CorPipe at CRAC 2026: Empty Nodes and Cross-Lingual Transfer in Multilingual Coreference Resolution

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

Summary

CorPipe 26 is the winning submission to the CRAC 2026 Shared Task on Multilingual Coreference Resolution, a competition focused on comparing generative LLMs and specialized systems, which also introduced 5 new datasets and 2 new languages. This system, an enhanced version of CorPipe 25, incorporates a novel approach that simultaneously predicts empty nodes, mentions, and coreference links using a single model. CorPipe 26 demonstrated superior performance, surpassing all other entries in the LLM track by 2.8 percent points and all submissions in the unconstrained track by 9.5 percent points. The research further includes ablation experiments exploring various model sizes, empty node prediction techniques, and cross-lingual zero-shot evaluation. Its source code and trained models are publicly accessible.

Key takeaway

For NLP Engineers developing multilingual coreference resolution systems, CorPipe 26 offers a validated, high-performing architecture. You should consider its single-model approach for predicting empty nodes alongside mentions and coreference links, as it significantly outperformed other systems in the CRAC 2026 task. Leverage the publicly available source code and trained models to benchmark your own specialized systems or to integrate its methodology for improved accuracy in diverse linguistic contexts.

Key insights

CorPipe 26 integrates empty node prediction with coreference resolution, achieving top performance in multilingual tasks.

Principles

• Single-model prediction for empty nodes and coreference improves performance.
• Ablation studies are crucial for understanding model components.
• Cross-lingual zero-shot evaluation is a key research area.

Method

CorPipe 26 uses a single model to predict empty nodes, mentions, and coreference links, building upon CorPipe 25's architecture. It involves ablation experiments for optimization.

In practice

• Access CorPipe 26's code for multilingual coreference.
• Compare specialized systems against generative LLMs.
• Explore empty node prediction in NLP tasks.

Topics

• Multilingual Coreference Resolution
• Empty Node Prediction
• Cross-Lingual Transfer
• CRAC 2026 Shared Task
• CorPipe 26
• Generative LLMs

Code references

• ufal/crac2026-corpipe

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.
• Multilingual Coreference Resolution via Cycle-Consistent Machine Translation — Coreference resolution in low-resource languages can be significantly improved by cycle-consistent machine translation data augmentation.
• Closing the Gap at CRAC 2026: Two-Stage Adaptation for LLM-Based Multilingual Coreference Resolution — Two-stage fine-tuning and optimized input formats enable LLMs to achieve state-of-the-art multilingual coreference resolution.
• Plug-and-Adapt: Multimodal Coreference Resolution at First Sight with a Pretrained Alignment Model — Plug-and-Adapt repurposes a pre-trained vision-language alignment model for zero-shot multimodal coreference resolution, outperforming existing methods.
• CLP-Transfer: Cross-Lingual and Progressive Transfer Learning — CLP-Transfer simplifies cross-lingual model transfer using token overlap and a small helper model, but shows limited downstream task performance.
• End-to-end Neural Coreference Resolution in spaCy — spaCy now features an experimental end-to-end neural coreference resolution component, enhancing natural language understanding.

Open in AIssential →

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