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

2026-05-28 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

CorPipe 26, the winning submission to the CRAC 2026 Shared Task on Multilingual Coreference Resolution, introduces an improved system that predicts empty nodes alongside mentions and coreference links within a single model. This fifth edition of the task focused on comparing generative LLMs and specialized systems, incorporating 5 new datasets and 2 new languages. CorPipe 26 significantly outperformed all other submissions, achieving a 2.8 percentage point lead in the LLM track and a 9.5 percentage point lead in the unconstrained track. The authors also conducted ablation experiments on model sizes, empty node prediction methods, and cross-lingual zero-shot evaluation. The source code and trained models are publicly available at https://github.com/ufal/crac2026-corpipe.

Key takeaway

For NLP Engineers or AI Scientists focused on advancing multilingual coreference resolution, CorPipe 26 demonstrates a critical architectural improvement. Its integrated approach to predicting empty nodes, mentions, and coreference links within a single model significantly outperforms both generative LLMs and other specialized systems. You should consider adopting this unified empty node prediction strategy or leveraging the publicly available CorPipe 26 models and source code to enhance your own multilingual NLP pipelines.

Key insights

CorPipe 26 integrates empty node prediction with mention and coreference resolution for improved multilingual performance.

Principles

• Integrating empty node prediction enhances coreference resolution.
• Specialized systems can outperform generative LLMs.
• Cross-lingual transfer is crucial for multilingual NLP.

Method

CorPipe 26 employs a single model to predict empty nodes, mentions, and coreference links, building upon CorPipe 25's architecture for multilingual coreference.

In practice

• Explore empty node prediction for coreference tasks.
• Evaluate specialized models against LLMs for specific NLP.
• Utilize CorPipe 26's public code for multilingual coreference.

Topics

• Multilingual Coreference Resolution
• Empty Node Prediction
• Cross-Lingual Transfer
• Generative LLMs
• Specialized NLP Systems
• CRAC Shared Task

Code references

• ufal/crac2026-corpipe

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

Related on AIssential

• Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis — Cross-lingual transfer strategies for ABSA vary in effectiveness based on model architecture and resource availability.
• 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.
• Just Use XML: Revisiting Joint Translation and Label Projection — Joint translation and label projection with XML tags can improve both translation quality and label transfer.
• 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.
• Event and Entity Coreference Across Five Languages: Effects of Context and Referring Expression — Crosslingual coreference biases show personal pronouns favor entities, demonstratives favor events, influenced by antecedent complexity.
• Casual as an Anchor: Resolving Supervision Misalignment in Formality Transfer Dataset — Existing formality transfer benchmarks misalign supervision, leading models to generate pseudo-formal text; a graded "casual" anchor resolves this.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.