I'm more-or-less in agreement with everything you say below (although it seems to me that the superclasses of a type are static only when the set of types is static, without new knowledge becoming available, and the fact that "correct" and "bizarre" interpretations may both be present is still a tremendous barrier to developing practical applications), but the "meta-discussion" is really what's at stake, I think.
In other words, if Bibframe is a language for discussion and information exchange only amongst catalogers and their professional colleagues, my objections to requiring inference evaporate. In that situation, I'm all in favor of using all technological capabilities to the fullest. If, on the other hand, Bibframe is a language for publishing bibliographic information to the wild wide Web,
then Bibframe triples could be consumed by anyone and anything, and the assumption of inference (even given the points you make below) seems to me to be too much.
---
A. Soroka
The University of Virginia Library
On Nov 7, 2014, at 3:35 PM, Simon Spero <[log in to unmask]> wrote:
> On Nov 7, 2014 12:35 PM, "[log in to unmask]" <[log in to unmask]> wrote:
> > In some cases, it may be possible for Bibframe applications to do a _better_ interpretation if types are present or can be inferred, but it should never be the case that an application can make no sense or can only offer a bizarre or nonsensical interpretation of some Bibframe triples without typing information.
>
> I'm going to stay out of the meta-discussion and just make some technical remarks.
>
> 1. If a set of assertions where all entities are fully typed has a model, then removing an assertion does not eliminate the model ; it just allows for other models. So removing type information cannot leave *only* the bizarre models.
>
> 2. *Some* inferences for *some* axioms can be found extremely cheaply in *some* implementations. (Asterisks indicate bold, italic, comic sans, and blink).
>
> One example is inferring a type of an entity that is the subject (value) of a property with explicit simple domain (resp. range) statements, when querying a data store that is mapped from a relational database.
> In this situation, some type information comes straight from the mapping, so it can be done "for free" . This inferencing is sound but not complete.
> Of course, entities mapped from an rdb will usually have a type, though it may not be the one you want.
>
> Finding the superclasses given a type only needs to be done once. This does not find all the classes of an individual but it does a lot.
>
> 3. Restricted profiles of owl can perform some inference tasks in real time. Production quality databases using appropriate owl profiles operate at scale in mission critical applications.
>
> 4. Adding redundant assertions that could have been inferred cannot make a dataset wrong. It does give more places to add wrong data, but if they are added automatically then that is less of a concern.
>
> 5. If inferable assertions are added to a dataset in order to avoid end systems doing inference themselves, it is probably necessary to add all inferences. This approach is used by triple stores that compute inferences at load time, and can cause the dataset to become larger. Data compression can give high compression ratios here, but the receiver still has to deal with more assertions.
>
> Simon
>
|