Section 5.
In the discussion section, Shannon has compared the performance of different EB models and then discussed ways to improve those models. She has concluded that the Posterior Biasing model with regional effects is the best performing model based on predictive evalution criteria. Overall, I like the organization of this section. There aren't any logical gaps and enough explanations for me to grasp what she is concluding. My only comment is about the use of lengthy sentences. Splitting long sentence into several small sentence by exploting Gopen and Swan's idea of topic position and stress position would have made the reading easier. 

Here are some specific comments on each sub-section.

5.1

last line paragraph 1: It will be helpful to say what is the basis of differentiation here. Is it the type of dependencies in the data?

 
paragraph 2: You could perhaps remind us what the evaluation criteria or metric is here. (Is itCV error?) Think about the short term memory of the readers mentioned in Lebrun. Though you do talk about CV error later in this section, I think it will be useful to express it explicitly here.


section 5.2 


page 22 paragraph 1 (sentence starting with 'Posterior Biasing is ...'): You could try splitting one long sentence into smaller sentences. There are lot of ideas floating in the same sentence and it is hard for me to see the stress position (as discussed in Gopen and Swan) in this long sentence.


page 22 paragraph 2: here you have mentioned two ways to improve the model. It might be helpful to be explicit about it.

page 22 last paragraph, first sentence: Again I think there are many important ideas in one long sentence. It will be better if you split them into smaller sentences.