Elamur Research⌕
☰
home / papers / 2010 / Arabic Parsing Using Grammar Transforms

Arabic Parsing Using Grammar Transforms

Summary

The paper asks whether standard treebank-trained parsers for Arabic can recover grammatical functions (subject, object, predicate) well enough to support deep, LFG-style predicate-argument analysis. Plain PCFG training usually throws away function tags and morphology. The authors keep this information by changing the training data instead of the parsers.

They use two grammar transforms on the Penn Arabic Treebank (ATB). First, function tag masking fuses function labels into phrasal categories. Second, case percolation copies nominative or accusative case from the words up into the tree. They train two parsers on the transformed data: Bikel (lexicalised, Collins Model 3) and Berkeley (latent-variable PCFG). After parsing, the fused labels are split apart again. The Arabic LFG annotation algorithm (A3) then turns the trees into f-structures.

Combining function and case information modestly improved both the trees and the dependency-level f-structures for both parsers. Bikel did slightly better. The work matters because it shows that cheap data-side transforms, with no change to the parsers, can give off-the-shelf parsers the functional detail that deep grammar extraction needs.

Key points

  • Data:

    • The ATB has 23,611 MSA newswire sentences.
    • Its tagset has 22 phrasal tags, 20 functional tags and 24 basic POS tags (497 POS tags once morphology is included).
    • The split was 80/10/10, and parsing used gold POS tags.

  • Transforms:

    • Masking grows the phrasal tag set from 20 to 150 symbols.
    • Adding case grows the POS set to 40 tags.
    • -SBJ, -OBJ and -PRD together make up 68% of functional tag tokens in the ATB.

  • Best scores:

    ParserEVALB labelled bracketing f-scoreLFG f-structure dependency f-score
    Bikel73.5888.95
    Berkeley71.5688.5

    Function tags were removed before the EVALB evaluation.

  • Upper bound: on gold ATB trees, the A3 annotation algorithm reaches 95% f-score against a gold f-structure bank.

  • Errors:

    • Accusative and genitive plurals share the same form, and this ambiguity makes the parsers overuse -OBJ.
    • In verbless (nominal) sentences, -SBJ and -PRD get confused.
    • Sentences that open with a copula or a subordinating conjunction lead to -PRD/-OBJ mix-ups.

Interesting ideas

  • I1. Function information can be taught to a parser that knows nothing about it: glue the label onto the category during training, then split it off after parsing. Nothing in the parser itself has to change.
  • I2. Morphological evidence that sits on the words can be copied upward to the phrase nodes, so that the decision about grammatical role sees it directly.
  • I3. The pipeline keeps statistical syntax separate from a deterministic, rule-based mapping into predicate-argument structure. This keeps the semantic step transparent and easy to inspect.
  • I4. The annotation algorithm is scored both on gold trees and on parser output. This separates errors made by the rules from errors inherited from the parser.
  • I5. The error analysis traces systematic tagging mistakes back to underspecified features in the treebank itself, rather than to the model.

For Mova

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

  • I1 (tree): gparse or the perceptron parser could be trained on labels fused with a feature (e.g. voice or case on pronouns), with the suffix stripped back to plain UD afterwards, to test whether the richer labels help attachment.
  • I2 (tokens): English pronoun case (he/him) and agreement features from the tagger could be passed as explicit hints to the parser and to subject/object decisions.
  • I3 (frames): the UD-to-frames step in prag could stay rule-based and explainable, with parser quality treated as a separate, swappable input.
  • I4 (learn): every frames or world stage could be evaluated on gold UD trees and on parsed trees, so that regressions can be blamed on the right component.
  • I5 (learn): when building silver data, features that are genuinely ambiguous should be flagged rather than forced to one value, so the parser is not trained to overpredict one label.

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: Linguistics & languages

Topics Semitic languages Syntax & grammar formalisms

 Institutions

🎓Dublin City UniversityDublin · Ireland48 works · 8,753 citations

Cite

Cite as Lamia Tounsi and Josef van Genabith. 2010. Arabic Parsing Using Grammar Transforms. Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC'10).

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

@inproceedings{tounsi2010arabic,
  title = {{Arabic Parsing Using Grammar Transforms}},
  author = {Lamia Tounsi and Josef {van Genabith}},
  booktitle = {Proceedings of the Seventh International Conference on Language Resources and Evaluation (LREC'10)},
  year = {2010},
  doi = {10.63317/4j59udrzaxf8},
  url = {https://doi.org/10.63317/4j59udrzaxf8}
}

