https://doi.org/10.5061/dryad.9w0vt4brx

Description of the data and file structure

The AAFAQ Dataset was collected and validated through experimental efforts focused on fine-tuning Arabic NLP models, such as AraBERT, for multi-label question classification. Additional experiments included integrating the dataset with generative answering systems like Alpaca + Gemma-9B Unsloth to enhance metrics for multi-attribute classification. The AAFAQ Dataset is a rich and comprehensive Arabic dataset designed for semantic and cognitive question classification in Modern Standard Arabic (MSA). The dataset consists of 5,009 records annotated with a variety of attributes, including Question Tool, Intent, Answer Type, Cognitive Level, and Temporal Context, among others. It serves as a benchmark resource for research in Arabic NLP, advancing fields such as education, cognitive research, and multilingual AI systems.

The name AAFAQ (Arabic: آفَاق) means "horizons," symbolizing the goal of expanding understanding and pushing the boundaries of Arabic Question Answering Systems (QAS) toward deeper comprehension of the language's complexities.

Files Included

AAFAQ_Dataset_utf_formatted.csv: The primary dataset file in CSV format, containing 5,009 rows and 15 columns.

• Columns:
  1. QuestionID: Unique numerical identifier for each question.
  2. QuestionText: The text of the question.
  3. QuestionTool: The interrogative tool or word used in the question (e.g., "What," "Why").
  4. QuestionToolType: Type of interrogative tool (e.g., imperative, implicit).
  5. QuestionType: Categorization as factoid or non-factoid.
  6. List: Indicates whether the expected answer is a list.
  7. AnswerType: The type of answer expected (e.g., number, description).
  8. Intent: Purpose or intent of the question (e.g., informational, explanatory).
  9. CognitiveLevel: Cognitive skill level required to answer the question.
  10. Category: The domain or field of the question (e.g., health, education).
  11. Subjectivity: Whether the question is subjective or objective.
  12. TemporalContext: Time frame referenced in the question (e.g., past, present).
  13. PurposeContext: Goal of the question (e.g., problem-solving, decision-making).
  14. AnswerSourceText: Source text from which the answer is derived.
  15. Answer: The actual answer to the question.

AAFAQ_Dataset_utf_formatted_Translated_UTF8_BOM_1.csv is the English translated version. However, please note that the following fields were not translated: QuestionText, QuestionParticle, and Answer. These fields contain linguistically sensitive content, and translating them may result in a loss of meaning or misrepresentation of the original context. To preserve the integrity and accuracy of the data, we have kept it in the original Arabic.

Dataset Features

• Size: Approximately 2.82 MB
• Format: CSV
• Annotations: Rich manual annotations validated with high inter-annotator agreement (85%).
• Coverage: Semantic, cognitive, and contextual intricacies of Arabic questions

Code/software Requirements

• Python (Version 3.7 or higher)
• Libraries: Pandas, NumPy, Matplotlib (for data exploration and visualization).
• NLP Frameworks:
  • AraBERT for fine-tuning language models.
  • Hugging Face Transformers for model integration.