NLP for Arabic – The case of lemmatization

Arabic is a complex language for NLP tasks, even for simple ones like lemmatization.

There are several reasons for this:

Arabic creates words based on roots: for example, the word کتاب (kitab, “book”) is derived from ك ت ب (k t b). Many related words are derived from the same root.

Arabic can create words, similar to “compounds” but more limited, combining certain POSes, such as prepositions, conjunctions and pronouns, with nouns, adjectives and verbs. For example, وكتابي (wakitabi, “and my book”) is composed of و (wa, “and”) + كتاب + ي (i, “my”).

More often than not, Arabic speakers omit vowels when writing, which makes hard to determine the real lemma of the words.
Finally, Arabic is written in one canonical form across countries (MSA, Modern Standard Arabic) and has different variants, both in those same countries (Egyptian, Najdi…) and across countries (Gulf Arabic is used Kuwait, UAE, Qatar…); additionally, some regional dialects like Egyptian are known across most of the Arabic-speaking world due to their widespread use in media. Furthermore, there are different registers involved: Classical Arabic, used for old texts and reciting the Qur’an; MSA, used for writing, broadcasting or interviewing; the colloquial regional dialect, which is the everyday language used in informal contexts, and more.

As a result of this complexity, developing tools to do NLP in Arabic requires, and in particular lemmatization:

A good data source, comprehensive and accurate, that tags morphological attributes and also language variants. For example, the plural of سيارة (“car”) is different in MSA (سيارات) and in Gulf Arabic (سيايير).
A good data architecture that integrates different information sources: prefixes, suffixes, roots, forms…
A very efficient processing software designed to handle millions of different potential tokens that can be generated just in MSA, for example.