CHALIS - Challenging Language Identification Samples

CHALIS is a multilabel dataset composed of several sections aimed at testing language identification abilities in difficult scenarios.

Main contribution comes in the form of gathering and classifying by a human expert of sentences belonging to four language pairs (Spanish - Catalan, Portuguese - Galician, Danish - Norwegian, Czech - Slovak) These were sorted into two categories: joint for sentences which belong to both languages and single which belong to only one of the languages.

Sections

The dataset has the four following sections (with further subsections in case of transiteration):

• non-language - The non-language data such as emails, ISBNs, etc.
• single - The sentences which were assigned only to one language during dataset gathering. This creates a fairly challenging dataset for recognizing the less prevalent language of the pair.
• joint - The sentences which were assigned to both cousin languages during dataset gathering.
• transliteration
  • czech - Transliteration of Russian into the Czech alphabet.
  • cyrilic - Transliteration of Cyrilic written sentences into latin characters.
  • arabic - Transliteration of Kazakh language written in Cyrilic into Arabic.
  • latin - Transliteration of diacritic heavy language sentences into purely latin characters.
  • antspeak - Transliteration of English sentences into antspeak.
  • leet - Transliteration of English sentences into leetspeak.
  • random - Substitution of random letters in English, Czech, and Slovak sentences for ones from Russian and Greek alphabets which preserve the human readable renderings of the sentences.