Elamur Research⌕
☰
home / papers / 2011 / ArbTE: Arabic Textual Entailment

ArbTE: Arabic Textual Entailment

Paper · 2011 · Maytham Alabbas

Summary

This workshop paper outlines ArbTE, a planned system for recognising textual entailment in Modern Standard Arabic. The author reports finding no prior TE work for Arabic. Arabic is very ambiguous: short vowels go unwritten, clitics attach freely, word order is flexible, subjects can be dropped, the copula is omitted, and construct phrases carry little marking. The author trained taggers (AMIRA 2.0, MADA 3.1, an in-house MXL) and dependency parsers (MaltParser, MST) on the Penn Arabic Treebank. To do this, he first converted its phrase-structure trees to dependencies with a semi-automatically built head percolation table. Parsing accuracy stayed modest: about 80% unlabelled and about 70% labelled. Voting among parsers and taggers raised reliability. The proposed pipeline has four parts: forward inference rules taken only from trusted subtrees, a tree edit distance in which subtree operations cost less than their parts, entailment pairs mined from news headlines, and later, polarity marking. The work matters because it shows how to build entailment on top of parsers that cannot be fully trusted.

Key points

  • PATB (over 5000 phrase-structure trees) was converted to dependencies using a head percolation table in the style of Xia & Palmer and Collins. Related tags were grouped, and their possible heads were ordered by hand, with 'CONJ' given high priority.
  • Tagger accuracy with a coarse tagset was AMIRA 89.59, MXL 95.17 and MADA 94.05. On tokens where all three agree, precision reaches 0.9947, with recall 0.83 and F-score 0.9049.
  • Two voting schemes combined MST, MALT1 (arc-eager) and MALT2 (stack-eager):
    • TECHNIQUE1 uses majority voting; when the parsers disagree, the head comes from MST and the label from MALT1.
    • TECHNIQUE2 keeps the output where MALT1 and MALT2 agree and otherwise falls back to MST.
    • Both beat every single parser, on retagged input and on gold-standard tags alike.
  • The author reports that adding more training data had stopped improving accuracy, so better parsing alone is not expected to fix the problem.
  • Text:hypothesis pairs are harvested by sending news headlines to the Google API and taking the first sentence of the top N pages. Two filters apply: word overlap must stay below a threshold, and headlines must exceed a minimum length.

Interesting ideas

  • I1. When independent taggers agree, the result is far more precise than any single tagger, so their agreement can mark which tokens are trustworthy.
  • I2. Inference rules are applied only to subtrees that the parser has proven to get right consistently; less reliable regions are left to later matching.
  • I3. In the modified tree edit distance, removing or swapping a whole modifier subtree costs less than editing each of its nodes separately, so edits respect linguistic units.
  • I4. Entailment test pairs are mined by pairing headlines with the opening sentences of matching articles, with overlap filters to block near-copies; this avoids pairs biased by the developers' own intuitions.
  • I5. Words like "doubt" flip the polarity of their complement, so an inference that is valid under "believe" becomes invalid under "doubt". Matching must therefore track polarity.

For Mova

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

  • I1 (learn): Use unanimous agreement among witness taggers and parsers as the filter for silver data, and track the precision/recall trade-off explicitly as this paper does.
  • I2 (frames): Build frames only from tree fragments with high agreement or confidence, and mark frames built from shaky attachments as tentative so the explanation can flag them.
  • I3 (answer): When checking whether an answer follows from a story, compare trees with subtree-level costs, so that dropping a whole adjunct counts as one cheap, explainable step.
  • I4 (learn): Mine headline/lead-sentence pairs from English news as unbiased entailment checks, and use high-overlap pairs as negative controls.
  • I5 (world): Mark polarity on embedded clauses in the world model, so that upward inferences such as woman → human being are allowed only in positive contexts.

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

 Institutions

🎓University of ManchesterManchester · United Kingdom87 works · 44,938 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 works

🧭 Topics 2Research subjects and languages the work is about.Reasoning · 1,426 worksSemitic languages · 2,908 works

Cite

Cite as Maytham Alabbas. 2011. ArbTE: Arabic Textual Entailment. Proceedings of the Second Student Research Workshop associated with RANLP 2011.

