A POS Tagging Model Designed for Learner English

What it is about

Taggers trained on edited English fail on learner English, and the authors trace the failures to unknown words (often misspellings), to capitalisation errors, to missing determiners, and to a different tag distribution (the word "like" is a preposition in 82% of the native corpus but a verb in 94% of the Konan-JIEM learner data). Their model is a bidirectional LSTM over word embeddings, character embeddings and a native-language vector, trained on Penn Treebank sections 00–18 and on learner text pre-trained without tags. It reaches 0.964 accuracy on the Treebank of Learner English test set, above the Turbo tagger's 0.958 with learner training data, and 0.948 on the Konan-JIEM corpus, where a CRF gets 0.940. The character module adds 1.0 point, and pre-training adds 0.3 to 0.6 points. For "like" the CRF reaches 0.635 on 304 instances and the model 0.927. The native-language module has almost no effect; the authors report this as a negative result.

Ideas

• I1. en/src/ptag.rs: a count table of tags per word form in learner text, written as a BTreeMap, used as a prior only when the word's native-corpus and learner-corpus tag distributions differ strongly. Start with the "like" case and report it per word.
• I2. Character-level features for misspellings: a closed set of edit-distance-1 mappings (Becose to Because, thier to their) to the nearest lexicon form, stored as explicit rules with a reason string, so each recovered tag says which correction produced it.
• I3. Capitalisation rule: a sentence-initial bare noun with no determiner ("Flower is ...") is a candidate for a missing determiner, not a proper noun. Test it as a rule in en/src/tag.rs on the paper's examples and on held-out learner text.
• I4. The paper's native-language module is a negative result. Before any native-language table is built, run the ablation (module off, same data, same seeds) and report the difference, as the value-ablation practice requires.

For Mova

The code points are en/src/ptag.rs (perceptron PTB tagger, where the word prior and character features would sit), en/src/tok.rs for tokens with capitals, and en/src/tag.rs. The chapters are 06. Spelling rules: inflection, derivation, compounds and national standards and 10. Word classes. The measure: tag accuracy on a learner set Mova owns, split into the "like" cases and the rest, with the ablation of each rule run twice for determinism.

Topics: Part-of-speech tagging

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/w18-6106

@inproceedings{nagata2018pos,
  title = {{A POS Tagging Model Designed for Learner English}},
  author = {Nagata, Ryo and Mizumoto, Tomoya and Kikuchi, Yuta and Kawasaki, Yoshifumi and Funakoshi, Kotaro},
  booktitle = {Proceedings of the 2018 EMNLP Workshop W-NUT: The 4th Workshop on Noisy User-generated Text},
  year = {2018},
  doi = {10.18653/v1/w18-6106},
  url = {https://doi.org/10.18653/v1/w18-6106}
}

References

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

2. Jan Aarts and Sylviane Granger. 1998. Tag sequences in learner corpora: a key to interlanguage grammar and discourse. Longman, New York. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/nagata-2018-pos-tagging-model-adapted-learner/