Scaling Arabic NLP Research at Cairo University with Theta EdgeCloud

This is the first in a series of case studies looking at how academic and enterprise teams are using Theta EdgeCloud in their work. Each piece will showcase a single research group or organisation, the problems they were trying to solve, and how access to distributed GPU infrastructure has shaped what they have been able to build.

Arabic is one of the most widely spoken languages in the world, yet it remains comparatively underserved by natural language processing (NLP), the field of AI concerned with how machines read, interpret, and generate human language.

Part of the reason is technical because Arabic morphology is complex, with words derived from three-letter roots that can produce hundreds of related forms, and meaning often hinges on structural cues that simpler approaches miss.

Part of the reason is to do with infrastructure because building serious language AI systems requires significant compute, and researchers working on under-resourced languages frequently find themselves waiting in queues for shared graphics processing units (GPUs) while their counterparts at the most well-funded labs train freely.

A research lab at Cairo University has been working against both constraints, and recently published a paper showing how they did it.

The SummARai Project

SummARai is a web-based summarisation system designed to take Arabic documents and produce summaries that are concise and faithful to the source. The team’s approach combines three components in a hybrid pipeline.

The first stage uses TextRank, a graph-based algorithm that identifies the most important sentences in a document by mapping how they relate to one another. This is an extractive method, meaning it pulls existing sentences out of the source rather than generating new ones.

The second stage uses Text-to-Text Transformers for Arabic Language Generation (AraT5), a transformer model adapted for Arabic that the team fine-tuned on their own datasets.

Transformers are the family of AI models behind systems like ChatGPT, and they are particularly well suited to generating fluent natural language. This stage is abstractive, meaning the model generates new sentences that capture the meaning of the source rather than simply quoting from it.

The third stage uses a large language model (LLM) to smooth the output, improving fluency and making the final summary read more naturally.

The system was evaluated using Bidirectional Encoder Representations from Transformers Score (BERTScore), a metric used in NLP research to measure how closely generated text matches reference summaries. Rather than checking for exact word matches, BERTScore compares meaning at a semantic level, which makes it a more useful measure for systems that paraphrase. SummARai scored 70.71% on BERTScore F1, a competitive result for Arabic summarisation and one that confirms the hybrid approach works in practice.

The team published the work in a peer-reviewed paper with the Institute of Electrical and Electronics Engineers (IEEE, one of the leading publishers of computer science research, where they acknowledged Theta Labs for providing the compute infrastructure that made the experiments possible.

Beyond SummARai

The lab is now applying similar workflows to a second project called InspaceAI, a platform-agnostic user interface testing tool built around vision-based autonomous agents. Most UI automation tools work by inspecting the underlying code of a web page or app to find buttons and form fields.

InspaceAI’s agents take a different approach, interpreting interfaces visually and navigating applications the way a human tester would, by looking at the screen. The workload is inference-heavy and latency-sensitive, which makes the flexibility of distributed GPU access particularly valuable.

According to the team at Cairo’s Faculty of Computers and Artificial Intelligence, Theta EdgeCloud made a big difference to their research:

“It enabled us to scale experiments beyond what was possible locally, especially when fine-tuning transformer models like AraT5 on large Arabic datasets. We were able to run multiple experiments in parallel, reducing turnaround time and accelerating iteration across different pipeline components such as extractive methods, chunking strategies, and smoothing techniques.”

For researchers working on under-resourced languages or applied AI problems outside the largest labs, the lesson from Cairo University’s work is that the bottleneck is often less about ideas than about access. When access opens up, the research follows.

If you are a university or research team looking for access to affordable compute via Theta EdgeCloud, reach us at partners@thetalabs.org

Scaling Arabic NLP Research at Cairo University with Theta EdgeCloud was originally published in Theta Network on Medium, where people are continuing the conversation by highlighting and responding to this story.