Hello!
As with text based sentiment problems, there are a few routes that can be taken. The first question I must ask is what would you consider as being a successful deliverable? A list of Tweets with corresponding labels and/or probabilities (depending on approach), for you data? A production type app that can be trained to identify labeled data that you can train and continue to use?
From experience, working from a sampling of labeled data to train a probabilistic metric (positive / negative), usually produces the best results regardless of language. What this would require, is a list of example text (Tweets in this case), that are labeled as positive or negative from you to start. Otherwise, we could explore doing very basic analysis to give you a broad view of sentiment.
Also, depending on what your goals are, I usually do some sort of latent index classification / topic model with these sorts of projects and usually find it useful to understand what keywords and groups of words are most representative of a collection of texts.
One challenge I should mention is that I don't speak Arabic at all. My lack of understanding of the Arabic language could present challenges that are unique to the application that may require more correspondence than is typical for this type of project.