Researchers release 1.5-million-instance Hebrew relation extraction dataset

Researchers released HeRE, a 1.5-million-instance dataset for training and evaluating systems that identify relationships in Hebrew text. Their preprint describes a method that filters noisy matches between knowledge bases and texts by testing whether each text supports a proposed relationship. The researchers assessed five filtering strategies. The best open-weights model, Gemma-4-31B-it, scored F1 = 0.915 against GPT-5.4's 0.921. On a fully held-out test set annotated twice, it reached F1 = 0.844. HeRE uses automatically derived labels, a distinction from the separately annotated test set.