Elamur Research⌕
☰
home / papers / 2023 / Unpacking Ambiguous Structure: A Dataset for Ambiguous Implicit Discourse Relations for English and Egyptian Arabic

Unpacking Ambiguous Structure: A Dataset for Ambiguous Implicit Discourse Relations for English and Egyptian Arabic

Summary

The paper builds a small controlled resource for studying implicit discourse relations that are truly ambiguous. These are two-clause items with no connective that can be read either as causal ("so") or as concessive ("but"), and that stay unresolved without outside context. The authors propose construction principles, chiefly that the first argument should be evaluatively neutral so that context can tilt it positive or negative. They then write paired items in which the same target sentence follows two different preceding context sentences, each pushing toward one reading. The English items were translated into Egyptian Arabic. Humans validated the data by picking the fitting connective from a list, with iterative cleanup between rounds. Human readers resolved the items easily. Masked language models (BERT, ELECTRA, CAMeLBERT, AraELECTRA) barely changed their predictions between the two contexts, and in Arabic they often proposed words that are not connectives at all. The resource is meant to isolate semantics from prosody, so that future experiments can test whether intonation alone disambiguates.

Key points

  • Each example is a minimal pair. Arg1 and Arg2 are identical, and only the preceding context differs, licensing either "so" (causal) or "but" (concessive). An example is "The car is very cheap / very expensive. It's 100,000, __ I'll buy it."
  • The construction principles are:
    • The relation must be implicit, with no structural connective.
    • Arg2 must not contain anaphoric "this/that" pointing back at Arg1.
    • Arg2 should elaborate on Arg1.
    • Arg1 should be neutral and shaped by context.
    • The relation should be inferable only from context.
  • The authors tried contronyms (e.g., "crazy prices") as Arg1 but rejected them. Readers could resolve them from background assumptions alone.
  • The LM probing put a [MASK] between Arg1 and Arg2 and inspected the top-3 predictions. English used bert-base-uncased and electra-base. Egyptian Arabic used CAMeLBERT-Mix and AraELECTRA-base-generator.
  • The Arabic top-3 outputs contained 25 (BERT) and 33 (ELECTRA) distinct words, many of them non-connectives. "And"/proclitic "wa" showed up often, which the authors relate to relations conveyed through "and"-conjunction.

Interesting ideas

  • I1. Hold the target sentence fixed and vary only the preceding context, so any change in a model's output can be attributed purely to context; this is a clean causal test of context sensitivity. — For Mova (learn): This could become a negative-control suite in which a frames/world step that returns the same relation under both contexts is flagged as ignoring discourse.
  • I2. The ambiguity comes from an evaluatively neutral Arg1 whose polarity is assigned by a context-supplied preference, so the relation follows from preference plus fact. — For Mova (world): The world model could store stated preferences as states and derive "expected vs. unexpected" for the next event, choosing cause or concession explainably.
  • I3. The authors drop candidate items such as contronyms when readers resolve them from background knowledge alone, keeping only items that context truly decides. — For Mova (answer): Comprehension test items could be screened the same way, discarding questions answerable without reading the story, so scores reflect real text understanding.
  • I4. Fill-the-connective is used as a probe of discourse understanding, scoring whether the predicted word is even a connective before judging whether it is the right one. — For Mova (back): Generation could pick between "so" and "but" from the meaning graph's expectation state and be checked on such minimal pairs.
  • I5. Implicit relations can surface as bare "and" or plain juxtaposition, so "and" between clauses should not be assumed to mean pure addition. — For Mova (frames): Where an UD cc "and" links two clauses, prag could keep the relation underspecified and let world-level reasoning resolve it to result or concession.

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: Discourse analysis

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2023.codi-1.16

@inproceedings{ruby2023unpacking,
  title = {{Unpacking Ambiguous Structure: A Dataset for Ambiguous Implicit Discourse Relations for English and Egyptian Arabic}},
  author = {Ruby, Ahmed and Stymne, Sara and Hardmeier, Christian},
  booktitle = {Proceedings of the 4th Workshop on Computational Approaches to Discourse (CODI 2023)},
  year = {2023},
  doi = {10.18653/v1/2023.codi-1.16},
  url = {https://doi.org/10.18653/v1/2023.codi-1.16}
}

