Combining Unsupervised and Supervised Methods for PP Attachment Disambiguation
What it is about
The paper asks what to do when only a small treebank exists for prepositional phrase (PP) attachment. The data are German: 10,000 NEGRA sentences give the supervised examples, and about 5.5 million words of a computer magazine give raw co-occurrence counts. A cascade of count rules, with a noun factor of 5.47 or 5.97 to balance verb and noun counts, reaches 81.67% on the decidable cases (90.13% coverage) of a 4469-item test set. A back-off model trained on the small treebank reaches 73.98%. Ordering the two sources by reliability gives 80.98% on all 4469 items. With the roles swapped, the combination gives 71.95% against 68.29% for back-off alone.
Ideas
- I1.
en/src/parse.rs: add a cascade step for PP arcs (obl,nmodandadvclwith a PP head). Each level is a count table with a threshold, and the level that decides is recorded. Measure coverage and accuracy on the EWT test PP attachments against the current perceptron choice. The tables come from EWT train only. - I2. Verb/noun bias factor: compute the ratio of overall verb-preposition to noun-preposition co-occurrence in the training data (the paper's noun factor) and scale the noun score before comparing. Measure on the two-fold check (
en folds), not on the test set, and run twice to confirm identical output. - I3. Support-verb units as the first level: light-verb frames such as "take into account" or "make use of" with a PP, taken from the MWE list in 19. The lexicon: vocabulary, multiword expressions and register. Check the precision on EWT train before the rule is used. In the paper this level is very reliable (97 cases at 100% on German); the English count is unknown until measured.
For Mova
The chapters are 22. Phrases: noun, adjective, adverb and prepositional phrases and 23-complementation.md (verb or noun choice for a PP complement). The code is en/src/parse.rs, with en/src/explain.rs for the per-level reason. The German figures do not carry over to English; the first English number must come from EWT test, with the two-fold check for decisions. Count tables are built from EWT train, which is CC BY-SA and may be used for training.
Cite
@inproceedings{volk2002combining,
title = {{Combining Unsupervised and Supervised Methods for PP Attachment Disambiguation}},
author = {Volk, Martin},
booktitle = {COLING 2002: The 19th International Conference on Computational Linguistics},
year = {2002},
doi = {10.3115/1072228.1072232},
url = {https://doi.org/10.3115/1072228.1072232}
}References
- E. Brill and P. Resnik. 1994. A rule-based approach to prepositional phrase attachment disambiguation. In Proceedings of COLING, pages 1198–1204, Kyoto. ACL. find
- M. Collins and J. Brooks. 1995. Prepositional phrase attachment through a backed-off model. In Proc. of the Third Workshop on Very Large Corpora. find our card
- D. Hindle and M. Rooth. 1993. Structural ambiguity and lexical relations. Computational Linguistics, 19(1):103–120. correct incorrect accuracy 2400 469 83.65% 1219 381 76.19% 3619 850 80.98% 4469 (of 4469) coverage: 100% threshold 0.020 0.109 number 97 6 269 845 2677 71 81 202 210 11 4469 coverage 2… find
- Adwait Ratnaparkhi. 1998. Statistical models for unsupervised prepositional phrase attachment. In Proceedings of COLING-ACL-98, Montreal. find
- W. Skut, T. Brants, B. Krenn, and H. Uszkoreit. 1998. A linguistically interpreted corpus of German newspaper text. In Proc. of ESSLLI-98 Workshop on Recent Advances in Corpus Annotation, Saarbrücken. find
- J. Stetina and M. Nagao. 1997. Corpus based PP attachment ambiguity resolution with a semantic dictionary. In J. Zhou and K. Church, editors, Proc. of the 5th Workshop on Very Large Corpora, Beijing and Hongkong. find
- Martin Volk. 2001. Exploiting the WWW as a corpus to resolve PP attachment ambiguities. In Proc. of Corpus Linguistics 2001, Lancaster, March. find
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2002-volk-combining-unsupervised-supervised-methods-pp/