In the table at the bottom of page 1, at the bottom of the first column, the rule "P -> P NP" should read "P -> IN NP." 

2c: I got asked a few times about what this problem means. The idea is that you can modify the training examples in the Treebank so that when you train a PCFG from that data, it will allow us to distinguish between higher and lower attachments. I.e., imagine you are writing a program that alters the trees in the Treebank in a systematic way to achieve the desired result. This program may introduce new nonterminals to the trees. The PCFG trained from your modified treebank (using the estimation routine) should have the desired effect of distinguishing high and low attachments.

4: It appears that the exponential prior on morph length actually leads to the same ranking as the Gamma prior, so ignore the comment "... and we'll see why that's a Bad Thing shortly" on page 5 and the last question of part e, "Why do the different priors lead to different results?"