Elamur Research⌕
☰
home / papers / 2019 / Q-learning pour la résolution des anaphores pronominales en langue arabe (Q-learning for pronominal anaphora resolution in Arabic texts)

Q-learning pour la résolution des anaphores pronominales en langue arabe (Q-learning for pronominal anaphora resolution in Arabic texts)

Summary

The paper tackles pronominal anaphora resolution for Arabic. Arabic is morphologically rich, pronouns are frequent (the authors count 18.2% of words in literary texts), and corpora annotated with anaphoric links are scarce. The covered pronouns are personal subject and object pronouns, demonstratives and relatives. The pipeline has two steps:

  • Identification: pronouns are found by their grammatical values, and rule-based filtering removes non-referential ones.
  • Resolution: a candidate antecedent is picked from a window of preceding words.

Resolution is framed as reinforcement learning. A Markov decision process is built whose actions are morphosyntactic preference criteria and whose states are combinations of those criteria. Q-learning searches, per pronoun, for the criterion sequence that best ranks the candidates. Its reward matrix is seeded in a pre-learning phase from a few labelled texts.

On small test sets of literary, technical-manual and journalistic texts, the system reaches 70–77% precision. In every genre it beats a Mitkov-style weighted-criteria baseline run on the same texts. The work matters because it shows a low-resource alternative to fully supervised coreference.

Key points

  • Seven criteria are used: "Définition" (definite NPs), "Thème" (sentence subjects), "Distance", "Tête de paragraphe", "Nom propre", "Répétition" (repeated lemmas) and a new one, "Antécédent pronom précédent".
  • The new criterion favours the antecedent chosen for the previous pronoun. About 35% of antecedents in Arabic literary texts satisfy it.
  • Precision of the Q-learning system versus the Mitkov baseline:
GenreWordsPronounsQ-learningMitkov
Literary161521870.11%64.67%
Technical manuals171413672.8%53.04%
Journalistic8728277.14%50%
  • The baseline uses the same criteria, weighted with Mitkov's fixed relevance values (-1, 0, 1, 2).
  • If the correct antecedent only needs to be among the top two candidates, precision reaches 84.29% on journalistic texts. Among the top three, it reaches 87.7% on technical manuals.
  • Planned future work: larger journalistic and technical corpora, and semantic criteria.

Interesting ideas

  • I1. Instead of fixed global weights, the system learns which subset and order of salience criteria to apply for each pronoun context, because a criterion's usefulness shifts with pronoun type and genre. — For Mova (frames): The coreference step in prag could keep its explainable preference rules but learn per-context rule sequences, so every antecedent choice still traces back to named criteria.
  • I2. A cheap discourse-continuity heuristic, "reuse the previous pronoun's antecedent", covered about a third of cases in narrative text. — For Mova (frames): This fits fables and tales well, where one protagonist dominates. It could be added as an explicit, inspectable coreference preference.
  • I3. The baseline is given exactly the same criteria and differs only in how they are combined, so the gain is attributed to the learning scheme rather than to richer features. — For Mova (learn): When testing a learned combiner over hand rules, the hand-weighted version over the same features can serve as the control.
  • I4. Reporting accuracy at ranks 1, 2 and 3 for the correct antecedent shows how much a later reranking step could still recover. — For Mova (frames): A coreference evaluation could log top-k antecedent hit rates, showing whether world-level checks (states, holdings) could fix the remaining errors.
  • I5. The reward model is bootstrapped from only a few annotated texts, and learning continues from there, which suits languages or domains without large coreference corpora. — For Mova (learn): Small hand-checked or LLM-teacher-labelled samples could seed a reward signal for tuning discourse rules, instead of requiring a full gold corpus.

