Put comments here for "Statistical relational learning: Four claims and a survey" by Jennifer Neville, Matthew Rattigan, and David Jensen.
From James Cussens:
first page, 2nd col, 1st para: It's often claimed that relational learning involves a massive model space on the grounds that with a richer representation you can define more models. However, in much ILP the first-order framework is often used to tightly constrain the model space. Most of my work on grammar learning with ILP is expressing linguistic constraints in Prolog. I remember some work my colleagues did at Oxford (see p255-256 of 
this paper) were once the model space was correctly constrained, learning became possible.
It seems a pretty deep question as to which attributes are really intrinsic to an object, related to the philosophical distinction between internal and external relations. Forming "relational features" amounts to pretending that some extrinsic property is intrinsic. In practice, this issue seems to be decided according to what is convenient.
2nd page The lack of non ad hoc measures of confidence is a real drawback of predictions from theories produced by ILP.
The Bernstein et al quote is correct. Hierarchical Bayes approaches (see Gelman et al) exploit this view a lot.
Section 4: Yes, examining particular sorts of relations seem necessary. "Relational model" is so all-encompassing that it makes it difficult to say much of any import.
Section 5: Of course, an acknowledgement that data instances may not be independent is not new to SRL.