Studying paper of Li Juanzi

This morning I am reading Li Juanzi's paper 《语言模型中的一种改进的最大熵方法及其应用》which is published on JOURNAL OF SOFTWARE.

In her paper, she used an updated method combining maximum entropy, mutual information and Z-test to choose the best feature of context for a multivocal word and then used IIS algorithm to optimize the parameters of the linear model for Word Sense Disambiguation.

The experiment results are displaying the advantage of this approach. But I think the paper has two flaws. Firstly, the experiments for WSD is not enough. Secondly, Z-test is usually used to test normal distribution for large scale's samples. And in this paper there is a connotative hypothesis that the mutual information between the feature set and the category a multivocal word is followed normal distribution. The experiments did not prove this hypothesis.

And I think this experiments could be done more fully, and the experiment should do the hypothesis test.

OK. There is an idea. We can arrange some little experiments of our current information corpus to do some hypothesis test. En, this idea should be thought more and more.