UAE: Pakistani university graduate uses AI to combat propaganda on social media

Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) graduate Muhammad Umar through his research has made a significant contribution in detecting propaganda on social media platforms, especially in cases where there is a mixture of low and high-resource languages.

While globally, there is vast amounts of research and time being spent on languages other than English for preservation, education, and language models, Umar, who hails from Pakistan, is contributing in his native language of Roman Urdu, among others.

Umar, a Master of Science in natural language processing (NLP), recognises the power of language in shaping opinions and influencing public discourse.

“Propaganda is a pervasive tool used to manipulate public opinion, and it is a growing concern in the digital age, especially in bilingual communities where little to no work has been done to detect it. Most propaganda detection work has been done on high-resource languages, such as English, leaving low-resource languages largely unexplored,” said Umar, who is part of the university’s first cohort of NLP graduates.

Umar noted that code-switching, which involves mixing multiple languages in the same text, is common in low-resource language communities and can make propaganda detection more challenging.

“In linguistics, code-switching refers to the practice of alternating between two or more languages or language varieties in a single conversation or text. In the context of my thesis, code-switched social media text specifically refers to social media text that uses a mixture of different languages, including English and Roman Urdu.”

Despite graduating, Umar is continuing his research and hopes to submit a paper related to detecting propaganda techniques in code-switched text at the 2023 Empirical Methods in Natural Language Processing (EMNLP) conference, one of the primary high impact NLP and artificial intelligence conferences for NLP research.

His model can be extended to other underrepresented or low-resource languages.

“This is one of the advantages of using deep learning models for NLP tasks as they can be trained on large amounts of text data and then fine-tuned for specific tasks. I ran experiments using several pre-trained mono-lingual, multilingual, and cross-lingual language models, fine-tuning them on the annotated code-switched dataset I prepared, and evaluated their performances. I found XLM RoBERTa, fine-tuned on Roman Urdu, outperformed all other baseline NLP models on our task and dataset.”

Umar said he feels privileged to be part of the groundbreaking programme, which is leading the way in the field of NLP.

“Being a part of the first batch has given me the opportunity to work closely with world-class faculty members, cutting-edge technologies, and a diverse group of fellow students who share a passion for NLP.”

Positive, negative impacts of NLP

Umar noted that NLP is a rapidly growing field with significant potential for positive and negative impacts.

• UAE: University gets Middle East's first-ever biocomputing innovation research lab