Elamur Research⌕
☰
home / papers / 2026 / LLMBridge: An LLM Pipeline for End-to-end Referential Bridging Resolution in English

LLMBridge: An LLM Pipeline for End-to-end Referential Bridging Resolution in English

Paper · 2026 · Lauren Levine and Amir Zeldes

Summary

Bridging is when a new entity is understood through its link to an earlier one, as when "the door" is read as belonging to a house just mentioned. Automatically finding such anaphors and their antecedents from raw text has been very hard: earlier systems stayed below about 40% F1 for spotting anaphors and below about 30% for linking them. The authors present LLMBridge, which they describe as the first LLM-driven pipeline for this end-to-end task. It surrounds an LLM's inference ability with rule-based pre- and post-processing, and it takes predicted mentions and coreference from Stanza models. They test it on the three English corpora that follow the referential, information-status view of bridging: ISNotes, BASHI and GUMBridge. LLMBridge sets new state-of-the-art results on all three, both in the End-to-end setting and in a Basic setting where the gold anaphor is supplied. The error analysis shows that long-distance links and abstract relation types remain difficult. The code is released, which gives a stronger baseline for implicit entity relations, a phenomenon relevant to QA, generation and fact checking.

Key points

  • Evaluation corpora. All three are English:
    • ISNotes: WSJ news, about 40k tokens, 916 bridging instances.
    • BASHI: WSJ news, about 58k tokens, 459 instances.
    • GUMBridge: 24 genres, about 291k tokens, about 5.7k instances.
    • ARRAU was left out because it mixes lexical and referential bridging.
  • Three subtasks. The paper separates anaphor recognition, anaphor resolution (antecedent selection) and subtype classification.
  • Three settings. These differ in how much gold mention information the system receives: Basic, Full and End-to-end.
  • What is reported.
    • P/R/F1 in the End-to-end setting.
    • Accuracy in a modified Basic setting that gives the gold anaphor but no gold mentions.
    • Subtype scores on GUMBridge.
  • Mentions and scoring. Predicted mentions come from Stanza coref models trained on GUM and on OntoNotes. Scoring is at the entity level: any non-pronominal mention of the gold antecedent's cluster counts as correct.
  • Error analysis.
    • The strongest backend was Gemini-3.1-pro-preview.
    • ENTITY-ASSOCIATIVE, ENTITY-PROPERTY and SET-MEMBER instances are often missed.
    • RESULTATIVE was labelled ASSOCIATIVE 23 times.
    • Extra, unwarranted labels are mostly COMPARISON.
    • Evaluation cost about 1000 USD. LoRA fine-tuning of a mid-sized LLM is suggested as a cheaper direction.

Interesting ideas

  • I1. Pair deterministic pre/post-processing with an LLM that only does the inference step, so the hard semantic judgement goes to the model and span bookkeeping stays rule-based.
  • I2. Treat a bridging link as a typed relation between entities (part-whole, associative, resultative, comparison) rather than as a plain pointer.
  • I3. Score antecedents at the entity level, crediting any non-pronominal mention in the gold coreference cluster, so a correct link is not punished for choosing a different span of the same entity.
  • I4. Report results under graded amounts of gold input (gold anaphor given vs. raw text only), which shows how much error comes from upstream components.
  • I5. Use a confusion matrix with a "none" gold row to catch hallucinated labels, and to spot narrow classes absorbed by a broad one, which signals where the prompt needs more examples.

For Mova

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

  • I1 (frames): The coreference module could add a bridging pass in which rules propose candidate definite NPs and earlier entities, and an LLM teacher only confirms the link, keeping each decision traceable.
  • I2 (world): In the text world, "the door" after "a house" could be stored as a meronymy edge to the house entity, so later questions about whose door it is can be answered from the graph.
  • I3 (learn): Evaluation of coreference and frame-argument linking could use the same rule, separating real errors from span-choice noise.
  • I4 (learn): Each pipeline stage could be scored with gold versus predicted trees and mentions, revealing whether a failure belongs to the parser or to the frame layer.
  • I5 (learn): The same check on LLM-teacher outputs for VerbNet roles or speech acts would show which narrow categories collapse into a catch-all and need extra guidance.

