benefits | modules | events | modelling principles | test questions | performance | tools
Performance
For the sake of performance, some peripheral assertions have been moved into their own ontologies. These are still imported by all.owl.ttl
- species.owl.ttl - traits, anatomy etc
- manufacturers.owl.ttl - creation of ships/droids etc
Although all.owl.ttl works with Pellet, it is significantly more responsive
reasoning and querying with events.owl.ttl
Reasoners
Initial testing on all.owl using participatedIn some Murder:
| Reasoner | Works? | Notes |
|---|---|---|
| Pellet | Yes | Fast. See notes on datatype restrs * |
| Pellet (Incremental) | Yes-ish | DL Query not working? see above for other issues |
| HermiT | Yes | Slow. 66-72s |
| JFact | ? | Need to build from source? |
| ELK | Partially | fast but v. incomplete - finds Ochi, Ziro but ignores all results from Events. Does not support DataPropertyAssertions, some DataHasValue, some Unions. Some tabs do not work |
Pellet note - min/max Exclusive causing error in unknown circumstances:
InternalReasonerException: Adding type to a pruned node "anon(222)" http://www.w3.org/2001/XMLSchema#int
Expressivity
Currently reported as:
- RRESTR role domain/range
- C concept complex negation (disjoints, differentFrom)
- U concept union
- CINT concept intersection
- UNIVRESTR universals
- E full existential - (non Top fillers)
- R role complex (refl chains, hierarchy, transitivive)
- O nominals
- I role inverses
- Q = (N + F) cardinality and functional roles
- (D) datatypes/roles
Speed Tests
We have some basic metrics being checked - see ConsistencyTest
There could be a lot of other queries testing
Expensive modelling (Pellet)
Cardinality
In some specific cases, cardinality on oneOfs cause a large hit. They are normally used as a syntax shortcut and should be avoided. See Issue #3
Characterising species
36-38s full classification in Protege of all.owl.ttl
- Moving hasPart usage (heavy use of cardinality) to another ontology not directly imported = 26-27s classification for all.owl.ttl
- Moving hasTrait usage only saves a second, but fits better in the new ontology
- Moving sentience usage drops classification to 15s
- Moving movesWith usage drops it all the way to 12s
Assertions on individuals have been retained in star-wars.owl.ttl
Characterising vehicles/roles
Leaving movesWith usage for 6 type of vehicles and troopers in base.owl.ttl adds 2/3s to classification
visited as a property chain
Nice to have this to show where characters have been, but it can be used at query time instead as its slow:
participatedIn o location -> visited
Property Domain/Range
Adding property domain/ranges base.owl Classification time of rebels.owl reduced from 23.5s to 18s
Expensive modelling (HermiT) - to revise using Pellet
allDifferent
There seems to be no significant overhead using Pellet (unlike HermiT).
If we do local allDifferent (each species/planets/etc) then we can get more cardinality queries, eg:
Event and (participant min 2 (hadRole some Mandalorian))
Property Domain/Range
Adding domain and range on all object properties increased the query time from 10s to 28s.
However, we do lose helpful consistency checking (as we have top level disjoints).
eg participants in Events are Actors, not Roles - leaving this range on participant is not too bad.
Whereas, adding a range to during automatically adds a few seconds on.
Domains/ranges were added to a top level test ontology which has now been removed using pellet.
Cardinality
Cannot add participant min 14000 to the battle of Exegol - it breaks the reasoner!
Different individuals
If we add an allDifferent (and remove Darths/Rens) what else can we assert / query?
For a start, it grinds the reasoner to a halt