I asked my self this simple question while reading "Comment Abuse Classification with Deep Learning" by Chu and Jue. Indeed, they say at the end of the that
It is clear that RNNs, specifically LSTMs, and CNNs are state-of-the-art architectures for sentiment analysis
To my mind CNNs were only neurons arranged so that they correspond to overlapping regions when paving the input field. It wasn't that recurrent at all.