Elamur Research⌕
☰
home / papers / 1997 / Probabilistic Parsing of Unrestricted English Text, With a Highly-Detailed Grammar

Probabilistic Parsing of Unrestricted English Text, With a Highly-Detailed Grammar

Paper · 1997 · Ezra Black, Stephen Eubank, Hideki Kashioka et al.

Summary

The paper addresses broad-coverage statistical parsing of open-domain English, aiming for much richer output than the shallow phrase labels common in treebanks of the time. The authors build a parser that uses probabilistic decision trees. It is trained under supervision on a 676,000-word, highly varied portion of the ATR/Lancaster Treebank of General English. In that treebank, each node is labelled with the ATR English Grammar rule that produced it, and each word has a combined syntactic-semantic tag. Tying nodes to grammar rules lets the models query many percolated feature values, head information and rule constraints. The models can also ask about raw words, word strings and properties of the whole sentence. All of these queries are written in a dedicated question language, and the answers are turned into binary strings for the trees. The parser needs no closed vocabulary or fixed domain, and it outputs detailed semantic as well as syntactic analyses. The authors also describe a statistical method for converting less-detailed treebanks into this detailed scheme, so that much larger training sets become available.

Key points

  • Training and testing use 676,000 words from the one-million-word ATR/Lancaster Treebank of General English, which includes varied genres such as brochures and take-out food fliers.
  • The ATR English Grammar has 1155 rules, which serve as non-terminal labels, and a tagset of 2843 combined syntactic-and-semantic tags. Each node carries 66 features with about 12 values per feature on average.
  • Prediction follows the decision-tree approach of Jelinek et al. (1994) and Magerman (1995). A parse is assembled through a sequence of partial-parse states.
  • Every question answer (Boolean, categorical, integer or set-valued) is mapped to a bit string. The grammarian clustered tags and rule labels by hand, while a vocabulary of about 60,000 words was clustered automatically with the Brown et al. (1992) algorithm, implemented by Ushioda.
  • A statistical treebank-conversion procedure maps coarser treebanks into the detailed ATR format to increase training data.

Interesting ideas

  • I1. Global sentence cues work well as predictors: overall length, the first and last word, and where the single comma falls (for example in the first quadrant after "Since") point to structures like a premodifier followed by the main clause, and very short fragments usually lack finite verbs.
  • I2. Context is defined through a recursive navigation language over the current parse state rather than a fixed window, so moving from a bigram to a trigram model only means adding new questions.
  • I3. Semantic tag decisions are supported by contexts strongly associated with a category, for example a first name, then a quoted token, then a last name, which signals a NickName.
  • I4. Coarse treebanks can be statistically converted into a richer annotation scheme, which multiplies the training data available for the detailed target format.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • The grammar includes features built specifically to help prediction, such as np_modification, which records at the top of a noun phrase how much further modification it can likely accept; this gives attachment decisions a compact signal that would be hard to induce from raw trees. — (tree): The perceptron parser and gparse could compute a similar per-head "saturation" feature (already postmodified, closed by a comparative phrase, determinerless) to improve PP and clause attachment.
  • I1 (tokens): The tagger could use sentence-level features such as length, presence of a finite verb and comma position to separate preterites from adjectival participles in headlines, titles and speaker labels.
  • I2 (learn): Writing parser and repair-rule features as declarative tree-navigation queries would make each feature inspectable and easy to extend, which fits Mova's requirement that every step be explainable.
  • I3 (world): Entity building could use such hand-listed surface templates to assign fine-grained types (nickname, title, organisation) before linking to the ontology, each with a traceable reason.
  • I4 (tree): The same idea applies to UD dialect conversion, where Mova could learn mappings from other treebanks into its own enriched UD variant instead of re-annotating them by hand.

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, topics and concepts

📚 Shelves 2The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 works

🧭 Topics 1Research subjects and languages the work is about.English · 2,379 works

Cite

Cite as Ezra Black, Stephen Eubank, Hideki Kashioka and David Magerman. 1997. Probabilistic Parsing of Unrestricted English Text, With a Highly-Detailed Grammar. Fifth Workshop on Very Large Corpora.

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

@inproceedings{black1997probabilistic,
  title = {{Probabilistic Parsing of Unrestricted English Text, With a Highly-Detailed Grammar}},
  author = {Ezra Black, Stephen Eubank, Hideki Kashioka and David Magerman},
  booktitle = {Fifth Workshop on Very Large Corpora},
  year = {1997},
  url = {https://aclanthology.org/W97-0105/}
}

