Annotating tense, mood and voice for English, French and German

What it is about

The paper presents an open-source rule tool that labels each verbal complex (finite and non-finite verb groups) with tense, mood and voice, for English, French and German. It reads the dependency trees of the Mate parser and the POS tags of the verbs; the rules are applied after the verb group has been collected through dependency links. There are about 32 rules for English and 26 each for German and French. Manual checks on random Europarl sentences (151 English, 157 German, 137 French verbal complexes) give about 76% accuracy for the three features together; for English, tense is 81.5%, mood 88.1% and voice 86.1%. Most of the errors that were analysed come from the parser, for example a gerund being added to the verbal complex as a present participle. Disambiguation between a stative passive and a perfect uses lists of verbs that take "be" as their auxiliary.

Ideas

I1. en/src/morph.rs and en/src/parse.rs: a closed enum of English tense-mood-voice combinations (the paper's Table 1 lists them, from present to "would have been working"), assigned from the auxiliary chain. Test on a hand-annotated set of 150 verbal complexes, held out until the release test, with a control that assigns tense at random.
I2. Feed the verbal complex's tense into event time in world/src/events.rs, so that "will have read" is recorded as a future-perfect event; check the effect on the bAbI temporal tasks in world/src/babi.rs.
I3. Parser-error guard: a test list in en where a gerund after an auxiliary is not added to the verb group ("is being examined" versus "examining"). Report the rate on UD EWT before and after.

For Mova

The English chapters are 13. Verbs (auxiliary chains), 28-voice-and-passive.md (the passive test) and 29-ellipsis-word-order-information-structure.md if needed. The code is en/src/morph.rs and en/src/parse.rs, and events go to world/src/events.rs. The measure is accuracy on a hand-checked set, plus the control. The paper's accuracy is the baseline to beat, not a target.

Cite

The paper itself, not this card. BibTeX from the library's record of the work.

@inproceedings{ramm2017annotating,
  title = {{Annotating tense, mood and voice for English, French and German}},
  author = {Ramm, Anita and Loáiciga, Sharid and Friedrich, Annemarie and Fraser, Alexander},
  booktitle = {Proceedings of ACL 2017, System Demonstrations},
  year = {2017},
  url = {https://aclanthology.org/P17-4001/}
}

References

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

Uladzimir Sidarenka, Andreas Peldszus, and Manfred Stede. 2015. Discourse segmentation of German texts. Journal for Language Technology and Computational Linguistics 30(1):71–98. find
Bernd Bohnet and Joakim Nivre. 2012. A transition-based system for joint part-of-speech tagging and labeled non-projective dependency parsing. In Proceedings of the 2012 Joint Conference on EMNLP. Jeju Island, Korea. find
Naushad UzZaman, Hector Llorens, Leon Derczynski, James Allen, Marc Verhagen, and James Pustejovsky. 2013. SemEval-2013 Task 1: TempEval-Events, and Temporal Relations. In Proceedings of the SemEval 2013. Atlanta, Georgia. find
Francisco Costa and António Branco. 2012. Aspectual type and temporal relation classification. In Proceedings of the 13th Conference of the EACL. Avignon, France. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/ramm-2017-annotating-tense-mood-voice-english/