TunTap: A Tunisian Dataset for Topic and Polarity Extraction in Social Media

The massive usage of social networks has recently opened up new research avenues in the fields of data mining and decision-making. One of the most relevant forms of data generated by users in social media is an unstructured text that identifies their emotions on a given topic. Analyzing this new form of writing to extract valuable information is a challenging task, and could be of great interest in several fields such as healthcare, business intelligence, marketing strategies, ... to name but a few. This article considers topic and polarity extraction in application to Online Social Media (OSM) analysis, in the benefit of numerous domain applications. Implementing sentiment analysis and topic extraction algorithms for the purpose of detecting the polarity of a given comment towards a certain topic requires a sophisticated machine and deep learning supervised models and, at the same time, collecting, preparing and annotating a huge amount of data to train those models. In this paper, we propose a special dataset that can be used to extract both topic and polarity features from dialectical messages used in Tunisian daily electronic writing across the most popular OSM networks. We collected our data by crawling posts and comments' text from Facebook, Twitter and YouTube using related network graph API. In this work, we describe the whole pipeline used to prepare our corpus as well as the several extensive experiments setup and results conducted to evaluate the generated dataset. Up to our knowledge, the proposed multivariate Arabic dataset (Topic and Polarity) of Tunisian dialect is a first-time introduced in the NLP community up to now, and we made it publicly available on GitHub (https://github.com/DescoveryAmine/TunTap).