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

🎓Georgetown UniversityWashington · United States88 works · 4,732 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 3Research subjects and languages the work is about.Language models & pretraining · 3,320 worksEnglish · 2,263 worksCoreference resolution · 574 works

Cite

Cite as Lauren Levine and Amir Zeldes. 2026. LLMBridge: An LLM Pipeline for End-to-end Referential Bridging Resolution in English. arXiv:2605.29048.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2605.29048 · arXiv 2605.29048

@misc{levine2026llmbridge,
  title = {{LLMBridge: An LLM Pipeline for End-to-end Referential Bridging Resolution in English}},
  author = {Lauren Levine and Amir Zeldes},
  year = {2026},
  doi = {10.48550/arxiv.2605.29048},
  eprint = {2605.29048},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2605.29048}
}

References

36 references, 24 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Nicholas Asher and Alex Lascarides. 1998. Bridging. Journal of Semantics, 15(1):83–113. find
  2. Stefan Baumann and Arndt Riester. 2012. Referential and lexical givenness: Semantic, prosodic and cognitive aspects. Prosody and meaning, 25:119–162. ↗
  3. Lanni Bu, Lauren Levine, and Amir Zeldes. 2025. DiscoTrack: A multilingual LLM benchmark for discourse tracking. arXiv preprint arXiv:2510.17013. ↗
  4. Herbert H. Clark. 1975. Bridging. In Theoretical Issues in Natural Language Processing. find
  5. Marie-Catherine de Marneffe, Christopher D. Manning, Joakim Nivre, and Daniel Zeman. 2021. Universal dependencies. Computational Linguistics, 47(2):255– 308.
  6. 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. ↗
  7. Yanai Elazar, Victoria Basmov, Yoav Goldberg, and Reut Tsarfaty. 2022. Text-based NP enrichment. Transactions of the Association for Computational Linguistics, 10:764–784. ↗
  8. Biaoyan Fang, Timothy Baldwin, and Karin Verspoor. 2022. What does it take to bake a cake? the RecipeRef corpus and anaphora resolution in procedural text. In Findings of the Association for Computational Linguistics: ACL 2022, pages 3481–3495, Dublin, Ireland. Association for Computational Linguistics. ↗
  9. John A. Hawkins. 1978. Definiteness and indefiniteness: A study in reference and grammaticality prediction. Journal of Linguistics, 27:405–442. find
  10. Yufang Hou. 2020. Bridging anaphora resolution as question answering. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1428–1438, Online. Association for Computational Linguistics. ↗
  11. Yufang Hou, Katja Markert, and Michael Strube. 2014. A rule-based system for unrestricted bridging resolution: Recognizing bridging anaphora and finding links to antecedents. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 2082–2093, Doha, Qatar. Association for Computational Linguistics.
  12. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. ICLR, 1(2):3.
  13. Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and Omer Levy. 2020. Span-BERT: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64–77.
  14. Sopan Khosla, Juntao Yu, Ramesh Manuvinakurike, Vincent Ng, Massimo Poesio, Michael Strube, and Carolyn Rosé. 2021. find
  15. The CODI-CRAC 2021 shared task on anaphora, bridging, and discourse deixis in dialogue. In Proceedings of the CODI-CRAC 2021 Shared Task on Anaphora, Bridging, and Discourse Deixis in Dialogue, pages 1–15, Punta Cana, Dominican Republic. Association for Computational Linguistics.
  16. Hideo Kobayashi, Yufang Hou, and Vincent Ng. 2022a. Constrained multi-task learning for bridging resolution. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 759–770, Dublin, Ireland. Association for Computational Linguistics. find
  17. Massimo Poesio and Ron Artstein. 2008. Anaphoric annotation in the ARRAU corpus. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), Marrakech, Morocco. European Language Resources Association (ELRA). ↗
  18. Ellen F. Prince. 1981. Toward a taxonomy of given-new information. Radical pragmatics, pages 223–255. find
  19. Hideo Kobayashi, Yufang Hou, and Vincent Ng. 2022b. End-to-end neural bridging resolution. In Proceedings of the 29th International Conference on Computational Linguistics, pages 766–778, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.
