New benchmark HeRE boosts Hebrew relation extraction with LLM denoising

Researchers have developed a novel methodology to create a large-scale benchmark for Hebrew Relation Extraction (RE) using LLM-guided denoising of knowledge graph alignments. This approach addresses the lack of standardized benchmarks for low-resource languages by creating datasets that are less prone to English-centric biases. The study found that multilingual models performed better in few-shot settings, while Hebrew-specific encoders were competitive with fine-tuning. The resulting benchmark, HeRE, contains 1.5 million instances and aims to support data-intensive experimentation and model fine-tuning for Hebrew RE. AI

IMPACT Enables more robust evaluation of LLMs in low-resource languages, potentially accelerating development for non-English NLP tasks.

RANK_REASON The cluster describes a novel benchmark and methodology for relation extraction in a low-resource language, presented in an academic paper. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Gemma 4-31B-it
• GPT-5.4
• Hebrew
• Hebrew Relation Extraction
• HeRE
• Hugging Face
• Knowledge Graph Alignments

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →