Finding The Right One and Resolving it

What it is about

One-anaphora is the use of the word "one" or "ones" for a noun head, as in "the black one". The paper separates it from the determinative use of "one" as a numeral, a confusion that has weakened earlier computational work. The authors annotate one-anaphora with their antecedents in the British National Corpus and release a gold corpus of 921 instances. Template rules over POS strings, run on spaCy parses, reach 78.34% precision, 85.92% recall and 81.96% F1 for detection. Neural classifiers with fastText and BERT embeddings are then trained on the corpus. From the templates, 15,647 BNC sentences contain 18,669 one-anaphora words, so the sentence-level rate is 6.25%, far below the 12–15% reported in earlier annotation efforts.

Ideas

• I1. en/src/expert.rs: a deterministic detector for "one" and "ones" that uses the UD labels to tell the nominal head from the numeral. Report precision and recall on the 921-instance corpus, test only, since the BNC is NC.
• I2. Determinative split: count how many "one" tokens on EWT test are numerals (det with a number reading) and how many are nominal heads. The paper's point is that mixing the two inflates the counts, so the split should be reported first.
• I3. Coreference hook: link a one-anaphor to the nearest noun head with compatible modifiers, using the candidate set of the coref crate. Measure resolution accuracy on the test part only, and report the paper's own baseline next to it.

For Mova

The code is en/src/expert.rs, coref and en/src/ctx.rs. The chapters are 35. Reference and coreference and 22. Phrases: noun, adjective, adverb and prepositional phrases. The gold corpus comes from the NC BNC, so it is for measurement only. Measure first: the detector against the gold corpus, with the determinative split reported.

Topics: Coreference resolution

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2020.conll-1.10

@inproceedings{khullar2020finding,
  title = {{Finding The Right One and Resolving it}},
  author = {Khullar, Payal and Bhattacharya, Arghya and Shrivastava, Manish},
  booktitle = {Proceedings of the 24th Conference on Computational Natural Language Learning},
  year = {2020},
  doi = {10.18653/v1/2020.conll-1.10},
  url = {https://doi.org/10.18653/v1/2020.conll-1.10}
}

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.

1. Nameera Akhtar, Maureen Callanan, Geoffrey K Pullum, and Barbara C Scholz. 2004. Learning antecedents for anaphoric one. Cognition, 4:141–145. 65.49 64.22 78.87 79.35 71.74 89.35 71.76 67.77 83.78 55.24 58.97 63.07 61.29 65.74 72.33 58.11 62.17 67.38 find
2. Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2016. Enriching word vectors with subword information. arXiv preprint arXiv:1607.04606. 59.99 70.01 64.61 find
3. Joan Bresnan. 1971. A note on the notion “identity of sense anaphora”. Linguistic Inquiry, 2:589–597. find
4. Carl Lee Baker. 1978. Introduction to generative transformational syntax. Englewood Cliffs, NJ:: Prentice-Hal. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/lingbulk-2020-khullar-finding-right-one-resolving-it/