Generation of VP Ellipsis: A Corpus-Based Approach
What it is about
The paper asks when a generator should elide a verb phrase and leave an auxiliary in its place, as in "in 1987, only 9% did". The corpus is taken from Sections 5 and 6 of the Wall Street Journal part of the Penn Treebank: 126 examples in all, 111 negative cases where two identical verb phrases were not elided, and 15 positive cases. The authors code distance in words and sentences, the syntactic relation, voice, adjuncts, quotes and discourse structure. Distance and syntactic relation correlate with ellipsis, the length of the antecedent does not. A rule learner (Ripper) gives the lowest error when it runs after the realizer on surface features: a mean error rate of 7.5% in ten-fold cross-validation, against 11.9% for never eliding.
Ideas
- I1.
en/src/lin.rs: after word order is fixed, add a rule that marks a verb pair for ellipsis only under the conditions the paper learns: same sentence, comparative construction, or at most 14 words apart with an antecedent of at least 3 words. Measure precision and recall on identical verb pairs from EWT, on the two-fold check (en folds). - I2. Build a VPE set from EWT, which is CC BY-SA and may be used for training. Screen pairs of identical verbs with fewer than 10 intervening verbs, as the paper does, and label them by hand or with
annot, checking a sample by hand. Report how many pairs are elided. - I3. Stage test: run the elision decision at three points (on semantic features only, after parsing, after linearization) and compare error rates in
en lin-eval. The paper finds the error falls from 11.7% to 9.2% to 7.5% as the stage moves later. - I4. Use the 126 Penn Treebank examples as a test set only, never as training data for Mova.
For Mova
The chapter is 29. Ellipsis, word order and information structure, and the code is en/src/lin.rs with the document context in en/src/ctx.rs for the antecedent distance. Measure first how often identical verb pairs within 10 verbs are elided in EWT. The paper's data is WSJ text from the LDC-licensed Penn Treebank, so it is test only in any measurement. Its 126 examples are a small set, and its error rate is cross-validated, not held out.
Cite
@inproceedings{hardt2001generation,
title = {{Generation of VP Ellipsis: A Corpus-Based Approach}},
author = {Hardt, Daniel and Rambow, Owen},
booktitle = {Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics},
year = {2001},
doi = {10.3115/1073012.1073050},
url = {https://doi.org/10.3115/1073012.1073050}
}References
- Jean Carletta. 1996. Assessing agreement on classification tasks: The kappa statistic. Computational Linguistics, 22(2):249–254. find
- William Cohen. 1996. Learning trees and rules with set-valued features. In Fourteenth Conference of the American Association of Artificial Intelligence. AAAI. find
- Mary Dalrymple, Stuart Shieber, and Fernando Pereira. 1991. Ellipsis and higher-order unification. Linguistics and Philosophy, 14(4), August. find
- Robert Fiengo and Robert May. 1994. Indices and Identity. MIT Press, Cambridge, MA. find
- Daniel Hardt. 1997. An empirical approach to vp ellipsis. Computational Linguistics, 23(4):525–541. find our card
- Daniel Hardt. 1999. Dynamic interpretation of verb phrase ellipsis. Linguistics and Philosophy, 22(2):187–221. find
- Andrew Kehler. 1993. The effect of establishing coherence in ellipsis and anaphora resolution. In Proceedings, 28th Annual Meeting of the ACL, Columbus, OH. find
- Owen Rambow and Tanya Korelsky. 1992. Applied text generation. In Third Conference on Applied Natural Language Processing, pages 40–47, Trento, Italy. find
- Ivan A. Sag. 1976. Deletion and Logical Form. Ph.D. thesis, Massachusetts Institute of Technology. (Published 1980 by Garland Publishing, New York). find
- Marilyn A. Walker. 1996. Limited attention and discourse structure. Computational Linguistics, 22-2:255–264. find
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2001-hardt-generation-vp-ellipsis-corpus-based/