Slides were presented by Zhe Wu remotely, using Oracle conferencing software
Slides for this session: Media:zhe-f2f1.pdf
Slide: Oracle 10gR2 RDF
Some technical difficulties continued w.r.t slide presentation
Zhe Wu: many ways to insert data.
Zhe Wu: ...in 10r2 we also support some inferencing and rules. we use forward chaining approach
Zhe Wu: ...also query using a SPARQL-like syntax
Zhe Wu: ... this was all in 2005
Zhe Wu: this year new release with new features. faster loading, owl reasoning with proof generation
Zhe Wu: ... overhauled performance w.r.t. load and query
Zhe Wu: ... just recently added Jena / Oracle adapter
Zhe Wu: ... joint with HP
Zhe Wu: subset of owl is supported
Uli Sattler: i'm curious about what scalable and efficient means
Zhe Wu: i will show some numbers later
Zhe Wu: ... re: what is supported - forward chaining rules implementation for fast query answer
Zhe Wu: ... conclusion in ISWC 2006 paper was that existing reasoners had problems with large ABox data
Slide 7: OWL subsets supported
Zhe Wu: rdfs++ added as a "minimal" extension to RDFS
Zhe Wu: ...owl prime, what is now proposed as rdfs 3.0
Slide: semantics characterized by entailment rules
Zhe Wu: owl prime has ~50 rules
: does somebody know the exact literature reference for OWLPrime and can send it?
Slide: Applications of partial dl semantics
: It has more fine grained analyses, including in terms of AL, ALHF, SHIF, and SHOIN, on the one hand, and RDFS(DL), DL-Lite, EL++, and "non-tractable"
: It also discusses "repairable" OWL Full ontologies and (sketchily) how the non-repairable ones fall into OWL Full
: It's not at all clear to me how to map the analyses in that paper to OWL Prime (in part because I don't understand OWL Prime)
: paper also took a DL approach to the world, the raw data showed the great bulk of the stuff out there, pre-change, was low expressivity RDF or RDF with a little OWL - it's where the RDF 3.0 proposal came from
: Uhm...I don't knwo what you mean by "raw data" and "great bulk"
: In fact, I don't see that anything I said had anything to do with what fell into RDFS or not
: The repair had mostly to do with the nominally owl full documents.
: In fact, if you look at table 2 and table 3, the second part of your assertion is at least questionable
: """Of the 307 OWL Full documents that can be patched, 63% become OWL Lite documents, and just 37% become OWL DL. Two observations can be made. First, The majority (91%) of the OWL Full documents (from Table 2) can be turned into a decideable portions of the languages by adding type triples. Secondly, the majority of RDFS documents (95%) can transition to OWL easily by adding type triples and use OWL vocabulary instead of RDFS vocabulary."""
Slide: support semantics beyond owl prime
Jeremy Carroll: question about example being supported directly in the future
Zhe Wu: exactly
Achille Fokoue: question about updates to abox
Zhe Wu: i'll get to that later
Slide 13: Advanced options
Alan Ruttenberg: question about time, can we focus on questions now
Zhe Wu: ok, i'll quickly browse remaining slides, then go to questions
Slide: implementation in rules
Zhe Wu: I want to stress that we did not handle one property at a time
Zhe Wu: I'll jump to query answering slide
Zhe Wu: ...that's all I wanted to cover, open for questions
Post Presentation Q & A
Ian Horrocks: the tractable fragments doc describes fragments with known database mapping. wondering why you didn't choose one of those
Zhe Wu: we started by asking existing customers what they needed. most told us they just needed simple extension into owl from rdf
Zhe Wu: ... pretty much the approach was driven by customers and need to implement efficiently
Ian Horrocks: but, customers said you needed something small (rdf + a bit) which is exactly what the fragments are. instead you chose a large fragment and implemented incompletely
Zhe Wu: so far, for those other fragments we have not found a complete rule set (except PD*)
: (I find the "small fragment" vs "large fragment" language very confusing, because I don't know what the metric is. large number of terms? large number of users? large implementation effort needed?
Uli Sattler: I want to echo ian and point out that you don't allow intersection, but a clever user would have it
Uli Sattler: ...and to be complete complexity becomes a problem
: Complete with respect to one reasoning problem is sound with respect to another. ( In OWL? Really?? ) [Scribe assist by Sandro Hawke
Alan Ruttenberg: they're not trying to be complete
: the mentioned paper by ter Horst seems to be the following: Herman J. ter Horst, Completeness, decidability and complexity of entailment for RDF Schema and a semantic extension involving the OWL vocabulary, Web Semantics: Science, Services and Agents on the World Wide WebVolume 3, Issues 2-3, , Selcted Papers from the International Semantic Web Conference, 2004 - ISWC, 2004, October 2005, Pages 79-115.
: Keywords - Ontology; Semantics; Entailment; Completeness; Computational complexity
Boris Motik: echo ian, observes that fragments exist which can be implemented with a set of complete rules
Bernardo Cuenca Grau: i'm worried about soundness and worried about what "sound and complete" means here. I don't understand the semantics
Bernardo Cuenca Grau: ...b/c you haven't implemented the OWL semantics, you've chosen some of the OWL DL vocabulary
Zhe Wu: we do care about completeness, but don't consider it critical
Zhe Wu: ... completeness is evaluated w.r.t. query answering for some benchmarks, etc.
: From a spec perspective, this fragment seems to be *implementation* defined...which is a bit worrisome
: My test for this would be, without looking at thier rules or using your rule engine per se, can i write an implementation from a publically available description?
: I would point out that DB communities tend to do language/sublanguage without model theories very comfortably
Jeremy Carroll: what I hear from customers echos Zhe's comments.
Jeremy Carroll: ...I note that much of the questioning is hostile
Alan Ruttenberg: I agree
Jeremy Carroll: that may be b/c much of the questioning is coming from members with different user groups
: I would disagree with the assessment of tone...is it even relevant?
Carsten Lutz: I would strengthen Boris claim and say that most (if not all) other fragments admit forward chaining, which is sound and even complete,
Carsten Lutz: and the rules are easily derived.
: I also find that Web 3.0 companies, including the folks interacting with us on the billion triple challenge, also come from the perspective Zhe represents
Ian Horrocks: it wasn't intended to be hostile. I was trying to understand whether Oracle would be interested in more well understood and explainable fragments
Ian Horrocks: ...e.g., dl-lite which can be implemented in a database system, and also in a rule system
Carsten Lutz: Same applies to EL++
Discussion of PD* soundness and completeness in a rule based implementation, which scribe didn't capture
: fwiw, those fragments got little or no traction with the folks I consult for - they care about parallelizability and performance over the more understood stuf - their work is largely heuristic anyway
Ian Horrocks: the problem with PD* is that it doesn't implement a subset of OWL, it implements PD*
Jeremy Carroll: it depends on what you mean by fragment of OWL
Alan Ruttenberg: I hear interest in co-ordinating on database fragments with Oracle
Bijan Parsia: to standardize a fragment, we need a well defined specification that we can all understand
: Zhe, I might have sounded hostile, which wasn't intended: some of us simply have a specific reading for certain words like "reasoner", and I couldn't see how this could be possible.
: Zhe, also, if you want to see how IntersectionOf can be simulated with someValues and AllValues, ask Carsten.
IRC aside on specifications, definitions, and implementations
The following IRC conversation happened in parallel to the in room verbal conversation and some other IRC exchanges. It continued until Ivan the point in the minutes where Ivan requests an end to side conversations (just before Semantic Subsets)
: What I am arguing for is that there are some important communities out there to whom the fragments they care about are not those tied to Uli's definition of reasoner
: hendler, people are not really paying attention to IRC.
: Jim, I appreciate this -- but "reasoner" was used on Zhe's slides, and i simply wanted to know in which sense.
: Bijan - agree with needing a spec, but I'd point out most programming languages get by just fine with operational semantics - in fact, since you implement Pellet in JAva, in a certain sense you're trusting that they get it right in some sense -
: Jim, I guess what we would like to see is a consensus of what we mean by "Tool/reasoner X supports feature Y"
: First, I'm not saying anything about the specification style....but at the moment we don't have a clear spec. I don't know where to start other than by looking at Oracle's implementation
: Uli, how do we show "Java" supports "begin/end loops"? that seems to be something in the real world that we could model for some (not all) of our work
: this is the point!! thw Owl language features all have operational semantics that are good enough for many people in many situations - so when Oracle says we support X, why do we need more than that
: Second, there are differences between programming languages and ontology/data modeling languages. I hear your point, but find the analogy rather unconvincing.
: Jim, I would never dream of trying to do this - but i would like to try to say what it means for a reasoner to support feature X
: What I'm arguing is that RDFS 3.0, or OWL Prime, might be better looked at less as ontology languages (leave that to OWL DL) then as useful data analysis languages
: But I don't see why "useful data analysis langauges" don't need a clear spec
: this is what my nose is rubbed in when I attend the Sem Tech conference and places like that
: bijan - the question is what is the definition of a clear spec.
: Furthermore, model theory is pretty easy way to specify something...at the moment, no one has proposed anythign else
: if there is a set of rules defined as those in RDFS spec, is that clear?
: they may be informative, however, that is how most people understand semantics
: but there are no model theories for many things, and model theory is not the only way to spec other things - like these rule-based examples
: Zhe, we would call this "operational semantics" or such like and would be split about how clear this is
: Again, my test is can I write an interoperable implementation without looking at your implementation
: if we agree on a set of rules, then interoperability is not an issue
: Not at all since I don't know what you -> means
: For example, i might no use those rules *in* my implementation
: using Hendler's example rule,
: I might want to use a very different technique
: if we agree on that, then we are interoperable
: The reading of rules, for example, differ in whether you have contraposition or not
: Bijan, that is either easily defined, or can be left to philosophers trying to write PhDs, in the real world, lots og languages work this way - but if you want something better - okay, we'll use SCL
: and whether you "apply" it to all named individuals or to *all* individuals
: so, I agree that your defintion of inverse seems clear, but when you want to implement it, there are questions coming up
: complementOf not in RDFS 3.0 for precisely that reason
: Again, I find englisch often clearer than things like "->' or such like
: Or did you mean CLIF from the ISO standard, Common Logic
: In any case, I'm asking for a spec. We can beat on the spec and if we find problems we find problems
: but anyway, the point I'm making is not to oppose model theory - but the problem is to get the model heory right we have to put restrictions on the languge that some of us cannot live with easily
: this is why OWL LITE is a subset of DL - not of Full, so there is no fragments for Full - which is what i often get asked for
: Jim, model theory doesn't restrict things?
: I'm very skeptical about it, but I don't know because I don't know what the current fragment actually *is*
: Uli, so we could define the language fragment based on other concerns and then define it via model theory - that doesn't bother me at all - I'd be fine with that
: But I'm open to being convinced otherwise...but I'm more convincable by a proof of concept (at least) than high level discussion
: so Oracle has implemented OWL prime - what did I miss?
: And this is true for programming langauges as well
: Jim, what would you define via model theory? The fragment or its semantics?
: Uli - whichever you want - I'm not going to need to read that document anyway ;-)
: Jim, I think we simply disagree what it means to *implement* a fragment
: There are langauges defined by *specs* (including Java, Common Lisp, C, C++) and those defined by *implementation* Perl, Python
: So, frankly, I don't want to port Oracles implementation. That's probably not even legal
: fine - I want a fragment of OWL that is defined by *specs*
: I want a specification sufficient for independant implementation
: But jim, if the specs happen to do it by model theory and capture the language you want, what do you care?
: If the specs are clear enough for me I don't necessarily require model theory ( though it helps so we can understand the relation to existing OWL specs)
: bijan - the poiunt is I don't care - and I said that - what I care is what is in the fragment first, how to define it second
: How do you know what's *in the fragment* without a definitio of what the fragmetn is?
: Bijan - you're arguing circularly - but I can cut through it - I have created a wiki page with a description of exactly which language features I want to include - how to define it in a spec is something I'm happy to discuss
: Jim, we're talkin in the room now, and need to focus on this discussion in the voice channel, so no more chatter on IRC, please.
Bernardo Cuenca Grau presenting from slides in person
Bernardo Cuenca Grau: motivation of owl-lite was easier owl. b/c owl dl and full are rich and complex.
Bernardo Cuenca Grau: ...problem is owl-lite is broken b/c it doesn't address interactions between constructors
Bernardo Cuenca Grau: most features held out of owl-lite can be recovered through "back doors"
Bernardo Cuenca Grau: existing document includes fragments which
Bernardo Cuenca Grau: .... are well understood, documented, etc.
Bernardo Cuenca Grau: we don't expect users to go over recent literature on tractable fragments, so wanted a single document
Bernardo Cuenca Grau: most of the languages I will describe are "families" of languages, we decided to keep 1 from each
Bernardo Cuenca Grau: 1st is EL family
Bernardo Cuenca Grau: ...used in bio-medical already
Bernardo Cuenca Grau: stress that these fragments are not academic exercises, there are direct applications to existing ontologies
Bernardo Cuenca Grau: 2nd is DL-Lite family
Bernardo Cuenca Grau: ... designed for large number of instances in database technology
Bernardo Cuenca Grau: approach is similar to what zhe described, do work in tbox, then pass to database system for query answering
Carsten Lutz: reiterate bernardo, but contrast with zhe's approach. dl-lite change the ontology to use database technology, not change the database technology
: would be nice to get a literature reference to OWLPrima - the description on the slides was not clear enough
Alan Ruttenberg: another difference is in oracle you can query for classes, in dl-lite only instances
Bernardo Cuenca Grau: you can do tbox reasoning, but designed for abox answering.
Bernardo Cuenca Grau: I picked the particular dl-lite language b/c it is between rdfs schema and owl dl
: is there a scalable implementation of dl-lite? commercial tool?
: Uli, what kind of tool? what is the scalability?
: Zhe, yes ther is, I think: search for Quonto
: Zhe, because it translates queries into SQL queries and leaves everything in the DB.
Carsten Lutz: Zhe, I understand that you would like to do forward chaining. A lot of fragments can be captured in a sound
Carsten Lutz: (and even complete) way with this technique. I would like to learn what is your idea of tractability and
Carsten Lutz: scalability. Is it forward chaining per se, or is it a rule set that does not produce too many new facts?
: Carsten, the scability and performan requires are determined by the market.
: Carsten, people are asking for hundreds of millions of triples and beyond
Bernardo Cuenca Grau: next is Horn-SHIQ
Bernardo Cuenca Grau: ...can reason without disjunctions
Bernardo Cuenca Grau: ...and low complexity for query answering
Bernardo Cuenca Grau: other fragments dlp as a bridge to rules
Bernardo Cuenca Grau: ... but it may be more "hacky" that horn-shiq
Bernardo Cuenca Grau: questions for wg
Bernardo Cuenca Grau: ....1 do we fix owl lite
Bernardo Cuenca Grau: ....2 does that mean select one of these fragments
Bernardo Cuenca Grau: ....3 or do we present a menu of fragments?
Bernardo Cuenca Grau: not in slides - do we want semantic subsets of owl full?
Bernardo Cuenca Grau: ....e.g., owl full versions of these fragments? do we care about complexity of the full fragments? about compatibility?
Ivan Herman: request to drop side conversations
Ivan Herman: ... and focus
Alan Ruttenberg: little time, can we start with semantic subset of owl full?
Ian Horrocks: semantic subset means no change to syntax, but sanction smaller set of conclusions
Jeremy Carroll: example is pd*, which specifies what semantic rules are thrown away
Peter Patel-Schneider: pd* throws away *parts* of rules
Ian Horrocks: this is picky
: to Hendler: fragments of OWL 1.1, which: are the result of years of research, have “nice” computational properties, are already supported by tools
: But there are fragments which are not included that have all those things as well - Oracle Prime being a perfect example
: to hendler: Fragment Goals: suggest possible fixes to OWL Lite, inform the OWL community about recent research results, help users & tool designers
Alan Ruttenberg: how comfortable are people with this type of fragment
Alan Ruttenberg: ... does anyone want to say this is a lousy idea.
Peter Patel-Schneider: yes, its lousy b/c you can be arbitrarily picky
Ian Horrocks: its a lousy idea b/c it blows away the idea of interoperability
Bijan Parsia: qualm that methodological design principles are "unclear"
Bijan Parsia: ...guidance for making decisions seem more arbitrary, a dangerous rat-hole
Bijan Parsia: ... would rather people say they are incomplete than building incompleteness into fragments
Jeremy Carroll: in response to ian, any semantic subsetting would need to be clear that it is a subset of spec and an explicit, agreed semantic subset
Jeremy Carroll: ...e.g., oracle and hp would agree on semantic subset and interop on at-least the semantic subset
: +1 Jeremy -- "incompleteness" is fine When It's In A Specified Fragment, that is implementated in multiple places, etc.
Alan Ruttenberg: if we call this fragment or conformance level, it seems useful
Alan Ruttenberg: ...that baseline entailments are necessary, but additional entailments may be ok
Bijan Parsia: if we shift from language fragments to reasoner conformance I'm more comfortable
Sandro Hawke: Bijan, "Reasoner Conformance" might be a more useful notion here than "Language Fragments".
Sandro Hawke: ...I have examples of people specifying this at a tool level.
Jeremy Carroll: i'd be happy with such a rewording. i don't see it as notable
Alan Ruttenberg: does such a distinction help others
some affirmation to alan in room
Zhe Wu: ?
Alan Ruttenberg: he said it would be useful to say we support same entailments
Ian Horrocks: more comfortable defining conformance that fragments
: and jeremy's suggestion sounds like standardising implementations
: Jim, you still there. Hard to follow the IRC. I can read what you write if you want to respond.
Bernardo Cuenca Grau: users are comfortable with incomplete reasoning. swoop offering rdfs reasoner as a choice is an example of this
Bernardo Cuenca Grau: ...more comfortable with that than trying to specify semantic subsets
: bernardo also stressed the importance of a clear semantics as reference ...
Jeff Pan: i agree with bernardo and others.
Jeff Pan: ... implementation does not specify fragment.
Boris Motik: i just looked at pd* , this seems like definition. I think it is a useful fragment if evaluated a certain way.
Ian Horrocks: i didn't say pd* was bad, that we'd be standardizing an implementation. it was a reaction to jeremey's comments on what hp and oracle might do
Sandro Hawke: owl is unique to me b/c it doesn't specify what the tools do, people read into that. specifying the tools would be useful. as a customer I expect that and would like it
: ina: it is not just hp and oracle, owlim, allegrograph as well
Bijan Parsia: justifying discomfort - seems likely that over time fragments specified in such a way are likely to move
Bernardo Cuenca Grau: on sandro's comment - we should specify reasoning services
Bernardo Cuenca Grau: ... it's not in the spec for OWL DL. for fragments the services descriptions would be useful
Sandro Hawke: i don't know what the terms are, the market decides
: It should be customer driven. When they want to find on the shelf, those should be the things defined in the spec. [Scribe assist by Sandro Hawke
Ian Horrocks: its difficult to imagine semantic subsets not drifting apart
Ian Horrocks: ... it has been a success for owl that interoperability is so good, considering
Jeremy Carroll: responding to standardizing tools - yes. there is value to user if they know different tools perform the same
Jeremy Carroll: ... this wg could provide appropriate conformance levels where vendors and user community come together
Jeremy Carroll: ... clear that motivations from academic community are useful, but they aren't the only motivations
: +1 Jeremy: there is real value to the customers in knowing that a set of products will all do (at least) the same thing. It would be a service to the community for this WG to provide that.
Alan Ruttenberg: no one is saying market is unimportant
Uli Sattler: clarification on user needs?
Jeremy Carroll: users need some sort of specification, but don't need to know behavior is exact
Alan Ruttenberg: I want to poll for consensus on how to procede
Sandro Hawke: i don't understand
Alan Ruttenberg: I want to know if people think these fragments are useful
Alan Ruttenberg: ... defined as a minimum set of entailments
Bijan Parsia: reasoners can conform to the language to different degrees
: subset of language + conformance level, is something similar to the way current languages (e.g) doing, like, Deprecated apis + core language( and specialized apis)
Alan Ruttenberg: we should aim for something specified like pd*
Ian Horrocks: declarative...
Alan Ruttenberg: yes, declarative
: Q1 - The Working Group should (formally, precisely) define conformance levels, defining groups of reasoner which can do certain kinds of reasoning (all for a given OWL Fragment).
: Q1 - The Working Group should (formally, precisely) define conformance levels, defining groups of reasoners which can do certain kinds of reasoning (all for a given OWL Fragment).
Uli Sattler: we would later know e.g., what it would mean for a reasoner to conform to particular level?
Alan Ruttenberg: yes.
Jeff Pan: what does conformance level mean? is it in terms of benchmark?
Uli Sattler: provides example
: The Working Group should (formally, precisely) define conformance levels, defining minimum levels of inference that would be found?
Carsten Lutz: degrees of incompleteness?
: The Working Group should (declaratively) define conformance levels, defining minimum levels of inference that would be found?
Alan Ruttenberg: degree of completeness
Alan Ruttenberg: ...fragments are syntactic fragments
Alan Ruttenberg: ...conformance levels are distinct
: Q1 - The Working Group should (declaratively) define (one or more) conformance levels, defining minimum levels of inference which would be performed (for a given OWL Fragment).
: yes if we are talking about a declarative way of defining minimum levels of inference
Jeff Pan: there might be difference between alan's and uli's suggestions
Alan Ruttenberg: distinction is unimportant now
: when we, as a group define confromance levels, it is very useful to look at current market
: including HP, Oracle, AllegroGraph, OWLIM etc.
Alan Ruttenberg: reads Q1 as above
: lots and lots of people raise their hands in favor.