  20. Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. Stanza: A python natural language processing toolkit for many human languages. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 101–108, Online. Association for Computational Linguistics. ↗
  21. Hideo Kobayashi, Yufang Hou, and Vincent Ng. 2023. PairSpanBERT: An enhanced language model for bridging resolution. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6931– 6946, Toronto, Canada. Association for Computational Linguistics. ↗
  22. Ina Roesiger, Arndt Riester, and Jonas Kuhn. 2018. Bridging resolution: Task definition, corpus resources and rule-based experiments. In Proceedings of the 27th International Conference on Computational Linguistics, pages 3516–3528, Santa Fe, New Mexico, USA. Association for Computational Linguistics.
  23. Hideo Kobayashi and Vincent Ng. 2020. Bridging resolution: A survey of the state of the art. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3708–3721, Barcelona, Spain (Online). International Committee on Computational Linguistics. ↗
  24. Ina Rösiger. 2018. BASHI: A corpus of Wall Street Journal articles annotated with bridging links. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018), Miyazaki, Japan. European Language Resources Association (ELRA). ↗
  25. Hideo Kobayashi and Vincent Ng. 2021. Bridging resolution: Making sense of the state of the art. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1652–1659, Online. Association for Computational Linguistics. ↗
  26. Olga Uryupina, Ron Artstein, Antonella Bristot, Federica Cavicchio, Francesca Delogu, Kepa Joseba Rodríguez, and Massimo Poesio. 2019. Annotating a broad range of anaphoric phenomena, in a variety of genres: the ARRAU corpus. Natural Language Engineering, 26:95 – 128. ↗
  27. Lauren Levine and Amir Zeldes. 2025. Subjectivity in the annotation of bridging anaphora. In Proceedings of the 19th Linguistic Annotation Workshop (LAW-XIX-2025), pages 48–59, Vienna, Austria. Association for Computational Linguistics. ↗
  28. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30.
  29. Lauren Levine and Amir Zeldes. 2026. GUMBridge: A corpus for varieties of bridging anaphora. In Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026), pages 6823– 6837, Palma, Mallorca, Spain. European Language Resources Association (ELRA). ↗
  30. Ralph Weischedel, Sameer Pradhan, Lance Ramshaw, Martha Palmer, Nianwen Xue, Mitchell Marcus, Ann Taylor, Craig Greenberg, Eduard Hovy, Robert Belvin, and Ann Houston. 2011. OntoNotes release 4.0. LDC2011T03,Philadelphia, Penn.: Linguistic Data Consortium, 17. ↗
  31. Katja Markert, Yufang Hou, and Michael Strube. 2012. Collective classification for fine-grained information status. In Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 795–804, Jeju Island, Korea. Association for Computational Linguistics.
  32. Juntao Yu, Sopan Khosla, Ramesh Manuvinakurike, Lori Levin, Vincent Ng, Massimo Poesio, Michael Strube, and Carolyn Rosé. 2022. find
  33. The CODI-CRAC 2022 shared task on anaphora, bridging, and discourse deixis in dialogue. In Proceedings of the CODI-CRAC 2022 Shared Task on Anaphora, Bridging, and Discourse Deixis in Dialogue, pages 1–14, Gyeongju, Republic of Korea. Association for Computational Linguistics.
  34. Malvina Nissim, Shipra Dingare, Jean Carletta, and Mark Steedman. 2004. An annotation scheme for information status in dialogue. In Proceedings of the Fourth International Conference on Language Resources and Evaluation (LREC’04), Lisbon, Portugal. European Language Resources Association (ELRA). ↗
  35. Juntao Yu and Massimo Poesio. 2020. Multitask learning-based neural bridging reference resolution. In Proceedings of the 28th International Conference on Computational Linguistics, pages 3534–3546, Barcelona, Spain (Online). International Committee on Computational Linguistics. ↗
  36. Amir Zeldes. 2017. The GUM corpus: Creating multilayer resources in the classroom. Language Resources and Evaluation, 51(3):581–612. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/levine-2026-llmbridge-llm-pipeline-end-end/