Yes, we have. What we do is not keyword extraction, our tool suggests tags based on probabilistic algorithms. For example, if your document contains the terms Bush and Obama it should be tagged as politics even if that word is not present in it. Compare to the Yahoo Extraction Tool, for example. This approach will not add new keywords that would help in a search. It's only useful to have an idea of what the document is about.
The main problem is not the algorithm but the input data. Our system learns from millions of tagged blog posts among other sources. The quality of the tags varies a lot, and most of the work we do is about deciding what data to use for training.
The main problem is not the algorithm but the input data. Our system learns from millions of tagged blog posts among other sources. The quality of the tags varies a lot, and most of the work we do is about deciding what data to use for training.