n-gram models from foundations of SNLP

Today I finished reading the whole chapter of n-gram.First I felt that the book is not very good,because the sentences are usually very long,and the structure is not very in reason.Now when I looked back at it. There are many good ideas.
1.These methods for model smoothing are also for combining the information sources.
2.While some simple smoothing methods may be appropriate for exploratory studies,they are best avoided if one is hoping to produce systems with optimal performance.
3.use higher order n-gram models when one has seen enough data for them to be of some use,but back off to lower order n-gram models when there is not enough data.
Now I sumed up the chapter.
The most of the chapter is about the model smoothing.
1.MLE estimates
2.Add delta(laplace's law,lidstone's law, Jeffreys-Perks law)
3.held out estimator(Training set,held out set,test set)
4.Good-Turing estimation
5.Discounting method
6.linear interpolation
7.Backing-off(Kkneser-Ney back-off is the best(form Chen and Goodman 1998))
::Back-off models are in general not perfectly successful at simply ignoring inappropriately long contexts,and the models tend to deteriorate if too large n-grams are chosen for model building relative to the amount of data available.