References

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

  1. L. Bahl, P. Brown, P. deSouza, and R. Mercer. 1989. A tree-based statistical language model for natural language speech recognition. IEEE Transactions on Acoustics, Speech, and ,~ignal Processing, 36.7:1001-1008. find
  2. L. Baum. 1972. An inequ~llty and associated maximization ter]~ulque in statistical estimation of probabilistic functions of markov processes. Inequalities: 3:1-8. find
  3. E. Black, S. Eubank, H. Kashioka, R. Garside, G. Leech, and D. Magerman. 1996. Beyond skeleton parsing: producing a comprehensive large — scale general-English treebank with full grammatical analysis. In Proceedings of the 16th Annual Conference on Computational Lan#uistics, pages 107-112, Copenhagen. find
  4. E. Black. 1994. A new approach to evaluating broad-coverage parser/grammars of English. In Proceedings of the International Conference on New Methods in Language Processing. Manchester, UK. find
  5. E. Black, R. Garside, and G. Leech, Editors. 1993. Statistically-Driven Computer Grammars Of gnglish: The IBM/Lancaster Approach. Rodopi Editions. Amsterdam. find
  6. E. Black, F. Jelinek, J. Lafferty, D. Magerman, R. Mercer, S. Roukos. 1993. Towards History-Based Grammars: Using Richer Models For Probabilistic Parsing. In Proceedings of the 31st Annual Meeting of the Association for Computational Linguistics. Columbus, Ohio. Also in Proceedings of the DARPA Speech and Natural Language Workshop, 1992. ↗
  7. E. Brill. 1993. Automatic grammar induction and parsing free text: a transformation-based approach. In Proceedings of the 81st Annual Meeting of the Association .for Computational Langu/st/es, pages 259–265, Columbus, Ohio. ↗
  8. L. Breimau, J. Friedman, R. Olshen, and C. Stone. 1984. 6~assification and Regression Trees. Wadsworth & Brooks/Cole, Monterey, CA. find
  9. P. Brown, V. Della Pietra,. P. de 8ouza, J. Lai, R. Mercer. 1992. Class-based n-gram models of natural language. Computational Linguistics, 18.4:467–479. find
  10. M. Collins. 1996. A new statistical parser based on bigram lexieal dependencies. In Proceedings of the 34th Annual Meetin 9 of the Association .for Computational Languistics, Santa Cruz. find
  11. J. Eisner. 1996. Three new probabilistic models for dependency parsing: an exploration. In Proceedings of the 16th Annual Conference on Computational Languistics, pages 340–345, Copenhagen. ↗
  12. E. Eyes and G. Leech. 1993. Syntactic Annotation: Linguistic Aspects of Grammatical Tagging and Skeleton Parsing. Chapter 3 of Black et. al. 1993. tL Garside, G. Leech, G. Sampson, Editors. 1987. The Computational Analysis of English. London, Longman. find
  13. I~ Garside and A. McEnery. 1993. Treeba~kingz. The Compilation of a Corpus of Skeleton-Parsed Sentences. Chapter 2 of Black et. al. 1993. find
  14. D. Grinberg, J. Lafferty, and D. Sleator. 1995. A robust parsing algorithm for link gr'Amm~rs. In Proceedings of the .Wourth International Workshop on Parsing Technologies, Prague. find
  15. P. Harrison, S. Abney, E. Black, D. Flickenger, C. Gdaniee, R. Grishman, D. Hindle, R. Ingria, M. Marcus, B. Santorini, T. Strzalkowski. 1991. Evaluating Syntax Performance Of Parser/Grammars Of English. In Proceedings of the Workshop On Evaluating Natural Language Processing Systems, Association For Computational Linguistics. find
  16. J. Hughes, C. Souter, and E. AtwelL 1995. Automatic extraction of tagset mappings from parallel-annotated corpora. In Proceedings of SIGDAT Workshop, Dublin. find
  17. F. Jelinek, J. Lafferty, D. Magerman, R. Mercer, A. l~tnaparldai, S. Roukos. 1994. Decision tree parsing using a hidden derivation model In Proceedings of the ARPA Workshop on Human Language Technology, pages 260-265, Plainsboro, New Jersey. Advanced Research Projects Agsney. find
  18. F. Jelinek. 1969. A fast sequential decoding algorithm using a stack. IBM Journal of Research and Development, 13:675-685. find
  19. T. Kasami. 1965. An ettleient recognition and syntax algorithm for context-free languages. Scientific Report AFCRL~5–758. Air Force Cambridge Laboratory. Bedford, Massachusetts. 1~ Long. 1961. The Sentence and Its Parts. University of Chicago Press. Chicago. find
  20. D. Magerman. 1995. Statistical decision-tree models for parsing. In 33rd Annual Meeting of the Association .for Computational Linguistics, pages 276–283, Cambridge, Massachusetts. Association for Computational Linguistics.
  21. M. P. Marcus, B. Santorini, and M. A. Mareinkiewiez. 1993. Building a Large Annotated Corpus of English: The Penn Treebank. Computational Linguistics, 19.2:313-330. ↗
  22. J. R. Quinlan. 1986. Induction of decision trees. Machine Learning, 1:80-106. ↗
  23. Y. Sehabes. 1992. Stochastic lexiealized tree-adjoining gr~trnmaxS. In Proceedings of the 15th International Conference on Computational Linguistics, Nantes. find
  24. S. Seklne and R. Grishrnan. 1995. A corpus-based probabilistie grammar with only two nonterminals. In Proceedings, International Workshop on Parsing Technologies, 1995. find
  25. R. A. Sharman, F. Jelinek, and tL Mercer. 1990. Generating a grammar for statistical training. In Proceedings, DARPA Speech and Natural Language Workshop, Hidden Valley, Pe~n~lvania. find
  26. A. Ushioda. 1996. I-I.ierarehieal clustering of words. In Proceedings of the 16th Annual Conference on Computational Languistics, pages 1159-1162, Copenhagen. find
  27. A. Ushioda. 1996. Hieraxehieal clustering of words and application to NLP tasks. In Proceedings of the Fourth Workshop on Very Large Corpora, pages 28–41, Copenlmgen. find
  28. J. Wang, J. Chang, and K. Su. 1994. An a~xtomatie treebank conversion algortihm for corpus sharing. In Proceedings of the 3~nd Annual Meeting of the Association for Computational Linguistics, pages 248–254, Las Cruces, New Mexico. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/black-1997-probabilistic-parsing-unrestricted-english-text/