Rating inference using SVM or similar models

One interesting thing to do is using features from dependency parses. You can, for a small set of negation words, add features like WORD-is-headed-by-not or WORD-is-modified-by-ADVERB, and similar things. A simpler trick would be just using a known set of negation words and adding interaction features on all of them and all other words in sentences in which they occur. On the other extreme you have the dependency tree-based sentiment paper in ACL 2010.

In addition to your bag-of-words baseline, I'd also use a baseline that uses word bigrams (two consecutive words) as features. These features capture simple forms of negation. J. Dillards thesis on valence shifters is also a recommended reading [1].

[1] http://www.tacoma.washington.edu/tech/docs/research/gradresearch/ldillard.pdf

I suggest you take a look at the work by Isaac Councill et al. on negation detection. However, I would suspect that including negations have a relatively strong positive effect when using lexicon based methods, while the effect is somewhat smaller if you do supervised training on a BoW, or richer, representation. The reason is that sentences such as "I do not love this product" are not that common. Instead of negating a strictly positive word, often another construction is used, which typically involves other words that are not used to communicate positive sentiment. At least this is my experience from looking at review text.