Ellipsis Resolution as Question Answering: An Evaluation

What it is about

The paper treats ellipsis resolution as a reading-comprehension problem. The elided clause becomes a question, and the antecedent becomes the answer span in the preceding text. The authors train BERT-based models on sluice ellipsis and VP ellipsis, first alone and then jointly with question-answering and coreference data. For sluice ellipsis, the F1 rises from 70.00 to 86.01. For VP ellipsis it rises from 72.89 to 78.66. Joint training with coreference data adds up to 2.9 points for VP ellipsis. Combining the two ellipsis datasets gives about one point to each. The reported results average three runs with different random seeds.

Ideas

• I1. en/src/ctx.rs: for each ellipsis site, build the question as a string from the wh-remnant and the auxiliary, then accept a candidate antecedent only if its tree answers that question (same head, compatible clause type). This is a deterministic check of the paper's idea, with no neural model at test time. Measure on a test set whose licence allows it.
• I2. Offline silver data: a large model may label ellipsis sites in batches, offline, for an annot gold check. It may not answer at test time. Check a sample of 100 labels by hand before they are used.
• I3. Split the report by ellipsis type, sluice and VP, as the paper does, in every ellipsis report, so that a gain on one type cannot hide a loss on the other.

For Mova

The code is en/src/ctx.rs, prag and annot. The chapter is 29. Ellipsis, word order and information structure. The paper's models are neural, so only the question form and the check are a Mova analogue, not a port. Measure first: the question-based check against the current candidate choice, on a test set we may use.

Topics: Coreference resolution

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2021.eacl-main.68

@inproceedings{aralikatte2021ellipsis,
  title = {{Ellipsis Resolution as Question Answering: An Evaluation}},
  author = {Aralikatte, Rahul and Lamm, Matthew and Hardt, Daniel and Søgaard, Anders},
  booktitle = {Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume},
  year = {2021},
  doi = {10.18653/v1/2021.eacl-main.68},
  url = {https://doi.org/10.18653/v1/2021.eacl-main.68}
}

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.

1. ## Related Work Mostafa Abdou, Cezar Sas, Rahul Aralikatte, Isabelle Augenstein, and Anders Søgaard. 2019. X-WikiRE: A large, multilingual resource for relation extraction as machine comprehension. In Proceedings of the 2nd Workshop on Deep Learning Approaches for Low-Resource NLP (DeepLo 2019), pages 265–274, Hong Kong, China. Association for Computational Linguistics. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/eacl-2021-aralikatte-ellipsis-resolution-as-question-answering/