Prepositional Phrase Attachment through a Backed-off Model

What it is about

The paper treats the choice between noun and verb attachment of a prepositional phrase as a language-modeling problem. Each case is a quadruple of verb, first noun, preposition and second noun, taken from Wall Street Journal verb phrases in the IBM split (20,801 training, 3,097 test, 4,039 development). The method is the backed-off estimate used in n-gram speech models: it uses the full four-word counts when they are large enough, and backs off to triples and pairs when they are not, with a count threshold at each level. It reaches 84.5% accuracy on the test set. Ignoring any event seen fewer than five times drops it to 81.6%, which the authors read as evidence that low counts matter. A human given only the four head words is at about 88%.

Ideas

• I1. PP-attachment probe in en: extract the quadruples from the en parse of EWT test, score attachments with a backed-off count model, and compare with the gold attachments; this tells how many of the en errors are PP attachment errors and whether a count model would fix them (test only; the WSJ split itself is NC).
• I2. Back-off features in the perceptron: add the backoff stage (which of the verb-preposition, noun-preposition or full counts exist, bucketed) as features in en/src/parse.rs, and measure LAS on EWT test against the current 80.24.
• I3. Count threshold check on the lexicon: test whether removing rare-word features in en/src/ptag.rs hurts ATIS and EWT differently, which is the paper's low-count claim transferred to tagging.

For Mova

The code is en/src/parse.rs (features) and the en parse-eval command for the probe. The chapters are 22. Phrases: noun, adjective, adverb and prepositional phrases (what a PP attaches to) and 23-complementation.md. Measure first: the share of EWT test errors that are PP attachments, using the quadruple extraction; that count decides whether the idea is worth building.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.1007/978-94-017-2390-9_11

@inproceedings{collins1995prepositional,
  title = {{Prepositional Phrase Attachment through a Backed-off Model}},
  author = {Collins, Michael and Brooks, James},
  booktitle = {Third Workshop on Very Large Corpora},
  year = {1995},
  doi = {10.1007/978-94-017-2390-9\_11},
  url = {https://doi.org/10.1007/978-94-017-2390-9_11}
}

References

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

1. W. Gale and K. Church. Poor Estimates of Context are Worse than None. ht Proceedings of the June 1990 DARPA Speech and Natural L~mguage Workshop, ftidden Valley, Pennsylva.nia. find
2. Daniel Karp, Yves Scha,bes, Martin Zaidel and Dania Egedi. A Freely Available Wide Coverage Morphological Analyzer for English. In Proceedings of the 15th, International Conference on Computational Li'nguistics, 1994. D. IIindle and M. Rooth. Structural Ambiguity and Lexical Relations. Computational Linguistics , 19(1):103-120, 1993. find
3. S. Katz. Estima.tion of Probabilities fi:om Sparse Data for the Language Model Component of a. Speech Recogniser. IEEE Transaetion,s on Acoustics, Speech, and Signal Processing, Vol. ASSP-35, No. 3, 1987. find
4. M. Marcus, B. Santorini and M. Marcinkiewicz. Building a Large Annotated Corpus of English: the Penn Treebank. Computational Linguistics, 19(2), 1993. A. Ratnaparkhi, J. Reyna.r and S. Roukos. A Maximum Entropy Model for Prepositional Phrase Attachment. In Proceeding s of the A R P A Workshop on Human Language Technology, Plainsboro, N J, March 1994. [RRR941 find our card

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-1995-collins-prepositional-phrase-attachment-through-backed/