A general framework for information extraction using dynamic span graphs

Written for the Research library (claude-opus-5-5) from the paper text. Original — aclanthology.org/N19-1308.pdf

Summary

Entity recognition, relation extraction and coreference resolution all label text spans. Earlier joint or multi-task systems let these tasks interact only through a shared lower LSTM layer. In addition, BIO-style taggers cannot recognise nested entities. The authors propose DyGIE (Dynamic Graph IE), which enumerates every within-sentence span up to a fixed length. Each span is represented from BiLSTM states built over character, GLoVe and ELMo inputs. During training, the model builds a graph on the fly: beam pruning keeps the most confident spans as nodes, and edges are weighted by predicted coreference and relation scores. Span vectors are refined iteratively by gated updates, passed first along coreference links and then along relation links, before the final predictions are made. The method needs no external syntactic tools. DyGIE was tested on news, scientific and biomedical/lab-protocol data. It achieves new state-of-the-art results for joint entity–relation extraction and for overlapping entity recognition. This shows that explicitly sharing cross-sentence context improves span-based IE across domains.

Key points

The joint entity and relation extraction experiments cover four benchmarks: SciERC, the Wet Lab Protocol Corpus, ACE 2004 and ACE 2005. On ACE05, DyGIE reports relative gains of 5.7% (entities) and 9.9% (relations) over the previous best.
For overlapping entities, DyGIE improves on Wang and Lu (2018) by 11.6% on ACE04-O, 11.3% on ACE05-O and 1.5% on GENIA, a gain of up to 8 F1 points.
In ablations, relation propagation helps both tasks, particularly in sentences with many entities. Coreference propagation mainly helps entities on ACE05 and gives two-iteration-optimal results (N = 2).
Coreference propagation raises ACE05 pronoun entity F1 from 73.2 to 78.0, reducing GPE/PER and PER/ORG confusions. It has little effect on SciERC, where pronouns are labelled Generic.
Code and pre-trained models are released at github.com/luanyi/DyGIE.

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: Natural language processing

Topics: Coreference resolution

Cited by

116 papers of our library cite it (of the 140673 whose bibliographies we read); 5 of them have a page here, 111 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/n19-1308 · arXiv 1904.03296

@inproceedings{luan2019general,
  title = {{A general framework for information extraction using dynamic span graphs}},
  author = {Luan, Yi and Wadden, Dave and He, Luheng and Shah, Amy and Ostendorf, Mari and Hajishirzi, Hannaneh},
  booktitle = {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)},
  year = {2019},
  doi = {10.18653/v1/n19-1308},
  eprint = {1904.03296},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/n19-1308}
}

References

3 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Giannis Bekoulis, Johannes Deleu, Thomas Demeester, and Chris Develder. 2018. Adversarial training for multi-context joint entity and relation extraction. In Proc. Conf. Empirical Methods Natural Language Process. (EMNLP), pages 2830–2836. find
This research was supported by the Office of Naval Research under the MURI grant N00014-18-1-2670, NSF (IIS 1616112, III 1703166), Allen Distinguished Investigator Award, Samsung GRO and gifts from Allen Institute for AI, Google, Amazon, and Bloomberg. We also thank the anonymous reviewers and the U… find
Yee Seng Chan and Dan Roth. 2011. Exploiting syntactico-semantic structures for relation extraction. In Proc. Annu. Meeting Assoc. for Computational Linguistics (ACL). \#\# Conclusion find