Let me elaborate a bit more on how the app computes sentiment. For a particular word, its sentiment is the average of the sentences' sentiments which contain both the word and the brandname (in order to identify the sentiment targeted at the brand, not just the overall sentiment).

For example, in the case of Mazda where you say that "regret" is classified as positive, if you look into which message it comes from you can see the original sentence: "Buy a Mazda, you won't regret it :)"

I agree with you that the word cloud is not useful on its own, and this is why you can click on a word to see the actual messages. Think of the word cloud as merely an entry point into a more detailed analysis by a human.

Thanks for the feedback.