Elamur Research
☰
home / papers / 1998 / PCFG Models of Linguistic Tree Representations

PCFG Models of Linguistic Tree Representations

Written for the Research library (claude-opus-5-5) from the paper text, in our own words: the license "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" does not allow redistributing the text itself. Original — aclanthology.org/J98-4004.pdf

Summary

Johnson asks how the shape of treebank trees affects a parser built from a PCFG estimated on them. Each node label is the only link between a subtree and its surrounding context, so the number and labeling of nodes determine the independence assumptions of the induced grammar. The study combines analysis of toy corpora with experiments on the Penn II treebank Wall Street Journal (WSJ) corpus, preprocessed following Charniak (1996) so that POS tags serve as terminals. It introduces a transformation/detransformation method. Training trees are rewritten (flattened, converted to Chomsky adjunction, or relabeled with each node's parent category). A PCFG is then estimated and used for CKY parsing, and its outputs are mapped back to the original scheme for scoring. Parent annotation raised average precision and recall by about 8%, closing roughly half of the gap to the strongest parsers of the time, which use lexical information. The linguistically favoured Chomsky adjunction did worse than representations with weaker independence assumptions. The results show that choosing a tree representation is a real modeling decision governed by a bias/variance trade-off.

Key points

  • Data: WSJ portion of the Penn II treebank, with sections 2–21 used to estimate the grammar. Preprocessing removed words, function tags and empty nodes, added ROOT nodes, and tagged auxiliaries as AUX.
  • In the baseline treebank PCFG, 1,327 of its 14,962 productions (just under 9%) can never appear in a maximum-likelihood parse, because cheaper combinations of shorter rules outscore them. For example, NP → NP PP (0.112) beats NP → NP PP PP (0.006).
  • As a result, NP-attached PPs in the parser's output always come out as Chomsky adjunctions, even though the treebank itself uses a two-level scheme.
  • The parent annotation transform gives about an 8% gain in average precision and recall. Selective application shows that most of the gain comes from NP and S nodes. Separating root from non-root S alone adds about 3%, so the improvement may have little to do with PP attachment.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Topics: Treebanks

Cited by

136 papers of our library cite it (of the 140673 whose bibliographies we read); 9 of them have a page here, 127 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@article{johnson1998pcfg,
  title = {{PCFG Models of Linguistic Tree Representations}},
  author = {Johnson, Mark},
  journal = {Computational Linguistics},
  year = {1998},
  url = {https://aclanthology.org/J98-4004/}
}

References

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

  1. Aho, Alfred V. and Jeffery D. Ullman. 1972. find
  2. The Theory of Parsing, Translation and Compiling; Volume 1: Parsing. Prentice-Hall, Englewood Cliffs, NJ. find
  3. Bies, Ann, Mark Ferguson, Karen Katz, and Robert MacIntyre. 1995. Bracketing Guidelines for Treebank II style Penn Treebank Project. Linguistic Data Consortium. find
  4. Charniak, Eugene. 1993. Statistical Language Learning. MIT Press, Cambridge, MA. find
  5. Charniak, Eugene. 1996. Tree-bank grammars. In Proceedings of the Thirteenth National Conference on Artificial Intelligence, pages 1031-1036, Menlo Park. AAAI Press/MIT Press. find
  6. Charniak, Eugene. 1997. Statistical parsing with a context-free grammar and word statistics. In Proceedings of the Fourteenth National Conference on Artificial Intelligence, Menlo Park. AAAI Press/MIT Press. find
  7. Charniak, Eugene and Glenn Carroll. 1994. Context-sensitive statistics for improved grammatical language models. In Proceedings of AAAI "94, pages 728-733. find
  8. Chi, Zhiyi and Stuart Geman. 1998. Estimation of probabilistic context-free grammars. Computational Linguistics, 24(2): 299-305. find
  9. Cohen, Paul R. 1995. Empirical Methods for Artificial Intelligence. MIT Press, Cambridge, MA. find
  10. Collins, M. J. 1996. A new statistical parser based on bigram lexical dependencies. In The Proceedings of the 34th Annual Meeting, pages 184-191, San Francisco. Association for Computational Linguistics, Morgan Kaufmann. find
  11. Culy, Christopher. 1985. The complexity of the vocabulary of Bambara. Linguistics and Philosophy, 8(3):345-352. find
  12. Gazdar, Gerald, Ewan Klein, Geoffrey Pullum, and Ivan Sag. 1985. Generalized Phrase Structure Grammar. Basil Blackwell, Oxford. find
  13. Geman, Stuart, Elie Beinenstock, and Rend Doursat. 1992. Neural networks and the bias/variance dilemma. Neural Computation, 4:1-58. find
  14. Hindle, Donald and Mats Rooth. 1993. Structural ambiguity and lexical relations. Computational Linguistics, 19(1):103-120. find
  15. Krotov, Alexander, Robert Gaizauskas, Mark Hepple, and Yorik Wilks. 1997. Compacting the Penn Treebank grammar. Technical report, Department of Computer Science, Sheffield University. find
  16. Magerman, D. M. and M. P. Marcus. 1991. Pearl: A probabilistic chart parser. In Proceedings of the European ACL Conference. find
  17. Marcus, Mitchell P., Beatrice Santorini, and Mary Ann Marcinkiewicz. 1993. Building a large annotated corpus of English: The Penn Treebank. Computational Linguistics, 19(2):313-330. find our card
  18. Shieber, Stuart M. 1985. Evidence against the context-freeness of natural language. Linguistics and Philosophy, 8(3):333-344. find