ceiling of idea I2, since they decide whether idea I1 is worth a run.

Topics: Dependency parsing · Treebanks

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/k17-3016 · arXiv 1707.03228

@inproceedings{vilares2017non,
  title = {{A non-projective greedy dependency parser with bidirectional LSTMs}},
  author = {Vilares, David and Gómez-Rodríguez, Carlos},
  booktitle = {Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies},
  year = {2017},
  doi = {10.18653/v1/k17-3016},
  eprint = {1707.03228},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/k17-3016}
}

References

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

1. Željko Agić and Natalie Schluter. 2017. How (not) to train a dependency parser: The curious case of jackknifing part-of-speech taggers. In The 54th Annual Meeting of the Association for Computational Linguistics (ACL 2017). find
2. Giuseppe Attardi. 2006. Experiments with a mul-tilanguage non-projective dependency parser. In Proceedings of the Tenth Conference on Computational Natural Language Learning. Association for Computational Linguistics, pages 166–170. http://dl.acm.org/citation.cfm?id=1596276.1596307. ↗
3. Danqi Chen and Christopher Manning. 2014. A fast and accurate dependency parser using neural networks. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). Doha, Qatar, pages 740–750. http://www.aclweb.org/anthology/D14-1082. \#\# Conclusion ↗ our card

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/ud-2017-vilares-non-projective-greedy-dependency-parser/