Making sense of Arabic machine translation

Shutterstock/LightField Studios

• Machine translation is difficult
• Three forms and many dialects
• Translation services worth $6.6bn

Some estimates indicate the Arabic language consists of more than 12 million words. While the exact figure is much debated by academics, to put it into perspective the Oxford English Dictionary contains around 273,000.

Arabic “has dozens of dialects and millions of unique words, and there are many ways to say the same thing,” says Nizar Habash, professor of computer science at New York University Abu Dhabi.

As a result, Habash believes Arabic is a tough language for computers to figure out, especially when it comes to machine translation.

• For Gulf businesses, chatbots are a difficult conversation
• Reality still beats deepfake technology … for now
• Demand for data centres accelerates across Mena

Machine translation, the process of automatically translating text from one language to another without human involvement, gained prominence in the 1990s and has evolved ever since.

While strides have been made over the last three decades, Arabic remains a major challenge for machine translation developers to crack, experts say. 

The motivation is the potential size of the opportunity. Estimates vary, but one figure given is that 375 million speak Arabic globally.

Between 2017 and 2022, the market size for translation services worldwide increased by 7 percent a year on average, reaching $6.6 billion in 2022, according to US research firm IBISWorld.

In the UAE, the rise in international business and legal disputes has increased the demand for translation services, Dubai translation and interpretation provider Prime Legal Translation said in a blog post.

Another translation services platform, Aburuf Legal, says the Gulf state’s hosting of international events and its growing tourism sector has necessitated the demand for translation in various fields, such as legal, medical and technical sectors.

Why is it difficult to translate Arabic?

One of the challenges is that the language has three forms: Classical Arabic, which is common in literature and writing such as the holy book Quran; Modern Standard Arabic, the official written language; and local dialects – informal, communicative day-to-day language that varies from region to region and city to city.

“We learn language through immersion and repetition, constantly absorbing and picking up linguistic cues from our social environment,” Habash said in a research paper.

For computers to produce accurate results, computational linguists need to input more information about the language, research by Translation Journal, a digital online journal for translators, has found.

Arabic includes terms that lack direct equivalents in English, adding another layer of challenge to translation. Such words are usually culture-bound, the journal says.

Examples include suhoor, a meal eaten before dawn for fasting; aqiqah, a goat slaughtered and its meat distributed to the poor on the occasion of having a new baby; and salat al-istisqa, a prayer performed during times of drought or when in need of rain. 

According to NYU Abu Dhabi, Arabic verbs have up to 5,400 conjugations. This, coupled with the lexical differences – the degree to which a pair of languages’ word sets differ – and the absence of standard spelling rules in Arabic dialects, hinders computational linguists from achieving perfection in translation.

Progressive improvements

However, Avneesh Prakash, co-founder and CEO of Dubai-based AI translation platform Camb.AI believes improvements in Arabic translation have been made over the past decade.

“From low-quality text-to-text literal translations to AI models that can perform voice-to-voice colloquial and contextual dubbing while retaining voices and emotions, we’ve come a long way,” he tells AGBI.

“It used to take weeks, sometimes months, to get an hour-long video to be dubbed in multiple languages. Today, it can be done in minutes.”

Ahmed Mahmoud, founder and CEO of Egyptian AI startup DXwand explains that the adoption of neural machine translation (NMT), such as Google Translate, has led to dramatically improved fluency and accuracy compared with older phrase-based models. 

“NMT engines are now being trained on specialised datasets related to specific fields like healthcare, finance or legal documents, resulting in more accurate translations,” Mahmoud says.

Organisations are also designing wearable devices for real-time translation and interpretation.

US-based consumer and electronics manufacturer Humane last year introduced an AI pin, a device that can be clipped to clothes. It responds to touch, voice and hand gestures, with the primary aim to translate information in real-time.

Although there are some improvements, translation of dialectal Arabic, as on social media, or cultural artefacts such as dialectal songs and novels, remains underdeveloped, Habash says.

Experts say there is a need to develop and train AI models on massive datasets specific to Arabic dialects for real progress in machine translation to be made.

“Addressing potential bias in AI training data and translation algorithms is crucial for achieving culturally sensitive and inclusive translations,” Mahmoud adds.