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: Coreference resolution

Cite

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

@inproceedings{bouzid2019q,
  title = {{Q-learning pour la résolution des anaphores pronominales en langue arabe (Q-learning for pronominal anaphora resolution in Arabic texts)}},
  author = {Bouzid, Saoussen Mathlouthi and Zribi, Chiraz Ben Othmane},
  booktitle = {Actes de la Conférence sur le Traitement Automatique des Langues Naturelles (TALN) PFIA 2019. Volume II : Articles courts},
  year = {2019},
  url = {https://aclanthology.org/2019.jeptalnrecital-court.25/}
}

References

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

  1. Mitkov R. (1998). Robust pronoun resolution with limited knowledge. Montreal, Canada: Proceedings of the 18.th International Conference on Computational Linguistics (COLING' 98)/A CL' 98. find our card
  2. Mitkov R., Belguith L., Stys M. (1998). Multilingual robust anaphora resolution, Granada, Spain. find our card
  3. Schmolz H., Coquil D., Döller M. (2012). In-Depth Analysis of Anaphora Resolution Requirements. 2012 23rd International Workshop on Database and Expert Systems Applications. Vienna, Austria. find
  4. Seminck O., Amsili P. (2017). A Computational Model of Human Preferences for Pronoun Resolution. Proceedings of the Student Research Workshop at the 15th Conference of the European Chapter of the Association for Computational Linguistics, pages 53–63, Valencia, Spain. find
  5. Ashima A., Mohana B. (2016). Improving Anaphora Resolution by Resolving Gender and Number Agreement in Hindi Language using Rule based Approach. Indian Journal of Science and Technology, Vol 9(32), august 2016. find
  6. Aone C., Bennet S. (1995). Applying machine learning to anaphora resolution. International Joint Conference on Artificial Intelligence IJCAI 1995: Connectionist, Statistical and Symbolic Approaches to Learning for Natural Language Processing pp 302-314, 1995. Elghamry K., Al-Sabbagh R., El-Zeiny N. Arabic Anaphora Resolution Using Web as Corpus. Proceedings of the seventh conference on language engineering, pp. 1-18. find
  7. Weissenbacher D., Nazarenko A. (2007). A bayesian classifier for the recognition of the impersonal occurrences of the it pronoun. In Proceedings of DAARC’07, 2007. 29, 32, 43, 72, 120. find
  8. Hammami S. (2016). La résolution automatique des anaphores pronominales pour la langue arabe, thèse de doctorat, Université de Sfax, Faculté des Sciences Economiques et de Gestion, Sfax, Tunisie. find
  9. Mathlouthi S., Ben Fraj Trabelsi F., Ben Othmane Zribi C. (2016). A Novel Approach Based on Reinforcement Learning for Anaphora Resolution. 28th IBIMA Conference, November 2016. find
  10. Mathlouthi S., Ben Fraj F., Ben Othmane C. (2017). How to combine salience factors for Arabic Pronoun Anaphora Resolution. 2017 IEEE/ACS 14th International Conference on Computer Systems and Applications. find
  11. Lappin S., Leass H. J. (1994). An Algorithm for Pronominal Anaphora Resolution. London, Computational Linguistics, 20(4), 535-561. find our card
  12. Li D., Miller T., Schuler W. (2011). A pronoun anaphora resolution system based on factorial hidden markov models. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, Portland, Oregon, June 19-24, page 1169–1178, 2011. 42, 115. find our card
  13. Aktas B.¸ Scheffler T., Stede M. (2018). Anaphora Resolution for Twitter Conversations: An Exploratory Study. Proceedings of the Workshop on Computational Models of Reference, Anaphora and Coreference, pages 1–10. New Orleans, Louisiana, June 6, 2018. find our card
  14. Charniak E., Elsner M. (2009). EM works for pronoun anaphora resolution. In Proceedings of EACL, pp. 48-156. find our card
  15. Kamune K., Agrawal A. (2015). Hybrid Approach to Pronominal Anaphora Resolution in English Newspaper Text. International Journal of Intelligent Systems and Applications, 02: pp. 56-64. find
  16. Published Online January 2015 in MECS. DOI: 10.5815/ijisa.2015.02.08. ↗
  17. Abolohom A., Omar N. (2015). A Hybrid Approach to Pronominal Anaphora Resolution in Arabic, Journal of Computer Sciences 11(5): 764-71 DOI: 10.3844/jcssp.2015.764.771. ↗

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/bouzid-2019-q-learning-pour-la-resolution/