Lightweight Multilingual Coreference Resolution without LLMs @CRAC2026

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

Summary

A new multilingual coreference system, developed for the CRAC 2026 unconstrained track, introduces a unified, single-model architecture based on Conditional Random Fields (CRFs). This system supports 20 languages and notably achieves multilingual resolution without relying on large language models (LLMs) or pretrained weights. Unlike resource-intensive neural methods, the proposed model is highly efficient, designed for deployment on standard CPU hardware. It leverages linguistic and contextual features to accurately capture coreference relations across languages with diverse syntactic and morphological properties. Training utilized official data distributions from the CRAC 2026 shared task. This methodology offers a robust and scalable solution for multilingual Natural Language Processing, proving particularly useful in resource-constrained environments. The results confirm that feature-driven structured models remain effective for complex cross-lingual tasks, with consistent performance between development and test data.

Key takeaway

For NLP Engineers building multilingual systems in resource-constrained environments, this research suggests re-evaluating traditional feature-driven models. You should consider Conditional Random Fields (CRFs) as a viable alternative to large language models for coreference resolution across 20 languages. This approach enables efficient deployment on standard CPUs, significantly reducing hardware requirements and operational costs. Explore integrating linguistic and contextual features into structured models to achieve robust, scalable multilingual NLP solutions without the overhead of LLMs.

Key insights

A CRF-based system offers efficient multilingual coreference resolution for 20 languages without LLMs or pretrained weights.

Principles

• Feature-driven structured models remain effective for complex cross-lingual tasks.
• Efficient multilingual NLP is achievable without LLMs or pretrained weights.

Method

A unified, single-model architecture based on Conditional Random Fields (CRFs) uses linguistic and contextual features, trained on CRAC 2026 shared task data.

In practice

• Deploy coreference resolution on standard CPU hardware.
• Implement robust multilingual NLP in resource-constrained settings.

Topics

• Coreference Resolution
• Multilingual NLP
• Conditional Random Fields
• Resource-Constrained AI
• CPU Deployment

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

Related on AIssential

• 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.
• 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.
• Approximate Structured Diffusion for Sequence Labelling — Combining discrete diffusion with a Mean-Field approximated CRF denoiser improves sequence labelling accuracy and scalability.
• Multilingual and Multimodal LLMs in the Wild: Building for Low-Resource Languages — Multimodal LLMs must integrate text, speech, and vision for low-resource languages, moving beyond English-centric benchmarks.
• Macro: Enhancing Multilingual Counterfactual Explanations through Alignment-as-Preference Optimization — Preference alignment, specifically DPO, effectively balances validity and minimality in multilingual counterfactual explanation generation for LLMs.
• Cohere Launches Tiny Multilingual Open Weight Model — Tiny Aya offers open-weight, multilingual AI models to enhance linguistic diversity and edge device accessibility.

Open in AIssential →

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