Some cases are just difficult, but the overall accuracy could probably be improved considerably if the sentiment analyzer were calibrated for the domain. AFINN (linked above) is calibrated from newspaper articles, which almost certainly have a different distribution of word/sentiment correlations than Tweets do. It's not hard for me to imagine that "killed" is a better predictor of negative sentiment when classifying news articles.