GLiNER2: An Efficient Multi-Task Information Extraction System with Schema-Driven Interface

  • 2025-07-24 16:11:14
  • Urchade Zaratiana, Gil Pasternak, Oliver Boyd, George Hurn-Maloney, Ash Lewis
  • 0

Abstract

Information extraction (IE) is fundamental to numerous NLP applications, yetexisting solutions often require specialized models for different tasks or relyon computationally expensive large language models. We present GLiNER2, aunified framework that enhances the original GLiNER architecture to supportnamed entity recognition, text classification, and hierarchical structured dataextraction within a single efficient model. Built pretrained transformerencoder architecture, GLiNER2 maintains CPU efficiency and compact size whileintroducing multi-task composition through an intuitive schema-based interface.Our experiments demonstrate competitive performance across extraction andclassification tasks with substantial improvements in deployment accessibilitycompared to LLM-based alternatives. We release GLiNER2 as an open-sourcepip-installable library with pre-trained models and documentation athttps://github.com/fastino-ai/GLiNER2.

 

Quick Read (beta)

loading the full paper ...