Bringing ideas to life with a new writing tool for Arabic language learners

Some may question the value of writing well in an age where large language models (LLMs) can churn out nearly perfect essays on any topic in a matter of seconds. But for the developers of ARWI, a new tool designed to help Arabic language learners hone their writing abilities in Modern Standard Arabic (MSA), writing is about more than simply putting one word in front of another.

“Writing develops cognitive skills that are useful and are important for education more broadly,” explains Nizar Habash, professor of computer science and global network professor at NYU Abu Dhabi, consulting professor at MBZUAI, and one of the developers of ARWI. And while people can easily turn to LLMs to help them write today, the texts these systems produce often lack a personal voice. “People will always want to develop ways to stand out with their writing,” Habash says.

Another developer of ARWI, deputy department chair and professor of natural language processing at MBZUAI, Ted Briscoe, hopes that ARWI will be used widely by Arabic learners throughout the region. “It’s a free service for anyone with a browser and an internet connection,” Briscoe says.

Habash, Briscoe and their colleagues won the Diversity Award for ARWI at the recent Workshop on Intelligent and Interactive Writing Assistants (In2Writing) that was co-located with the Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics (NAACL).

Kirill Chirkunov and Chatrine Quaider of MBZUAI and Bashar Alhafini of NYUAD and MBZUAI are co-developers.

Helping language learners

ARWI stands for Arabic Write and Improve. When pronounced, “ARWI” sounds like an Arabic phrase with two poetically connected meanings: “I tell a story,” and “I water/irrigate,” which carries with it the sentiment of bringing an idea to life through cultivation.

The development of ARWI was inspired by a similar system Briscoe built called Write and Improve that was designed to help English language learners prepare for exams. Write and Improve quickly gained thousands of users who spoke a diverse array of native languages.

Similarly, ARWI is a web application that prompts users to compose essays on a variety of topics based on their skill level. The prompts are aligned with the Common European Framework of Reference for Languages (CEFR) skill levels, a widely used standard for language proficiency. Aligning the prompts to these skill levels helps learners prepare for exams that are benchmarked to these requirements.

The system features an Arabic text editor and detects errors and evaluates the quality of text composed in the app. It also tracks learner progress, providing users with personalized feedback that can help them improve over time.

Instead of immediately flagging errors, ARWI provides suggestions to users about corrections that need to be made when they submit their essays. This helps them learn from their mistakes, instead of automatically identifying and fixing errors.

While ARWI isn’t the first tool designed to help Arabic language learners write better, other systems don’t assess overall writing quality, and they aren’t able to identify nuanced errors that learners often make when writing in the language.

And while many students have turned to LLMs in recent years to help with writing, LLMs don’t identify errors in a way that promotes language learning. LLMs produce paraphrases of what users have written instead of making minimal corrections, often introducing subtle and unwanted changes in meaning, Briscoe explains.

Writing for a wide audience

The team developed ARWI to address a gap in writing tools for people learning to write MSA, which is used in the news media, academic and literary writing, and in government. People don’t typically speak MSA but interact with it through reading and writing.

Because Arabic dialects vary widely, MSA plays a key role in enabling communication across the Arabic-speaking world. One major difference between MSA and the dialects relates to the case system used in MSA, which is much more complex than it is in the dialects. “Some of these dialects are as different from MSA as French is from Latin,” Habash says.

ARWI could be helpful for a blogger trying to improve their writing in MSA so that they can reach a wide audience. “We often think of the best quality writer as one who writes well in MSA and who can write for the largest number of people,” Habash says.

Learning from mistakes

Since ARWI collects information about the kinds of mistakes users make, its data can show how mistakes vary by native speakers of different dialects and languages.

When beginning to use ARWI, users enter their native language or the Arabic dialect they’re familiar with and their proficiency level. This allows for more “targeted prompting and feedback” from the system.  

Studies by researchers who work on language education have found that if educators know the native language of learners who are studying a new language, the educators can provide better assessments of a learner’s ability. This is because features of the native language will seep into the language they are learning. The same is true for features of dialects and MSA. “Knowing the speakers’ dialects could help us provide better feedback and guidance for writing,” Habash says.

Continuing to cultivate language learning

Habash, Briscoe and their colleagues had students at NYUAD who were learning Arabic use ARWI and asked them to provide feedback about their experiences with the system. Overall, the students, who were all non-native speakers of Arabic, viewed it positively. The developers intend to use this feedback and other information they collect from users in the future to further improve the system.

While the main goal of ARWI is to help learners write better, the researchers note that the rich and diverse body of annotated Arabic texts that ARWI collects could also help with fine-tuning LLMs for Arabic. These systems require a huge amount of data before they become fully proficient in a language and studies have shown that LLMs don’t perform as well in Arabic as they do in so-called high-resource languages like English. The researchers plan to make the dataset collected by ARWI public — after removing any personal information — so that other developers can use it.

The researchers emphasize that ARWI is more than a web app — it’s also a platform with underlying infrastructure that can support additional capabilities. For example, its ability to evaluate writing samples could be used to provide language certificates that are needed for jobs that require proficiency in Arabic.

The team plans to add more functionality to the system in the future. “We want to make the system friendly and emphasize the positive aspects of writing and not just the ability to correct mistakes,” Briscoe says.

In helping learners find their voice in MSA, ARWI is a reminder that even in the age of LLMs, good writing still matters, not just for passing exams, but for thinking clearly, communicating precisely, and connecting across cultures. “Making us better writers makes us better readers, which also makes us better evaluators of the text that LLMs produce,” Habash says.