Researchers have introduced HeRE, a benchmark designed to evaluate relation extraction (RE) in Hebrew. The benchmark is constructed using LLM-guided denoising of knowledge graph alignments, a method that refines noisy alignments to produce a reliable test set. This addresses a clear gap: while LLMs show strong RE performance in English, standardized benchmarks for low-resource languages are scarce.

The authors note that the absence of such benchmarks makes it difficult to assess how well LLMs generalize beyond English. By providing a Hebrew-specific resource, HeRE enables more systematic evaluation of RE capabilities in a low-resource setting. The approach also demonstrates a practical way to create benchmarks using LLMs themselves to clean and structure data.

As a single-source announcement, the abstract does not compare HeRE to other benchmarks or discuss limitations in detail. The significance lies in its contribution to multilingual evaluation, offering a foundation for future work on Hebrew NLP and low-resource RE tasks.