Think of the word "awesome" - it generally correlates with positive sentiments (except in situations where awesome is used in its original meaning, then it's more neutral). Adding a "not" in front or behind the word "awesome" changes the sentiment of the phrase.
You can do sentiment analysis without using a NLP pipeline - there is no need to extract POS-tags, entities or dependencies if you have no need for them. In fact, you can just take a bunch of word vectors, and then average them over a sentence. Then softmax said average vector to your classes, and you'd get a very servicable sentiment analysis that would work 90% of the time. If you go down this path, I suggest using facebook's fasttext library which is..fast, and covers a lot of the challenges I mention below.