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

@inproceedings{alabbas2011arbte,
  title = {{ArbTE: Arabic Textual Entailment}},
  author = {Maytham Alabbas},
  booktitle = {Proceedings of the Second Student Research Workshop associated with RANLP 2011},
  year = {2011},
  url = {https://aclanthology.org/R11-2007/}
}

References

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

  1. Alabbas, M. and Ramsay, A. (2011). Evaluation of dependency parsers for long arabic sentences. In Proceeding of International Conference on Semantic Technology and Information Retrieval (STAIR’11), pages 243–248. IEEE. find
  2. Baptista, M. (1995). On the nature of pro-drop in Capeverdean Creole. Technical report, Harvard Working Papers in Linguistics, 5: 3-17. find
  3. Black, W., Elkateb, S., Rodriguez, H., and Alkhalifa, M. (2006). Introducing the Arabic WordNet project. In Proceedings of the Third International WordNet Conference (GWC-06). find
  4. Blackburn, P., Bos, J., Kohlhase, M., and De Nivelle, H. (2001). Inference and computational semantics. Studies in linguistics and philosophy, pages 11–28. ↗
  5. Collins, M. (1997). Three generative, lexicalised models for statistical parsing. In Proceedings of the 34th Annual Meeting of the Association for Computational Linguistics, pages 16–23, Madrid. ↗
  6. Results Entails Entails Entails Dagan, I., Magnini, B., and Glickman, O. (2005). The PAS-CAL recognising textual entailment challenge. In Proceedings of Pascal Challenge Workshop on Recognizing Textual Entailment. find
  7. Daimi, K. (2001). Identifying Syntactic Ambiguities in Single-Parse Arabic Sentence. Computers and the Humanities, 35(3):333–349. ↗
  8. Diab, M. (2009). Second generation tools (amira 2.0): Fast and robust tokenization, pos tagging, and base phrase chunking. In Proceedings of the Second International Conference on Arabic Language Resources and Tools, pages 285–288. find
  9. Habash, N. (2010). Introduction to Arabic Natural Language Processing. Synthesis Lectures on Human Language Technologies. Morgan & Claypool Publishers. find
  10. Habash, N., Rambow, O., and R., R. (2009). Mada+tokan: A toolkit for arabic tokenization, diacritization, morphological disambiguation, pos tagging, stemming and lemmatization. In Choukri, K. and Maegaard, B., editors, Proceedings of the Second International Conference on Arabic Language Resources and Tools, Cairo, Egypt. The MEDAR Consortium. find
  11. Heilman, M. and Smith, N. (2010). Tree edit models for recognizing textual entailments, paraphrases, and answers to questions. In Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, pages 1011– 1019. Association for Computational Linguistics.
  12. Hobbs, J. R. (2005). The handbook of pragmatics, chapter Abduction in Natural Language Understading, pages 724– 740. Blackwell Publishing. find
  13. Hobbs, J. R., Stickel, M. E., Appelt, D. E., and Martin, P. (1993). Interpretation as abduction. Artificial Intelligence, 63:69–142. find
  14. Hutchins, W. J. and Somers, H. L. (1992). An introduction to machine translation. Academic Press. find
  15. Kouylekov, M. (2006). Recognizing textual entailment with tree edit distance: Application to question answering and information extraction. PhD thesis. find
  16. Maamouri, M. and Bies, A. (2004). Developing an Arabic treebank: Methods, guidelines, procedures, and tools. In Proceedings of COLING, pages 2–9.
  17. McDonald, R. and Pereira, F. (2006). Online learning of approximate dependency parsing algorithms. In Proceedings of EACL, volume 6, pages 81–88.
  18. Nivre, J., Hall, J., Nilsson, J., Chanev, A., Eryigit, G., Kübler, S., Marinov, S., and Marsi, E. (2007). MaltParser: A language-independent system for data-driven dependency parsing. Natural Language Engineering, 13(02):95–135. ↗
  19. Xia, F. and Palmer, M. (2001). Converting dependency structures to phrase structures. In 1st Human Language Technology Conference (HLT-2001), pages 1–5, San Diego. find
  20. Zhang, K. and Shasha, D. (1989). Simple fast algorithms for the editing distance between trees and related problems. SIAM J. Comput., 18(6):1245–1262. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/alabbas-2011-arbte-arabic-textual-entailment/