Unifying the Scope of Bridging Anaphora Types in English: Bridging Annotations in ARRAU and GUM

What it is about

The two largest English bridging resources define bridging differently. GUM/GENTLE (16 + 8 genres, 1.9k bridging links) treat it as information status: a first mention that is interpretable only through a non-identical earlier entity, with one antecedent. ARRAU (news part 229k tokens, 3.7k links) treats it as a set of semantic relations (part, element, subset, other, each with an inverse), allows already-mentioned anaphors and split antecedents, and has subtypes. The authors remove ARRAU cases that GUM's definition excludes (297 split antecedents, 957 anaphors with an identity antecedent), unify entity types (person, place, organization, object, event, time, substance, animal, plant, abstract), train XGBoost pair classifiers on each corpus, and compare: in-corpus F 0.71 (GUM) and 0.67 (ARRAU), cross-corpus 0.56 and 0.22. Definiteness of the anaphor is the strongest cue in GUM, number in ARRAU. Genre-specific bridging appears in GUM only: a biography heading «Early life» bridging to the person, figure labels («[left]») to the figure, a reply by one forum user to a question addressed to many. They release harmonised test sets with ARRAU-style subtypes added to GUM.

Ideas

I1. A rule baseline for bridging anaphor detection on Mova's trees, in GUM's sense: a definite (the, possessive-less) common-noun mention that is first in its chain after coref resolution and not generic is a bridging candidate. Score it on GUM dev bridging (Bridge= in MISC, which coref resolve now drops) for precision and recall before any antecedent rule; it tells how much the cheap cue carries in Mova's genres.
I2. Build a semantic class lexicon for mentions with the paper's unified ten types, from licence-clean sources (WordNet's top categories, Mova's own gendered-noun and animacy lists), and use GUM's entity types on dev only to measure its accuracy. The coref README lists semantic classes as needed for «it» and relaxed_head; the paper's table gives the target inventory and the GUM mapping for scoring.
I3. Treat headings and captions as bridging to the document's main entity: in Mova's lessons a heading line («Habitat», «Early life») has the lesson's topic as its implicit owner. In the student (world/src/knowledge.rs) such a line should attach its frames to the topic entity, not create a new character — the same family of error as the references read as characters in Watching the SLM think: the diagnosis of its answers. Measure on the courses with headings.

For Mova

35. Reference and coreference §3.2 (bridging), §7 (annotating coreference). Code: coref/src/mention.rs (semantic class), coref/src/sieve.rs and coref/src/corefud.rs (bridging candidates, keep Bridge for scoring), world/src/knowledge.rs (headings). Measures: bridging detection P/R on GUM dev; class accuracy against GUM entity types on dev; course test counts. GUM is NC and ARRAU LDC — evaluation only, no lexicons from them.

Topics: Coreference resolution

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2024.crac-1.5 · arXiv 2410.01170

@inproceedings{levine2024unifying,
  title = {{Unifying the Scope of Bridging Anaphora Types in English: Bridging Annotations in ARRAU and GUM}},
  author = {Levine, Lauren and Zeldes, Amir},
  booktitle = {Proceedings of the Seventh Workshop on Computational Models of Reference, Anaphora and Coreference},
  year = {2024},
  doi = {10.18653/v1/2024.crac-1.5},
  eprint = {2410.01170},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.18653/v1/2024.crac-1.5}
}

References

1 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Tatsuya Aoyama, Shabnam Behzad, Luke Gessler, Lauren Levine, Jessica Lin, Yang Janet Liu, Siyao Peng, Yilun Zhu, and Amir Zeldes. 2023. GENTLE: A genre-diverse multilayer challenge set for English NLP and linguistic evaluation. In Proceedings of the 17th Linguistic Annotation Workshop (LAW-XVII), pages 166–178, Toronto, Canada. Association for Computational Linguistics. *Table 8: Distribution of bridging subtypes in harmonized GUM/GENTLE test and ARRAU WSJ test.* find our card

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/levine-2024-unifying-scope-bridging-anaphora-types/