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

2026-05-28

AI Summary

This work addresses the challenges of weak null element modeling and limited cross-lingual transferability in multilingual coreference resolution by proposing CorPipe 26, a unified framework that jointly predicts mentions, coreference links, and null elements. For the first time, null elements and coreference relations are co-modeled within a single architecture, leveraging multilingual pre-trained language models and a zero-shot cross-lingual transfer strategy to substantially enhance performance. Evaluated on the CRAC 2026 shared task, the system outperforms the second-best submission by 2.8 percentage points in the LLM track and by 9.5 percentage points in the unconstrained track, demonstrating its effectiveness and strong generalization capability across languages.

Abstract

We introduce CorPipe 26, our winning submission to the CRAC 2026 Shared Task on Multilingual Coreference Resolution. The fifth edition of this shared task focuses mainly on the comparison of generative LLMs and specialized systems; additionally, 5 more datasets and 2 new languages are introduced. CorPipe 26 is an improved version of CorPipe 25, with a new variant predicting empty nodes together with mentions and coreference links in a single model. Our system outperforms all other submissions in the LLM track by 2.8 percent points and all submissions in the unconstrained track by 9.5 percent points. Furthermore, we perform a series of ablation experiments with different model sizes, empty node prediction methods, and cross-lingual zero-shot evaluation. The source code and the trained models are publicly available at https://github.com/ufal/crac2026-corpipe.

Problem

• multilingual coreference resolution
• empty nodes
• cross-lingual transfer
• zero-shot evaluation

Innovation

• empty nodes
• multilingual coreference resolution
• cross-lingual transfer
• unified prediction model
• zero-shot evaluation