ImruQays/16-million-raw-arabic-words

Dataset Description

This dataset contains a collection of 16,052,878 unique Arabic words. These words were extracted from a large corpus of Arabic text originating from two primary sources: the Shamela library and the Hindawi library.

Key Characteristics:

• Unique Words: The dataset is focused on uniqueness. Each entry in the dataset represents a distinct Arabic word, and duplicates have been removed.
• Diacritic Sensitivity: Words with different diacritical markings are considered unique. For example, "كَتَبَ" and "كُتِبَ" are treated as separate words. This is an important feature for tasks involving diacritic-aware Arabic Natural Language Processing (NLP).
• Uncleaned Text: The words in this dataset are not fully cleaned. This means they may contain:
  • Punctuation marks: Attached to words (e.g., commas, periods, question marks).
  • Symbols: Various symbols that appeared in the original text.
  • Numbers: Numbers that were identified as separate tokens.
  • Other non-alphanumeric characters: Any other characters that were not specifically removed during the extraction process and were considered as part of a token by the word segmentation process (likely whitespace splitting).
• Line-Separated: The dataset is structured as a simple text file, with one unique word per line. This makes it easy to process and integrate into various NLP pipelines and tools.

Considerations and Limitations

• Uncleaned Nature: Users should be aware that the dataset is not cleaned and may contain noise in the form of punctuation, symbols, and other non-word characters. Depending on the intended use, some level of data cleaning or preprocessing may be necessary.
• Diacritic Variation: While diacritic sensitivity is a feature, it also means that morphologically related words with different diacritics are considered distinct. For some applications, you might need to consider normalizing or removing diacritics.
• Tokenization Method: The method used for word segmentation (likely whitespace splitting) might not be perfect and could lead to some edge cases in word tokenization.
• Source Domain: The vocabulary is heavily influenced by the domains covered in the Shamela and Hindawi libraries (Literature, Islamic studies, general knowledge). The dataset's vocabulary might be less representative of other domains like modern slang, social media language, or technical domains not well-represented in these libraries.