Multi-label sentiment analysis for Tunisian dialect on online social networks

Sentiment analysis is widely used NLP to automatically detect and analyze information related to user's opinions, attitudes, emotions, and feelings. Online consumers on e-commerce platforms and Online Social Network websites heavily trust and rely on online reviews posted by other users as much as personal recommendations. In this context, one largely used technique to process such data and give the appropriate insights is by deploying ML solutions in relation with sentiment detection. In fact, data collection and extraction, analysis and reporting methods has gained great attention in the context of NLP. However, the task in it self is quite challenging: understanding computer human sentences or linguistics within the text of these comments is dicult and remains prone to human errors mainly related to its comprehension and reporting. This issue was fortunately resolved for some languages such as English and French since many researches and work groups focused on treating and correctly structuring the content to deal with written documents in any NLP task. However, the problem is still pertinent when dealing with the Arabic language and especially one of its multiple dialects. In this master research, our goal was to investigate the issues related to emotion analysis for the Tunisian dialect to detect sentiments in OSN comments. This permits detecting not only positive or negatives trends or reviews posted online related to topics such as politics, economy, social life, culture, etc, but also provide the emotions such as Anger, happiness, sadness, fear ,etc that can be depicted in the content. In this work, we started by collecting the comments from Facebook as a reference source using a scrapping module. Besides, based on the literature review and the previous works performed in Arabic sentiment analysis we tried to identify the main steps required to process such type of unstructured data. In the pre-processing of the Tunisian dialect, two main resources were manually generated: the list of stop words and lemmatization and stemming dictionary that are mandatory to clean the collected text. In the experiments, the nal corpus generated was tested using various ML models and ML architectures to assess its eciency. Besides, it was also compared with a previous work proposed in the same context. All the generated materials from this study are made freely available for online use in a way to support the community to get engaged in the emotion analysis of the Tunisian dialect.