I have built my own sentiment lexicon using word embeddings and know I am stuck in how to add more features in order of train my machine learning algorithms.
First of all, the methods one would apply do not change much with a specific language. So, the answer to your question does not have to do anything with the Arabic. You might generalize your question and ask/research on a wider spectrum.
If you have a list of positive-negative words, you can basically count the numbers of those words in a given sentence or paragraph and normalize the numbers by dividing to the total number of words and present the result as the input to your machine learning algorithm. This approach is basically called "bag of words".
In order to be more precise, you can define some more classes rather than binary positive-negative classification. Some positive words might be much more positive. For example "marvellous" is much more positive than "decent".
You can also go beyond the 1-gram (one word) statistics by including 2-grams and 3-grams. For instance, adding the word "good" to your dictionary is 1-gram, if you add phrases like "on your feet" or "ear to ear" you would have 3-gram statistics in your set.
I studied hybrid analysis in my PhD thesis where I considered statistical sentiment analysis as language-independent since I mapped text words onto their counts. You can just extract BoW or N-grams as Tunc Guven Kaya said. However, you can also extract other features such as stylometric or grammatical although BoW is the most efficient.
In semantic sentiment analysis, you can use besides emotion words, also text grammar. Maybe, you have to parse texts automatically or scrutinize text grammatically. In this case, you additionally need linguistic expertise and analyze not only text grammar but also intensifiers and negotiations. Sorry, I don't know Arabic, otherwise I could have had other ideas.
Maybe, you find useful my PhD thesis that introduces own approaches to sentiment analysis and also an own hybrid approach that combines statistical and semantic approaches.