Evaluation of Query Rewriting Approaches for OWL 2 Hector - - PowerPoint PPT Presentation

Evaluation of Query Rewriting Approaches for OWL 2

Hector Perez-Urbina, Edgar Rodriguez-Diaz, Michael Grove, George Konstantinidis, and Evren Sirin

{ hector, edgar, mike, evren}@clarkparsia.com

SSWS+HPCSW 2012

Clark & Parsia, LLC

○ Semantic Research & Development firm since 2005

○ Washington, DC ○ Cambridge, MA

○ Leading supplier of innovative semantic technologies

○ Semantic Web ○ Web Services ○ Advanced AI technologies

○ Committed to development and adoption of standards

○ Participated in several W3C working groups ○ Organized OWLED workshops to foster relationships

between researchers and industry

Query Rewriting

Query Rewriting Algorithm TBox Conjunctive Query Rewritings

Why?

○ Independent of ABox in terms of size and computation time ○ No need to update the rewritings every time the ABox changes ○ Evaluation of rewritings can be delegated ○ The same rewritings can be evaluated over different instances of data without having to be recomputed

Not a Silver Bullet

○ Rewritings might get too big/complex to evaluate efficiently (EXP even for QL) ○ Rewritings might need to be recursive or disjunctive DQs(1) for more expressive fragments of OWL ○ Expressive ontologies might generate rewritings that are more difficult to evaluate

(1) Datalog Queries

State of the Art

Comparison

Evaluation Questions

○ Can UCQs(1) be evaluated more efficiently (faster) than DQs? ○ Can Requiem be made faster by enhancing it with various

• ptimizations (like some of the approaches in other

algorithms)?

(1) Union of Conjunctive Queries

Blackout

○ Stardog component ○ Highly optimized version of Requiem ○ Eager query containment ○ Data oracle

Empirical Evaluation

○ Empirical evaluation of Blackout (Stardog 1.0.3) and Clipper ○ Support the most expressive logics ○ Two different approaches (UCQs vs DQ) ○ Using LUBM benchmark ○ 14 queries ○ 3 different TBoxes (for QL, EL, and RL) ○ 4 different ABoxes of increasing size ○ Very simple hardware ○ 8 GB RAM ○ AMD 3.2 Ghz - 4 cores ○ First, examine performance of the rewritings computation (size/complexity, time) ○ Second, examine how fast the rewritings can be evaluated using two database systems (DLV, Stardog)

Rewriting Computation Results

Rewriting Evaluation Results

Overall Blackout/Stardog Performance

Conclusions

○ DQ-producing approaches do not necessarily produce smaller rewritings than UCQ-producing approaches ○ Some optimizations, such as the data oracle, should apply to DQs as well ○ It is not necessary to produce DQs for RL and EL in many cases ○ UCQs are amenable to straightforward parallel evaluation ○ Evaluating the rewritings dominates the overall query answering time at large scales ○ It is worth investing time on optimizing the computation of rewritings that can be evaluated more efficiently

Future Work

○ Add datalog evaluation to Stardog based on Query-SubQuery algorithm ○ Parallelization of UCQ evaluation within Stardog ○ Explore the performance implications of new SPARQL 1.1 features (sub-queries, transitivity, property paths) in rewritings ○ Determine if Blackout can be used to handle queries utilizing these SPARQL 1.1 features

Questions?