I agree to a point. For one, I would not group Finnish and Chinese together. Now, I can't speak much for Finnish, but I am positive a statistical approach to Chinese provides very poor results. However, with that said, I do agree that domain logic and a large relational data set is a good approach. Essentially relationships between words and symbols need to be established with false positive and false negative associations broken down and re-established in near real time to come close to a perfect solution. Basically, the problem is too expensive to solve.