Review Comment:
While I still have some concerns about the practicality of the system (as expressed in the previous review), I appreciate that the authors have added some discussion on these aspects, introducing some scenarios elaborating how it could be used, more in-depth treatment of strengths and weaknesses, as well as some discussion on optimisations. In general, I am quite happy with how the authors have addressed the comments and I believe that pushing the formal proofs to the Appendix was likewise a good choice in that the paper reads a lot better now. I still have some comments I'd encourage the authors to look at, but I don't see anything significant enough on my side to warrant holding the paper up with another revision cycle. Hence I recommend an accept.
I just have one technical doubt that the authors may wish to look at clarifying, regarding again the notion of an "extended reduction". When introduced, the authors state that "the extended reduction of an ontology K is required". The use of the definite article here suggests somehow that there is a unique extended reduction, which I don't immediately buy. I mentioned it in the previous review and although the text is now clearer, I still have questions over the uniqueness of extended reductions. In Example 3, the authors provide a case where a cyclical ontology leads to such a problem, and thus restrict to acyclical inputs. But I am still not convinced that these extended reductions are unique in acyclical cases, and what it would mean for the rest of the proposals if they are not. Take for example rule (e1) in the paper and the following K:
(x,dom,c) [F1]
(y,sp,x) [F2]
(z,sp,y) [F3]
This K is acyclical. Just considering (e1), we get in cl(K):
(x,dom,c) [F1]
(y,sp,x) [F2]
(z,sp,y) [F3]
(y,dom,c) [C1]
(z,dom,c) [C2]
Now let's consider applying (e1) in reverse.
1) If we apply it in reverse to (F1+F2->C1), and remove C1 first, then we cannot apply (e1) in reverse any more and we've reached a fixpoint at {F1,F2,F3,C2}.
2) If we apply it in reverse to (C1+F3->C2), and remove C2 first, we can still apply step 1) above, so we will reach the fixpoint of {F1,F2,F3}.
Probably this just requires a little more discussion to make clear. Specifically, in the left hand column of page 6, I would like the authors to make clear: is this extended reduction unique for acyclical cases, if so why, and if not, what effect does it have on the later relaxation/approximation steps.
MINOR COMMENTS:
* Page 1, right column, bad box with "(which"
* "arising form" -> "arising from"
* Page 6, right column, RELAX semantics, would be worth fixing that widow for the equation
* Table captions should go on top? Double check the SWJ style.
* "only on triple patterns containing a regular expression in which at least one constant appears" At first I read this as the regular expression must contain the constant. I would rather say triple patterns where the subject or object is constant.
* "We ran each query 6 times ..." I missed this the first time around but dropping the first run might be problematic in that the engine could (at least in theory) cache the query against the result, where the latter query runs would then just be retrieving the answers from the cache, rather than executing them. It would be worth mentioning why this is not problematic: i.e., does Jena TDB cache queries against answers? (This might be a more significant issue if it weren't for the fact that many of the queries are quite slow ... if they were all returning in milliseconds, I would suspect caching.)
|