Elamur Research⌕
☰
home / papers / 2021 / Sequential Cross-Document Coreference Resolution

Sequential Cross-Document Coreference Resolution

Summary

The paper addresses cross-document coreference resolution (CDCR), which groups mentions of the same entity or event across several texts. This matters for multi-document tasks such as summarization and reading comprehension. Earlier event coreference systems usually score mention pairs independently and then cluster them with tuned thresholds. The authors instead adapt the left-to-right sequential prediction approach, previously used for within-document entity coreference, to cross-document entity and event coreference.

Documents are processed in an ordered stream. Each new mention, encoded with BERT together with its event or entity arguments, is linked to one of the clusters already built from earlier documents, or marked as a singleton. Cluster representations are composed incrementally from mention and document-level (CLS) embeddings, which approximates higher-order inference in a single pass.

On ECB+ the model is competitive but not state of the art. It runs much faster than graph-based models and suits streaming settings, since new documents do not require rerunning the whole collection. Extensive ablations show which inputs matter most.

Key points

  • On the ECB+ test set, the best model (SeqXdoc+IC + adaptive pre-training) reaches CoNLL F1 of 75.3 for entities and 80.8 for events, below Caciularu et al. (2021) at 82.9 and 85.6.
  • Incremental candidate clustering (IC) adds about 8.9 F1 for entities and 2.8 F1 for events over the plain sequential cross-document baseline.
  • Event coreference depends heavily on argument information:
    • Replacing the argument BiLSTM with mean-pooling costs 2.5 F1.
    • Removing arguments entirely costs 2.1 F1.
  • Swapping BERT-base for RoBERTa-base hurts performance (entity −4.1, event −2.7). The authors attribute this to BERT's Next Sentence Prediction objective.
  • Using the newer SRL parser of He et al. (2018) makes the extra alignment rules of Barhom et al. (2019) unnecessary and even harmful. With the older Swirl parser plus those rules, event F1 drops by 2.1.

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

Cited by

11 papers of our library cite it (of the 140673 whose bibliographies we read); 5 of them have a page here, 6 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2021.emnlp-main.382 · arXiv 2104.08413

@inproceedings{allaway2021sequential,
  title = {{Sequential Cross-Document Coreference Resolution}},
  author = {Allaway, Emily and Wang, Shuai and Ballesteros, Miguel},
  booktitle = {Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  doi = {10.18653/v1/2021.emnlp-main.382},
  eprint = {2104.08413},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/2021.emnlp-main.382}
}

References

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

  1. A. Bagga and B. Baldwin. 1998. Algorithms for scoring coreference chains. find our card
  2. Shany Barhom, Vered Shwartz, Alon Eirew, Michael Bugert, Nils Reimers, and I. Dagan. 2019. Revisiting joint modeling of cross-document entity and event coreference resolution. In ACL. find our card
  3. C. Bejan and Sanda M. Harabagiu. 2010. Unsupervised event coreference resolution with rich linguistic features. In ACL. find our card
  4. Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. ArXiv, abs/2004.05150. find
  5. Anders Björkelund and Jonas Kuhn. 2014. Learning structured perceptrons for coreference resolution with latent antecedents and non-local features. In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 47–57. find our card
  6. Avi Caciularu, Arman Cohan, Iz Beltagy, Matthew E. Peters, Arie Cattan, and Ido Dagan. 2021. Cross-document language modeling. ArXiv, abs/2101.00406. — [arXiv:2101.00406](https://arxiv.org/abs/2101.00406) find
  7. Jie Cai and Michael Strube. 2010. Evaluation metrics for end-to-end coreference resolution systems. In SIGDIAL Conference. find our card
  8. Arie Cattan, Alon Eirew, Gabriel Stanovsky, Mandar Joshi, and I. Dagan. 2020. Streamlining cross-document coreference resolution: Evaluation and modeling. ArXiv, abs/2009.11032. find
  9. Prafulla Kumar Choubey and Ruihong Huang. 2017. Event coreference resolution by iteratively unfolding inter-dependencies among events. In EMNLP. find our card
  10. Kevin Clark and Christopher D Manning. 2016. Deep reinforcement learning for mention-ranking coreference models. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2256–2262. find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/allaway-2021-sequential-cross-document-coreference-resolution/