The topic of the research is mainly centered on sentiment analysis of arabic texts, and is applied to tweets extracted from the Levantine region (Lebanon, Palestine, Syria and Jordan). The goal of the project was to infer opinions and emotions from tweets, which is a highly subjective and challenging task. The first step in this process was to gather the data. After collecting more than 4000 Arabic tweets, the dataset was created and annotated based on topics (sports, politics, religion, etc.), emotions (very negative to very positive), as well as the regions. Consequently, Alaa and Professor Hazem developed a machine learning model that could accurately predict the sentiment of the tweets; however, they still needed to improve predictions pertaining to identifying the topic of the tweet. In order to do so, they worked on domain adaptation, which is basically transferring the knowledge from one task to another task.

Alaa thus proposed the usage of a new deep learning model called DARL, which is based on a combination of autoencoder and adversarial training, beating state-of-the art models at that time.

The research had major outcomes, such as creating an Arabic Tweets dataset that did not exist beforehand. This contribution is significant given that the Levantine region has fewer digital resources and less investment in research compared to the rest of the world. In addition, the development of advanced machine learning models for domain adaptation and transfer learning has a great impact, not only in analyzing Arabic text, but in a wide range of tasks across multiple languages. Creating datasets from scratch can be very costly and time consuming, but the more groundbreaking advancement is developing the capability to transfer information used in one task to another via these new transfer learning models.