1) The document proposes using semantic filtering of Twitter data to improve tracking and prediction of influenza-like illnesses compared to existing keyword-based methods. 2) A two-step filtering approach is used - first filtering tweets for syndrome-related keywords, then further filtering based on semantic criteria like negation, emoticons, hashtags etc. 3) Evaluation shows the semantic filtering approach improves correlation with CDC influenza surveillance data compared to state-of-the-art keyword filtering methods, with the best method achieving a 98.46% correlation.