Closing the Gap: Robust Multilingual Coreference Resolution with DAgger

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

DAggerCoref, a system submitted to the CRAC 2026 Shared Task on Multilingual Coreference Resolution, achieves a macro CoNLL F1 of 67.56 across 27 datasets and 19 languages. This three-stage cascade system is built on XLM-RoBERTa-large, incorporating a gap classifier for zero pronoun detection, a mention head classifier, and a coarse-to-fine antecedent scorer. A central contribution is the application of DAgger (Ross et al., 2011) to coreference resolution, fine-tuning the antecedent scorer on a 50/50 mix of gold and pipeline-predicted mentions. This approach closes the train/test distribution mismatch, improving development set macro CoNLL F1 by 1.10 points. The system also introduces Otsu adaptive thresholding for zero pronoun detection, which matches gold-tuned per-dataset thresholds without requiring gold supervision.

Key takeaway

For NLP Engineers developing robust multilingual coreference systems, you should consider integrating DAgger into your training pipeline. Applying DAgger to fine-tune antecedent scorers on a mix of gold and pipeline-predicted mentions can significantly improve macro CoNLL F1 scores by addressing train/test distribution shifts. Additionally, explore Otsu adaptive thresholding for zero pronoun detection to match gold-tuned performance without requiring explicit gold supervision.

Key insights

Applying DAgger to coreference resolution effectively mitigates train/test distribution mismatch, boosting performance.

Principles

• DAgger improves model robustness by fine-tuning on mixed gold and pipeline data.
• Otsu adaptive thresholding can replace gold-tuned thresholds for zero pronoun detection.

Method

DAggerCoref uses a three-stage cascade: a gap classifier, a mention head classifier, and an antecedent scorer. The antecedent scorer is fine-tuned with DAgger on a 50/50 mix of gold and pipeline-predicted mentions.

In practice

• Integrate DAgger for fine-tuning coreference antecedent scorers.
• Employ Otsu adaptive thresholding for zero pronoun detection.

Topics

• Coreference Resolution
• Multilingual NLP
• DAgger
• XLM-RoBERTa
• Zero Pronoun Detection
• Otsu Thresholding

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• CorPipe at CRAC 2026: Empty Nodes and Cross-Lingual Transfer in Multilingual Coreference Resolution — CorPipe 26 integrates empty node prediction with mention and coreference resolution for improved multilingual performance.
• Koshur Diacritizer: A Byte-Level Sequence-to-Sequence Model for Kashmiri Diacritic Restoration — Koshur Diacritizer is a ByT5-small model that restores Kashmiri diacritics using a new 23.7k sentence pair dataset.
• Code-Switching Reveals Language Anchoring in Multilingual LLMs — Grammatical frame dictates multilingual LLM internal language anchoring, impacting code-switched performance.
• Coreference Resolution in spaCy — spaCy's experimental "coref" component enables robust coreference resolution for NLP applications.
• Multi-Stage Training for Abusive Comment Detection in Indic Languages — Multi-stage training with language preprocessing and model ensembles improves abusive comment detection while minimizing false positives.
• A Larger Annotated Corpus of Portuguese Coreference — LLMs can annotate Portuguese coreference, but a new LLM-based system did not outperform existing rule-based methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.