Preprocessing for Sentiwordnet

[Hadsga]

Hi,

I want to perform a Sentiment-Analyses with SentiWordnet on Tweets. The analyses should be on the document level, i.e. I want to classify each Tweet as positive, negative or neutral. However, the quality (Accuracy) of the results are dependent on the preprocessing. While there are a bunch of papers dealing with the influence of preprocessing using machine learning applications (SVM, Randon Forest etc.) I found no paper which investigated the influence of preprocessing using Sentiwordnet. Is there a guideline which methods are essential to get good results?