References

20 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Y.Al-Raheb, A.Akrout, J.van Genabith, J.Dichy. 2006. DCU 250 Arabic Dependency Bank: An LFG Gold Standard Resource for the Arabic Penn Treebank The Challenge of Arabic for NLP/MT at the British Computer Society. find
  2. M.Attia. 2008. Handling Arabic Morphological and Syntactic Ambiguity within the LFG Framework with a View to Machine Translation. Ph.D. Thesis. University of Manchester, UK. find
  3. D.Bikel. 2004. Intricacies of Collins’ parsing model Computational Linguistics, 30(4), 2004. our card
  4. J.Bresnan. 2001. Lexical-Functional Syntax Blackwell, Oxford. find
  5. A.Cahill, M.Burke, R.ODonovan, J.van Genabith, A.Way. 2004. Long-Distance Dependency Resolution in Automatically Acquired Wide-Coverage PCFG-Based LFG Approximations. In Proceedings of ACL-04, pages 320-327, Barcelona, Spain. find
  6. A.Cahill, J.van Genabith. 2006. Robust PCFG-Based Generation using Automatically Acquired LFG Approximations. In Proceedings of ACL-06, pages 1033-1040, Sydney, Australia. find
  7. A.Cahill, M.Burke, R.O’Donovan, S.Riezler, J.van Genabith, A.Way. 2008. Wide-Coverage Deep Statistical Parsing using Automatic Dependency Structure Annotation. Computational Linguistics, Vol. 34, No. 1. find
  8. E.Ditters. 2001. A Formal Grammar for the Description of Sentence Structure in Modern Standard Arabic. Workshop on Arabic Processing, ACL/EACL’01. find
  9. N.Habash, G.Ryan, R.Owen, K.Seth, M.Marcus. 2007. Determining Case in Arabic: Learning Complex Linguistic Behaviour Requires Complex Linguistic Features. In Proceedings of EMNLP’07, Prague, Czech Republic. find
  10. N.Habash, F.Reem and R.Ryan. 2009. Syntactic Annotation in the Columbia Arabic Treebank. In Proceedings of MEDAR’09. find
  11. M.Johnson 1998. PCFG models of linguistic tree representations. Computational Linguistics, 24(4), pages 613-632. our card
  12. R.Kaplan , J.Bresnan. 1982. Lexical Functional Grammar, a Formal System for Grammatical Representation. The Mental Representation of Grammatical Relations. find
  13. D.Klein, C. Manning. 2003. Accurate Unlexicalized Parsing. In Proceedings of ACL03, Sapporo, Japan. our card
  14. S.Kulick S, G.Ryan, M.Mitchell. 2006. Parsing the Arabic Treebank: Analysis and improvements. In Proceedings of the Treebanks and Linguistic Theories Conference, Prague. find
  15. M.Maamouri, A.Bies. 2004. Developing an Arabic Treebank: Methods, Guidelines, Procedures, and Tools Proceedings of COLING 2004, Geneva, Switzerland. our card
  16. E.Othman, K.Shaalan, A.Rafea. 2003. A Chart Parser for Analyzing Modern Standard Arabic Sentence In proceedings of the MT Summit IX Workshop on Machine Translation for Semitic Languages: Issues and Approaches, Louisiana, USA. find
  17. S.Petrov and D.Klein. 2007. Improved inference for unlexicalized parsing In Proceedings of HLT-NAACL’07,pages 404-411, New York, USA. our card
  18. S.Sekine and M.J. Collins. 1997. EVALB bracket scoring program web page, http://nlp.cs.nyu.edu/evalb/. ↗
  19. L.Tounsi, M.Attia, J.van Genabith. 2009. Automatic Treebank-Based Acquisition of Arabic LFG Dependency Structures In Proceedings of EACL’2009, Athenes, Greece. find
  20. Z.Zabokrtsky, O.Smrz. 2003. Arabic syntactic trees: from constituency to dependency In Proceedings of EACL’03, pages 183-186, Budapest, Hungary. basic F-score Grammar 1 C-structure Grammar 2 C-structure Grammar 3 C-structure Grammar 4 C-structure Grammar 5 C-structure Grammar 5 F-structure Grammar 6 C-structure 64.10 67.75 69.63 70.86 71.31 n/a 70.99 C-structure F-structure 72.40 n/a Table 1: C-structure and F-structure evaluation for all sentence lengths. basic F-score Grammar 1 C-structure Grammar 2 C-structure Grammar 3 C-structure Grammar 4 C-structure Grammar 5 C-structure Grammar 5 F-structure Grammar 6 C-structure 67.97 71.18 73.02 74.37 75.07 n/a 74.61 C-structure F-structure 75.61 n/a Table 2: C-structure and F-structure evaluation for sentences ≤ 40 tokens in length. masked masked+ percol -SBJ Berkeley 64.10 64.48 67.75 68.43 69.63 70.23 70.86 71.19 71.31 71.78 87.98 88.50 70.99 71.34 Bikel 72.70 73.51 87.09 88.53 masked+ percol -SBJ -OBJ -PRD 64.84 68.83 70.65 71.48 71.82 87.60 71.56 73.58 88.95 masked masked+ percol -SBJ Berkeley 67.97 68.10 71.18 72.05 73.02 73.94 74.37 74.69 75.07 75.37 87.50 88.68 74.61 74.68 Bikel 75.85 76.53 87.12 89.20 masked+ percol -SBJ -OBJ -PRD 69.01 72.72 74.46 75.26 75.43 88.12 74.98 76.53 88.63 our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/tounsi-2010-arabic-parsing-grammar-transforms/