#1 - You can use lexicon and learning together. My most successful sentiment analysis work has used a malleable lexicon which itself is trained using learning techniques, rather than classifying the whole example naively.
The rest of the tips apply to most NLP..
#3 - There's 'neutral', there's 'mixed' and there's 'unclassifiable'. Depending on the application you might want to filter out stuff that's contradictory or just not usable rather than assign it the stronger label of 'neutral'. Also bear in mind that in some domains most examples will be heavily biased towards some strong opinion (reviews is a big one - people often don't leave "It was ok" reviews) so you might need to tune the degree as well as the direction. Point #9 mentions the probability of positive/neutral/negative being equal; I rarely found this to be true.
#7 - Domain specificity is really, really important for sentiment, but in two different ways: domain (topic) and platform/format. Twitter vs. Yelp reviews is one example of where you might want to use a totally different base classifier because the context, language, length and relationship between individual items is very different -- but you should also be looking at using different variables (lexicon, classifier weights, etc) between restaurant reviews, bar reviews, gym reviews and so on. A single word can have drastically different implications in different topic contexts, and I found that getting this right (or rather close to right) was the most important thing.
#8 - Many sentiment analysis techniques are not very good, but don't fall into the trap I did of over-optimising an algorithm without taking into account the acceptable level for the use case. One of the reasons my sentiment analysis startup didn't get market traction was that naive algorithms were 'good enough' for the market I was focusing on, and trying to sell a 5-10% improvement without the necessary market intelligence just didn't help anyone. Note that depending on your domain, humans will disagree anywhere as much as 25% of the time on sentiment. This automatically puts a cap on how good you can get programmatically. It's also worth considering whether you care more about false positives, false negatives, etc.
#9 - Admittedly I haven't looked at the current state-of-the-art in annotated corpora, but if you're working on noisy, modern data then most formally annotated corpora will be useless. Figure out a way to get real humans to annotate for you.
#10 - Ensemble learning doesn't work so well, but a classifier that combines different approaches is much better (in my experience) than over-tuning a single classifier.
Also, don't forget that sentiment analysis needs input. Depending what you're working on, your data feed might need filtering first. There's some cool information retrieval challenges involved in doing that, which can often (again) make far more difference than the quality of the sentiment analysis algorithm itself.
