A Roman Urdu Corpus for sentiment analysis.

This article focuses on advancing sentiment analysis for Roman Urdu, a low-resource language using the Roman script to write Urdu, by developing a large multi-domain corpus named RUrdu. The corpus combines three existing datasets with a newly scraped Foodpanda dataset, totaling over 51,000 annotated reviews across positive, negative, and neutral sentiments, addressing prior limitations of small and domain-specific datasets. The study benchmarks 12 machine learning (ML) and three deep learning (DL) models using word embedding techniques including word2vec and FastText trained from scratch on RUrdu, finding that an adaptive bidirectional long short-term memory network (BiLSTM) with FastText embeddings achieves the highest accuracy of 85.4%. This work contributes a standardized, large-scale dataset and comparative model analysis to support improved sentiment classification in Roman Urdu, facilitating further research and integration into language technologies such as chatbots and social media monitoring tools.