Joint Prediction of Morphosyntactic Categories for Fine-Grained Arabic Part-of-Speech Tagging Exploiting Tag Dictionary Information

What it is about

A fine-grained Arabic POS tag is a conjunction of values in up to 14 morphosyntactic categories (gender, case, mood, proclitics, enclitic and others), so the number of possible tags is very large. The authors train bi-LSTM taggers, first one model per category, then a joint multi-task model with shared hidden layers. They also add tag-dictionary information as a soft feature: for each word and category, the sum of embeddings of the tags an analyzer allows, instead of a hard filter. On the Penn Arabic Treebank the full-tag accuracy is 87.74 for independent models, 90.17 with the dictionary, 89.49 for the joint model and 91.38 for the joint model with the dictionary. That is 2.11 points above CamelParser (89.27). The same architecture is run on the Arabic UD treebank with 17 categories. The authors note that a hard analyzer constraint cannot recover a tag the analyzer did not return.

Ideas

• I1. Use the lexicon as a soft feature in the tools/hbo gparse input: for each token, the set of values per FEATS category that the lexicon allows (macula dict), added as a feature and never as a filter. A word with no entry gets an empty set, so the right analysis can still win. Where: tools/hbo, described in 13. Biblical Hebrew in Mova: the tools.
• I2. Report per-category accuracy and whole-analysis accuracy together, as the paper's tables do, on treebanks/hbo/ptnk and treebanks/he/. A gain in gender must not hide a loss in case; a gain is claimed only after two identical runs.
• I3. Use the per-category table as an error template for tools/hbo: which features are easy (clitics, prefixes) and which are hard (case, mood), and the share of errors that come from a missing candidate in the lexicon, which the paper reports as a large part of its baseline errors.

For Mova

I1 and I3 go into tools/hbo (the feature input and its error analysis). I2 is the evaluation for treebanks/hbo/ptnk and treebanks/he/. Case-related checks belong to The Noun, the Adjective and the Numeral.

Topics: Treebanks · Part-of-speech tagging

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/k17-1042

@inproceedings{inoue2017joint,
  title = {{Joint Prediction of Morphosyntactic Categories for Fine-Grained Arabic Part-of-Speech Tagging Exploiting Tag Dictionary Information}},
  author = {Inoue, Go and Shindo, Hiroyuki and Matsumoto, Yuji},
  booktitle = {Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017)},
  year = {2017},
  doi = {10.18653/v1/k17-1042},
  url = {https://doi.org/10.18653/v1/k17-1042}
}

References

3 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Joachim Bingel and Anders Søgaard. 2017. Identifying beneficial task relations for multi-task learning in deep neural networks. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers. Association for Computational Linguistics, Valencia, Spain, pages 164–169. http://www.aclweb.org/anthology/E17-2026. ↗
2. Bernd Bohnet, Joakim Nivre, Igor Boguslavsky, Richárd Farkas, Filip Ginter, and Jan Hajič. 2013. Joint Morphological and Syntactic Analysis for Richly Inflected Languages. Transactions of the Association for Computational Linguistics 1:415–428. find
3. Tim Buckwalter. 2002. Buckwalter Arabic Morphological Analyzer Version 1.0 LDC2002L49. Linguistic Data Consortium (LDC, Philadelphia US). find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/lingbulk-2017-inoue-joint-prediction-morphosyntactic-categories-fine/