Amit5674/hebrew-multilabel-correctness-error-detection-evidence-to-claim

Hebrew Multi-Label Error Type Classifier (Evidence-Based)

Model Description

Fine-tuned dicta-il/neodictabert for multi-label error type classification in Hebrew. Identifies specific factual error types in corrupted claims using evidence sentences extracted from source articles via semantic similarity.

Task: Multi-label error type classification (15-20 error types)
Language: Hebrew
Max Context: 512 tokens (evidence + claim)
Granularity: Sentence-level with article context

Key Features

• Evidence Extraction: Sentence-transformer (multilingual-MiniLM-L6-v2) for semantic similarity
• Similarity Threshold: 0.5 (minimum for evidence inclusion)
• Top-K Evidence: 2-3 most similar sentences per claim
• Error Types: Covers Hebrew-specific morphological errors (verb_gender_swap, hebrew_root_pattern_confusion) and cross-language errors (entity_swap, sentence_negation, temporal)

Architecture

• Base Model: dicta-il/neodictabert
• Classification Head: Multi-label (sigmoid activation per error type)
• Input Format: [CLS] evidence_sentences [SEP] corrupted_claim [SEP]
• Output: Binary vector (length = number of error types) with independent sigmoid probabilities

Training Configuration

• Learning Rate: 2e-5
• Epochs: 2.0
• Batch Size: 4 per device (effective: 32 with gradient accumulation)
• Max Sequence Length: 512 tokens
• Learning Rate Scheduler: Cosine
• Class Balancing: BCEWithLogitsLoss with pos_weight + WeightedRandomSampler

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification import torch from sentence_transformers import SentenceTransformer import numpy as np

Load model

model_name = "Amit5674/hebrew-multilabel-evidence-correctness-error-detection" tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True) model = AutoModelForSequenceClassification.from_pretrained( model_name, trust_remote_code=True, problem_type="multi_label_classification" ) model.eval()

Load evidence extractor

evidence_model = SentenceTransformer('sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2')

Extract evidence from article

def extract_evidence(article: str, claim: str, top_k: int = 3, threshold: float = 0.5): """Extract top-k most similar sentences from article to the claim""" article_sentences = article.split('.') claim_embedding = evidence_model.encode(claim) article_embeddings = evidence_model.encode(article_sentences)

similarities = np.dot(article_embeddings, claim_embedding) / (
    np.linalg.norm(article_embeddings, axis=1) * np.linalg.norm(claim_embedding)
)

top_indices = np.argsort(similarities)[-top_k:][::-1]
evidence_sentences = [article_sentences[i] for i in top_indices if similarities[i] >= threshold]
return ' '.join(evidence_sentences)

Example usage

article = "ישראל התחילה בהרעשה רגע אחרי הפסקת האש. הממשלה הודיעה על צעדים חדשים." corrupted_claim = "ישראל התחילה להתרגש רגע אחרי הפסקת האש"

Extract evidence

evidence = extract_evidence(article, corrupted_claim, top_k=3, threshold=0.5)

Tokenize and predict

inputs = tokenizer( evidence, corrupted_claim, return_tensors="pt", padding="max_length", max_length=512, truncation=True )

with torch.no_grad(): outputs = model(**inputs) logits = outputs.logits[0] probs = torch.sigmoid(logits)

# Get predicted error types (threshold = 0.5)
predicted_indices = (probs > 0.5).nonzero(as_tuple=True)[0].tolist()
predicted_labels = [model.config.id2label[i] for i in predicted_indices]

# Get all scores
all_scores = {
    model.config.id2label[i]: float(probs[i].item())
    for i in range(model.config.num_labels)
}

print(f"Predicted error types: {predicted_labels}") print(f"All scores: {all_scores}")

Input Format

• Premise: Original correct claim + top 2-3 evidence sentences from article (concatenated)
• Hypothesis: Corrupted claim with factual errors
• Evidence Extraction: Automatic via sentence-transformer similarity matching

Output Format

• Predicted Labels: List of error type strings (e.g., ["hebrew_root_pattern_confusion", "verb_gender_swap"])
• All Scores: Dictionary mapping error type → probability (0.0 to 1.0)
• Threshold: 0.5 (configurable)

Supported Error Types

Hebrew-Specific:

• hebrew_root_pattern_confusion, verb_gender_swap, noun_gender_swap, morphological_connective_confusion, construct_state_confusion, specificity_shift_errors, verb_tense_swap

Cross-Language:

• entity_person_swap, entity_location_swap, entity_organization_swap, sentence_negation, number_swap, measure_unit_swap, temporal

Limitations

• Max sequence length: 512 tokens (evidence + claim)
• Evidence extraction depends on sentence-transformer quality
• Performance may degrade if evidence sentences are not semantically relevant
• Class imbalance: rare error types may have lower recall

Citation

@misc{hebrew_evidence_multilabel_classifier, title={Hebrew Multi-Label Error Type Classifier with Evidence-Based Context}, author={Your Name}, year={2025}, publisher={Hugging Face} }