HeRE: Hebrew Relation Extraction benchmark

Data release for HeRE: A Novel Benchmark for Hebrew Relation Extraction via LLM-Guided Denoising of Knowledge Graph Alignments (Ron Keinan, Amir David Nissan Cohen, Reut Tsarfaty; AACL 2026). Code: https://github.com/Ronke21/HeRE.

Every row is a Hebrew Wikipedia passage aligned by distant supervision (CROCODILE) to a Wikidata triple (subject, predicate, object) whose two entities are mentioned in the passage. The label answers one question: does the passage express the stated relation between the subject and the object?

Configs

Config	Split	Rows	Label
gold	validation	500	human, dual-annotated and adjudicated; the set on which the denoising pipeline was selected
gold	test	500	human, dual-annotated and adjudicated; never used for selection (203 unseen-predicate rows, 97 depth rows, 200 silver-drawn rows)
silver	train	1,491,070	agree3: unanimous vote of Gemma-4-31B-it (few-shot), Gemma-3-27B-it (few-shot) and DictaLM-3.0-24B (HebNLI-fine-tuned NLI); 1,080,605 positive / 410,465 negative
silver_all_signals	train	2,564,034	every silver row with the raw score and hard prediction of all ten deployed signals and four consensus rules; agree3_label is -1 where the three models disagree

kg/here_kg_edges.tsv.gz (1,080,605 rows) is HeRE-KG: the positive agree3 edges as subject, predicate, object, docid, n_signals_pos, mean_score, i.e. a Hebrew knowledge graph in which every edge points to the passage (docid, join on silver) that expresses it. Statistics in kg/here_kg_stats.md.

Columns

Gold: item_id, docid, title, text, subject, predicate, object, label, annotator1, annotator2, notes, word_count, selection_reason (validation, or in the test set new_predicate 203, depth_boost 97, silver_validation_200 200). annotator1 and annotator2 are the two annotators' labels (annotator 2 annotated blind); label is the adjudicated label used throughout the paper (identical to annotator1, whose column absorbed the adjudication decisions); notes are annotator 2's free-text remarks. Inter-annotator agreement before adjudication: Cohen's kappa 0.809 (validation), 0.757 (test).

Silver: docid, title, text, subject, predicate, object, label (or agree3_label), then for each signal <family>__<model>__score (float) and <family>__<model>__pred (0/1; -5 or null where the signal abstains) for

Column stem	Strategy	Model
opensource_llm__gemma4_31b	4, few-shot	Gemma-4-31B-it
opensource_llm__gemma3_27b	4, few-shot	Gemma-3-27B-it
opensource_llm__dictalm3	4, few-shot	DictaLM-3.0-24B-Base
finetuned_nli__dictalm24b	2, LoRA NLI	DictaLM-3.0-24B
finetuned_nli__aya32b	2, LoRA NLI (threshold 0.9)	Aya-Expanse-32b
finetuned_nli__gemma2_9b	2, LoRA NLI (threshold 0.9)	Gemma-2-9B
encoder_nli__neodictabert	1, encoder NLI	NeoDictaBERT
encoder_nli__xlmroberta	1, encoder NLI	XLM-RoBERTa-large
cross_rc__neodictabert_k3	3, cross-trained	NeoDictaBERT, K=3
cross_rc__mmbert_k5	3, cross-trained	mmBERT, K=5

Consensus columns: vote_all__pred, vote7__pred, agree4__pred, agree2__pred (-5 = abstain).

Quality

Measured on the held-out gold test set: the best single deployed signal (Gemma-4-31B-it) reaches F1 0.844 (0.944 on predicates seen during development, 0.776 on unseen ones). The published agree3 label covers 66.8% of the test rows and on them reaches precision 0.899, recall 0.934, F1 0.916 (95% CI 0.886 to 0.944). Use the per-signal scores to choose a different precision/recall operating point.

On HeRE-KG, a TransE model trained on the denoised graph beats a size-matched random subset of the raw alignments by 24% relative MRR (three seeds). Every number behind the paper, with confidence intervals, is in the code repository under results/ (https://github.com/Ronke21/HeRE/tree/main/results).

Fine-tuned models

All HeRE resources are grouped in the Hub collection https://huggingface.co/collections/ronke21/here-hebrew-relation-extraction-6aa82bf21db8a9c955cf1cd1. The HebNLI-fine-tuned denoisers behind the released signals and the per-strategy winners:

Model	Role in HeRE	Repo
DictaLM-3.0-24B (Thinking) HebNLI LoRA	Strategy 2; one of the three agree3 models behind the published label	https://huggingface.co/ronke21/hebnli-dictalm-3.0-24b-lora
NeoDictaBERT HebNLI	Strategy 1; deployed signal encoder_nli__neodictabert	https://huggingface.co/ronke21/hebnli-neodictabert
XLM-RoBERTa-large HebNLI	Strategy 1; deployed signal encoder_nli__xlmroberta	https://huggingface.co/ronke21/hebnli-xlm-roberta-large
XLM-RoBERTa-XL HebNLI	Strategy 1 best (F1 0.779)	https://huggingface.co/ronke21/hebnli-xlm-roberta-xl
Qwen3.5-35B-A3B-Base HebNLI LoRA	Strategy 2 best (F1 0.873)	https://huggingface.co/ronke21/hebnli-qwen3.5-35b-a3b-lora
Gemma-4-31B HebNLI LoRA	Strategy 2 (F1 0.849)	https://huggingface.co/ronke21/hebnli-gemma-4-31b-lora
AlephBERT HebNLI	Strategy 1 (F1 0.618)	https://huggingface.co/ronke21/hebnli-alephbert

The Strategy 4 signals (Gemma-4-31B-it, Gemma-3-27B-it, DictaLM-3.0-24B-Base) are prompted public models; prompts are in the code repository.

Contamination note

The 500 test rows were drawn from the silver corpus; all 500 are withheld from every silver and KG file here. Do not add them back when training.

Sources and licenses

Text: Hebrew Wikipedia (early-2025 dump), CC BY-SA 4.0, which this release inherits. Triples: Wikidata, CC0. Alignment: CROCODILE (Huguet Cabot and Navigli, 2021). Text was cleaned of wiki markup; entity fields were normalised.

Citation

Keinan, R., Cohen, A. D. N., and Tsarfaty, R. (2026). HeRE: A Novel Benchmark for Hebrew Relation Extraction via LLM-Guided Denoising of Knowledge Graph Alignments. In Proceedings of AACL 2026.