References

27 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Ekaterina Jasinskaja. 2009. Pragmatics and Prosody of Implicit Discourse Relations: The Case of Restatement. Ph.D. thesis, Universität Tübingen, Germany. find
  2. Ines Adornetti. 2015. The phylogenetic foundations of discourse coherence: A pragmatic account of the evolution of language. Biosemiotics, 8:421–441. find
  3. Julia Lavid and Eduard Hovy. 2010. Towards a science of corpus annotation: a new methodological challenge for corpus linguistics. International Journal of Translation, 22:13–36. find
  4. Amal Alsaif. 2012. Human and automatic annotation of discourse relations for Arabic. Ph.D. thesis, University of Leeds, UK. find
  5. Song Lichao. 2010. The role of context in discourse analysis. Journal of Language Teaching and Research, 1:876–879. find
  6. Wissam Antoun, Fady Baly, and Hazem Hajj. 2021. AraELECTRA: Pre-training text discriminators for Arabic language understanding. In Proceedings of the Sixth Arabic Natural Language Processing Workshop, pages 191–195, Kyiv, Ukraine (Virtual). Association for Computational Linguistics. find
  7. Wanqiu Long, Bonnie Webber, and Deyi Xiong. 2020. TED-CDB: A large-scale Chinese discourse relation dataset on TED talks. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2793–2803, Online. Association for Computational Linguistics. find our card
  8. Fatemeh Torabi Asr and Vera Demberg. 2020. Interpretation of discourse connectives is probabilistic: Evidence from the study of but and although. Discourse Processes, 57(4):376–399. find
  9. Eleni Miltsakaki, Nikhil Dinesh, Rashmi Prasad, Aravind Joshi, and Bonnie Webber. 2005. Experiments on sense annotations and sense disambiguation of discourse connectives. In Proceedings of the Fourth Workshop on Treebanks and Linguistic Theories (TLT 2005), Barcelona, Spain. find
  10. El-Said Badawi. 1973. Mustawayat al- arabiyya al-mu asira fi misr (Levels of Contemporary Arabic in Egypt). Dar almaarif-cairo, Cairo, Egypt. find
  11. Elena Cabrio, Sara Tonelli, and Serena Villata. 2013. From discourse analysis to argumentation schemes and back: Relations and differences. In Computational Logic in Multi-Agent Systems, pages 1–17, Berlin, Heidelberg. Springer Berlin Heidelberg. find
  12. Robyn Carston. 1993. Conjunction, explanation and relevance. Lingua, 90(1):27–48. find
  13. Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. 2020. ELECTRA: Pre-training text encoders as discriminators rather than generators. In ICLR. find
  14. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171–4186, Minneapolis, Minnesota. Association for Computational Linguistics. ↗
  15. Nizar Y. Habash. 2010. Introduction to Arabic natural language processing, 1 edition, volume 3 of Synthesis Lectures on Human Language Technologies. Morgan and Claypool Publishers. find
  16. Go Inoue, Bashar Alhafni, Nurpeiis Baimukan, Houda Bouamor, and Nizar Habash. 2021. The interplay of variant, size, and task type in Arabic pre-trained Reinhard Muskens. 2000. Underspecified semantics. In Klaus von Heusinger and Urs Egli, editors, Reference and Anaphoric Relations, pages 311–338. Springer Netherlands, Dordrecht. find
  17. Ethan Nowak and Eliot Michaelson. 2020. Discourse and method. Linguistics and Philosophy, 43:119– 138. find
  18. Umangi Oza, Rashmi Prasad, Sudheer Kolachina, Dipti Misra Sharma, and Aravind Joshi. 2009. The Hindi discourse relation bank. In Proceedings of the Third Linguistic Annotation Workshop (LAW III), pages 158–161, Suntec, Singapore. Association for Computational Linguistics. find our card
  19. Emily Pitler, Mridhula Raghupathy, Hena Mehta, Ani Nenkova, Alan Lee, and Aravind Joshi. 2008. Easily identifiable discourse relations. In Coling 2008: Companion volume: Posters, pages 87–90, Manchester, UK. Coling 2008 Organizing Committee. find our card
  20. Rashmi Prasad, Nikhil Dinesh, Alan Lee, Eleni Milt-sakaki, Livio Robaldo, Aravind Joshi, and Bonnie Webber. 2008. The Penn Discourse TreeBank 2.0. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), Marrakech, Morocco. European Language Resources Association (ELRA). ↗ our card
  21. Ines Rehbein, Merel Scholman, and Vera Demberg. 2016. Annotating discourse relations in spoken language: A comparison of the PDTB and CCR frameworks. In Proceedings of the Tenth International Conference on Language Resources and Evaluation (LREC’16), pages 1039–1046, Portorož, Slovenia. European Language Resources Association (ELRA). find our card
  22. Hannah Rohde, Alexander Johnson, Nathan Schneider, and Bonnie Webber. 2018. Discourse coherence: Concurrent explicit and implicit relations. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2257–2267, Melbourne, Australia. Association for Computational Linguistics. find our card
  23. Joseph Tyler. 2014. Prosody and the interpretation of hierarchically ambiguous discourse. Discourse Processes, 51(8):656–687. find
  24. Arie Verhagen. 2000. Concession implies causality, though in some other space. In Elizabeth Couper-Kuhlen and Bernd Kortmann, editors, Cause - Condition - Concession - Contrast, pages 361–380. De Gruyter Mouton, Berlin, Boston. find
  25. Florian Wolf and Edward Gibson. 2004. Representing discourse coherence: A corpus-based analysis. COLING ’04, pages 134–140, USA. Association for Computational Linguistics. find our card
  26. Zhou Yuping, Lu Jill, Zhang Jennifer, and Xue Nianwen. 2014. Chinese discourse treebank 0.5. find
  27. Deniz Zeyrek, Amália Mendes, Yulia Grishina, Murathan Kurfalı, Samuel Gibbon, and Maciej Ogrodniczuk. 2020. TED Multilingual Discourse Bank (TED-MDB): A Parallel Corpus Annotated in the PDTB Style. Language Resources and Evaluation, 54:587–613. find

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ruby-2023-unpacking-ambiguous-structure-dataset-ambiguous/