ziaddddd/arabic-legal-embedding-dataset-nli

Arabic Legal Embedding Dataset (NLI & Semantic Similarity)

Overview

The Arabic Legal Embedding Dataset is a high-quality dataset designed for semantic similarity, Natural Language Inference (NLI), and sentence embedding tasks in Arabic. It focuses on legal, administrative, and general-purpose text, providing a rich set of sentence pairs annotated with semantic relationships. This dataset enables building robust AI models capable of understanding nuanced meanings in Arabic legal and administrative contexts.

Key Features

• Semantic Annotations: Each sentence pair includes a relation indicating its semantic type:

  • summary – concise summary of the first sentence
  • synonym – paraphrases or semantically equivalent sentences
  • paraphrase – sentences reworded differently but retaining the same meaning
  • contradiction – sentences with opposing meanings
  • weak – loosely related or contextually weak sentences
  • hard_negative – sentences appearing similar but semantically unrelated
  • entailment – one sentence logically follows from the other
  • passive – rephrasing in passive voice

• Similarity Scores: Each pair is scored between 0 and 1, quantifying semantic similarity.

• Legal & Administrative Focus: Ideal for training embeddings, NLI models, or semantic search systems for Arabic legal documents, contracts, administrative texts, and general-purpose applications.

Dataset Structure

Field	Type	Description
sentence_a	str	First sentence
sentence_b	str	Second sentence
score	float	Semantic similarity score (0–1)
relation	str	Type of relation (summary, synonym, paraphrase, contradiction, weak, hard_negative, entailment, passive)

Example Entries

{
  "sentence_a": "يرجى التوجه إلى كاتب الجلسة لوضع توقيعك في السجل الرسمي الذي يثبت حضورك للدفاع اليوم.",
  "sentence_b": "يجب التوقيع على إثبات حضور الجلسة في المحضر الرسمي لضمان الحقوق الإجرائية.",
  "score": 0.93,
  "relation": "summary"
}
{
  "sentence_a": "إحنا محتاجين مترجم محلف عشان نترجم المستندات دي.",
  "sentence_b": "إحنا محتاجين خبير فني يترجم الرموز اللي موجودة في كشوف الحسابات.",
  "score": 0.58,
  "relation": "hard_negative"
}
{
  "sentence_a": "الجلسة كانت زحمة جداً وما قدرناش نتكلم غير كلمتين.",
  "sentence_b": "شُهد ازدحام شديد في الجلسة ولم يُسمح إلا بحديث مقتضب.",
  "score": 0.85,
  "relation": "passive"
}
{
  "sentence_a": "عم نحاول نتواصل مع المندوب بس تلفونه مغلق.",
  "sentence_b": "هناك صعوبة مؤقتة في الوصول لجهة التوصيل المسؤولة.",
  "score": 0.82,
  "relation": "entailment"
}
{
  "sentence_a": "عم نحاول نتواصل مع المندوب بس تلفونه مغلق.",
  "sentence_b": "هناك صعوبة مؤقتة في الوصول لجهة التواصل المسؤولة.",
  "score": 0.82,
  "relation": "paraphrase"
}