manu-pac/SummIt-CorefUD

Dataset Card for Summ-it++ CorefUD

Dataset Summary

Summ-it++ CorefUD is a standardized, harmonized version of a 10% subset of the Summ-it++ corpus (the second largest coreference corpus for Portuguese). Adapted for evaluation in Universal Dependencies / CorefUD format, it contains two distinct variants designed for benchmarking entity coreference resolution models and LLMs in Portuguese:

1. Summ-it++ Original Harmonizado: A format-harmonized version of the original Summ-it++ annotations converted into CoNLL-U / CorefUD format without altering coreference annotations.
2. Summ-it++ Revisado (sem eventos): A manually revised version following OntoNotes 5.0 guidelines, focused strictly on entity coreference (excluding event coreference, bridging, and coordinated antecedents) to correct original annotation errors and omissions.

• Language: Portuguese (pt)
• License: Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0)
• Domain: News / Science Journalism (Caderno de Ciências, Folha de S. Paulo)

---

Dataset Modifications & Pipeline

To adapt Summ-it++ for modern NLP evaluation tools (specifically the CorefUD Scorer), the author performed several structural and linguistic modifications:

1. Morphosyntactic & Dependency Parsing (Harmonization)

• UDPipe Integration: Re-annotated tokenization, lemmatization, POS tags, and dependency trees automatically using UDPipe to align with Universal Dependencies (UD) standards.
• Head Assignment via Udapi: Original "Head" annotations were discarded due to structural limitations (lack of full dependency trees and multiple NP heads per mention). Instead, dependency trees from UDPipe were used alongside the Udapi Python framework (MoveHead block) to automatically assign mention heads.
• Removal of Entity Types: Named Entity labels (PER, ORG, etc.) were excluded from the CorefUD format because they were missing in most mentions and exhibited span mismatches relative to coreference mention boundaries.

2. Version Variations

Feature / Aspect	Summ-it++ Original Harmonizado	Summ-it++ Revisado (sem eventos)
Annotation Basis	Original Summ-it++ manual annotations	OntoNotes 5.0 guidelines adaptation
Coreference Chains	50 chains	55 chains (+10%)
Total Mentions	157 mentions	195 mentions (+24.2%)
Scope	Unrevised (contains raw annotations)	Strictly Entity Coreference (No Events, No Bridging)
Manual Corrections	None	Fixed overlapping chain IDs, missing mentions, unannotated entities, and bad spans

---

Detailed Version Descriptions

Version 1: Summ-it++ Original Harmonizado

This version preserves the raw manual coreference annotations from the original Summ-it++ corpus while placing them into the standardized CoNLL-U / CorefUD schema.

• Known Issues Preserved: Mentions assigned simultaneously to multiple chains (e.g. (28|29), omitted mentions in second occurrences, unindexed entities, and span boundary mismatches.
• Use Case: Baseline comparison for legacy Portuguese coreference systems.

Version 2: Summ-it++ Revisado (sem eventos)

This version undergoes a systematic manual revision based on the OntoNotes 5.0 coreference guidelines to eliminate noise and inconsistencies:

• Exclusion of Event Coreference: Event coreference was removed to align with standard entity coreference benchmarks.
• Exclusion of Bridging & Coordinated Antecedents: Complex indirect anaphora and multi-antecedent coordination were omitted for evaluation consistency.
• Fixes Applied:
  • Resolved conflicting double-chain assignments on single tokens.
  • Added missing coreference mentions to existing chains (e.g., adding omitted second occurrences of nominal phrases).
  • Created missing chains for valid coreferent entities previously left unannotated.
  • Adjusted mention spans to remove non-referential tail modifiers.
• Impact: When evaluating Large Language Models (e.g., GPT-4o, Custom GPTs), this revised version yielded significantly higher Precision scores without degrading Recall, confirming that original annotation omissions previously penalized valid model outputs.

---

Dataset Format & Fields

Data is provided in standard CoNLL-U / CorefUD format (10 tab-separated CoNLL-U columns + CorefUD Entity column):

1. ID: Word index (integer or range).
2. FORM: Word form or punctuation symbol.
3. LEMMA: Lemma or stem.
4. UPOS: Universal part-of-speech tag.
5. XPOS: Language-specific part-of-speech tag.
6. FEATS: List of morphological features.
7. HEAD: Head of the current word (dependency parent).
8. DEPREL: Universal dependency relation.
9. DEPS: Enhanced dependency graph.
10. MISC: CorefUD annotations including Entity=(e1-noun-...).

---

Citation & Attribution

If you use this dataset, please cite the original dissertation and the source corpora:

@mastersthesis{andrade2024engenharia,
  author       = {Manuela Pacheco de Andrade},
  title        = {Uso da engenharia de prompt em LLMs para a resolução de correferência na língua portuguesa},
  school       = {Universidade Federal de Santa Catarina (UFSC)},
  year         = {2024},
  address      = {Florianópolis, Brasil}
}

@inproceedings{antonitsch2016summit,
  title={Summ-it++: an enriched version of the Summ-it corpus},
  author={Antonitsch, Andr{'e} and others},
  booktitle={Proceedings of LREC 2016},
  pages={2047--2051},
  year={2016}
}

@inproceedings{nedoluzhko2022corefud,
  title={CorefUD 1.0: Coreference Meets Universal Dependencies},
  author={Nedoluzhko, Anna and others},
  booktitle={Proceedings of LREC 2022},
  pages={4859--4872},
  year={2022}
}