Arabic Speakers Deserve High-Quality Machine Translation

Not only that, but today’s sophisticated MT engines actually have roots in Arabic, building on the legacy of 9th-century Arab cryptographer Abu Yusuf al-Kindi (see “Throughlines of Genius” from MultiLingual magazine’s May 2024 issue). Three techniques that al-Kindi identified serve as the basis for modern MT:

• The quantitative technique, which compares the frequency of plaintext and ciphertext letters.
• The qualitative technique, which analyzes which letters can or cannot be combined and determines the proper order of letter sequences.
• The “probable word” technique, which leverages the predictability of language to fill in missing information with likely word choices.

Challenges of Arabic MT

Despite the historical link between MT and Arabic, in the 21st century, MT tools are underperforming for speakers of al-Kindi’s language. Several linguistic features of Arabic pose challenges for modern MT engines.

One factor is intra-linguistic diversity, referred to as “diglossia:” a phenomenon whereby a language has distinct varieties used in different contexts within the same community. In addition to the formal Modern Standard Arabic variant, there are at least 22 distinct dialects. These include, but are not limited to, the dialects of the Levant, the Persian Gulf, and Northern Africa. While diglossia enriches the collective linguistic repertoire of the Arabic-speaking world, it is a pain point in translation. By virtue of its very nature, professional translation transfers content not only cross-linguistically, but also cross-culturally; therefore, understanding the linguistic and cultural nuances of Arabic-speaking locales is essential to accurate translation.

Another translation challenge that poses itself when working with Arabic is the polysemy of words — where a word can bear multiple meanings or senses. For instance, as a native speaker of Jordanian Arabic, I grew up saying “Ya’ateek il afiyah,” an expression that literally translates into “May [God] give you good health.” It is one of many possible ways to express gratitude or even greet someone. In this phrase, afiyah means health, but the word does not have the same meaning in all Arabic-speaking locales; for example, afiyah means “fire” in the Moroccan (Darija) dialect. So, unless you want to see someone set on fire, you would want to use caution when using this expression in Morocco. This is but one example of how polysemy might cause translation challenges for human linguists and MT engines alike.

Arabic also features inconsistent term usage across dialects, which undermines the “probable word” cryptanalysis technique. In terminology management — the practice and study of collecting, defining, standardizing, translating, localizing, and maintaining terms for a domain-specific use in a terminology database or glossary — a term designates one and only one concept (1:1 ratio). For example, “target language” is a translation studies term that refers to the language into which a source text is being translated. Through consistent use, it has been standardized to mean exactly that, facilitating understanding among speakers in that specific domain.

Alas, the situation is more complex in Arabic. Consider the English term “Authority” (in the sense of an official governmental body, such as the “Federal Labor Relations Authority”), which has the following Arabic equivalents in three different locales:   

• Egypt: Maslaha
• Kuwait: Haya’ah
• United Arab Emirates: Jihaz

While these Arabic terms are all semantically acceptable translations of the English term “Authority,” they are not interchangeable across Arabic-speaking locales. This is an already challenging undertaking for human linguists, let alone MT engines. Prior to being trained on nuanced datasets, MT engines often mistranslated such terms — and in many cases, they still do. The irregularity in the usage of terms across Arabic dialects breaks down the predictability of language, leading to MT outputs that lack accuracy, consistency, and cultural sensitivity — all of which are hallmarks of successful translations that meet the specific needs of target readers.