README.md

metadata

license: apache-2.0
base_model:
  - dicta-il/neodictabert

---

license: apache-2.0 language: - he base_model: - dicta-il/neodictabert tags: - multi-label-classification - hebrew - fact-checking - error-detection pipeline_tag: text-classification library_name: transformers

Hebrew Multi-Label Error Type Classifier (Setup 2: Claim-Level)

Model Description

Fine-tuned dicta-il/neodictabert for multi-label classification of factual error types in Hebrew text at the claim/sentence level.

Task: Identify specific error types in corrupted Hebrew claims
Language: Hebrew
Max Context: 3,072 tokens
Granularity: Sentence-level (individual claims)

Input/Output

Input

• Premise: Original correct claim (Hebrew sentence)
• Hypothesis: Summary claim (Hebrew sentence)

Output

Multi-label classification with probability scores for each error type. Threshold: 0.5 (probabilities ≥ 0.5 indicate presence of that error type).

Supported Error Types

The model detects both cross-language and Hebrew-specific error types:

Cross-Language Errors

• entity_person_swap - Swapping person names
• entity_location_swap - Swapping place names
• entity_organization_swap - Swapping organization names
• entity_date_swap - Changing dates
• number_swap - Altering numerical values
• measure_unit_swap - Changing units of measurement
• sentence_negation - Adding/removing negation

Hebrew-Specific Errors

• hebrew_root_pattern_confusion - Morphological root/binyan errors
• morphological_connective_confusion - Logical connector errors (שכן ↔ כך ש)
• verb_gender_swap - Verb gender disagreement
• noun_gender_swap - Noun/pronoun gender errors
• homographic_gender_errors - Gender mismatches in homographic words (מורה, מרצה)
• specificity_shift_errors - Definite article changes (ה-)
• construct_state_confusion - Smichut (סמיכות) errors
• subject_verb_person_mismatch - Person agreement errors
• verb_tense_swap - Tense inconsistencies
• evidentiality_source_attribution_collapse - Attribution removal
• impersonal_to_personal_verb_errors - Voice changes (passive ↔ active)
• verb_template_agency_reversal - Agency reversals (הלביש ↔ התלבש)
• directional_preposition_swap - Directional errors (מ-X ל-Y ↔ מ-Y ל-X)