CorPipe at CRAC 2025: Evaluating Multilingual Encoders for Multilingual Coreference Resolution

TL;DR

This paper presents CorPipe 25, the winning entry in CRAC 2025 for Multilingual Coreference Resolution, featuring a full PyTorch reimplementation of the prior winning system and a three-stage pipeline that predicts empty nodes, detects mentions, and performs coreference linking via antecedent maximization. It introduces a new LLM track, reduced mini-dev/mini-test sets, and augmented data, along with extensive ablations and cross-lingual evaluations across multilingual corpora. The results show CorPipe 25 substantially outperforms all competitors by 7–8 percentage points, with the ensemble variant achieving the best overall performance; the work also compares PyTorch versus TensorFlow, demonstrating practical advantages of the PyTorch implementation. The authors release the source code, trained models, and three pretrained multilingual encoders, highlighting the value of multilingual pretraining and long-context modeling for coreference resolution in diverse languages.

Abstract

We present CorPipe 25, the winning entry to the CRAC 2025 Shared Task on Multilingual Coreference Resolution. This fourth iteration of the shared task introduces a new LLM track alongside the original unconstrained track, features reduced development and test sets to lower computational requirements, and includes additional datasets. CorPipe 25 represents a complete reimplementation of our previous systems, migrating from TensorFlow to PyTorch. Our system significantly outperforms all other submissions in both the LLM and unconstrained tracks by a substantial margin of 8 percentage points. The source code and trained models are publicly available at https://github.com/ufal/crac2025-corpipe.

Figures (2)

• Figure 1: The system architecture of the empty node prediction baseline. Every ReLU activation is followed by a dropout layer with a dropout rate of 50%.
• Figure 2: The CorPipe 25 model architecture.