> -----Original Message-----
> From: Z39.50 Next-Generation Initiative [mailto:[log in to unmask]]On Behalf Of
> Fabio Simeoni
> Sent: Wednesday, April 21, 2004 3:51 AM
> To: [log in to unmask]
> Subject: Re: metasearch
>
>
> Peter,
> thanks for your comments.
>
> > Semantic variation is a fact of life of any metasearch
> > activity. Even harvesting does not reduce the semantic
> > differences of the underlying documents.
>
> the point is not whether harvesting is immune from problems of semantic
> interoperability, for sure enough successfully exchanged data must still
> be uniformly understood.
> the point is *how much* semantic alignment between the mutually
> autonomous parties of a federation is required by harvesting and *how
> much* by distributed searching.

I would suggest the amount of semantic alignment between parties to
harvesting is much higher than for a metasearch. The harvesting is done "by
agreement" - even if the agreement is the one sided one of 'we want that
data because it fits in some way with the rest of what we have'. Thus the
harvesters have decided that the incoming data 'makes sense' in some fashion
within the context of their greater (harvested) collection. Strict semantic
harmonisation (say re-indexing through a uniform controlled vocabulary) will
be one extreme, the other will just be "its medical stuff". For
metasearching the sources are in no way aligned, they are chosen at query
time often on the flimsiest of evidence that there is some alignment. Thus
my favourite Mercury-the element and Mercury-the planet can easily be a
problem in a search of a very small number of scientific sources.

> you will agree that the interop issues raised by metadata disclosure are
> a subset of those raised by the specification and implementation of a
> full-fledged end-user service over that metadata.I cannot see any
> agreement required for harvesting which is not also required by
> distributed computing, regardless of the particular protocol
> implementations of the two modes.

I do agree.

>  In contrast, I can see a lot of
> specification and implementation details in full service provision which
> may compromise iinterop and which do not surface when parties
> participate only of metadata disclosure and service provisions has all
> the good properties of local implementations.
> It seems to me hard to deny that the complexity introduced by full
> service provision over metadata disclosure introduces more potential for
> semantic disalignment.

I disagree that the implementation details are the problem. Most of what is
required in the 'full service provision' (which is often not very "full")
for metasearching are mechanical details (protocols, addresses, syntax,
authentication, etc.) and have little to do with the semantics of the search
universe.

What does cause the problem is that there is absolutely no alignment on the
content of the sources or their treatment.

>
> > An aggregated
> > document set will be re-indexed and the indices will expand
> > to include the union of the disparate originals or the
> > originals will all be fitted into the controlled vocabulary
> > of the aggregation. The former only gives a mechanical
> > advantage and no search advantage (in fact a disadvantage as
> > in the original databases - which are presumed to be
> > semantically optimised in some way - the search can be mapped
> > to the specifics of the individual database semantics) the
> > latter fits disparate documents under 'awkward' terms and
> > reduces the specificity of those terms as measured by the
> > cohesiveness of the documents they index.
>
> I cordially disagree here. Participation to a community based on
> harvesting implies agreement on the semantics of the exchanged metadata,
> the looseness of which is only predicated on the metadata model of
> choice (loose for DC, tighter for MARC or IEEE, etc) and the cataloguing
> QoS at participating parties. Both observables do not disappear with
> distiributed searching and are in fact orthogonal to the architectural
> issues we are discussing. At most, harvesting encourages metadata
> enhancements in correspondence with 'weak' participations, whereas
> similar mediating services are much harder or altogether impossible to
> accommodate in real-time by services based on distributed searching.
>
> In particular, I cannot see the value you see in the 'specificity' of
> search terms with respect to local metadata collections, when the
> semantics of the original search terms is defined by the remote
> metasearch client and its users. The local interpretation of a non-local
> search term -- which may appear more precise from that local perspective
> --  can only but contribute to a loss of precision outside that local
> context, i.e. in the merged results, and of this fact it seems to me the
> literature provides abundant experimental evidence. Rather, the very
> value you attribute to local indexing policies applies intact at the
> metasearch client side, which offers a local service against originally
> remotely distributed data. In particular, indexing and other QoS aspects
> of the metasearcher -- whether poor or good -- are of concern to the
> metasearcher and the metasearcher only
>
We already have a situation where two contexts have to be considered for the
semantics of the search whether it is in a metasearch environment or a
harvested environment. One is the searchers context, the other is either the
single context of the harvested collection (assuming there is some amount of
semantic alignment), or the multiple contexts of the different metasearch
sources. So the mis-alignment of the searchers context with whatever the
target context(s) is/are is always present. You could argue (and you have)
that the alignment is better with the single harvest context as that context
itself is more uniform than the diverse (even divergent) contexts of the
metasearch sources. On average this should be so (or else the searcher has
chosen the wrong place to seek). However the very diversity of the
independent sources means that some will be better aligned with the search
and some worse than the 'average' context of the harvested collection. This
leads to the question of whether the searcher is better served by the
provision of a few 'really good' answers and a number of 'really bad' ones,
or a set of 'average' answers.

All of this, rather theoretical, discussion assumes some serious uniformity
of excellence and creating the data and services in the first place, and on
the part of the searcher in specifying the search. And, unfortunately, most
research and empirical evidence shows the real world has implementation
'noise' which outweighs the above effects. That doesn't mean we shouldn't
try.


Regards  -  Peter
Dr Peter Noerr
Chief Technical Officer
Museglobal, Inc.

tel: +1 801 208 1880
fax: +1 801 208 1889
cell:+1 801 910 4912

[log in to unmask]
www.museglobal.com

P.S this is getting somewhat out of context for Z39.50 (or at least overly
broad in scope) so maybe we should start another thread or take it private.


> My point here is not to criticise distributed computing model, of
> course, but try to understand the model invariants that limit its
> suitability to metasearching in the federation.
>
> regards,
>
> fabio simeoni
>
> ***************************************
> Fabio Simeoni
> Senior Research Fellow
> Centre for Digital Library Research (CDLR)
> Computer and Information Sciences Department
> University of Strathclyde
> Tel:0044-(0)141-5485855
>
>
>
> > I would argue that the user gets better retrieval from a
> > metasearch across a number of smaller specialist databases
> > than from the same documents aggregated in one database. The
> > variable is the accuracy of mapping the user's search to the
> > capabilities of the individual database. If it uses the
> > semantics of the database (a 'term list for example) then the
> > mapping is improved for each such database and the overall
> > result is improved. (I am talking mostly precision here,
> > given real life volumes recall is usually not the problem.)
> >
> > Peter Noerr
> >
> > -----Original Message-----
> > From: Z39.50 Next-Generation Initiative
> > [mailto:[log in to unmask]]On Behalf Of Fabio Simeoni
> > Sent: Tuesday,
> > April 20, 2004 10:28 AM
> > To: [log in to unmask]
> > Subject: FW: metasearch
> >
> >
> > >Currently it is not easy with SRU/W to broadcast the same
> > query to many
> > >SRU/W servers because one has to take into account all the
> > differences
> > >between different servers.
> > >
> >
> > >> Definitely true - this due to the fact that SRW is a client-server
> > protocol, whereas metasearch broadcasting as you are
> > describing is more a grid computing task.
> >
> > One wonders if such problems of semantic interoperability
> > between server-side implementations are an invariant of the
> > distributed computing model underlying SRW/SRU rather than
> > the protocol itself, Z39.50 and indeed any other
> > implementation of that model. In particular, one wonders
> > whether is the very requirement that servers align their
> > interpretation of service provision (here searching) that
> > introduces assumptions about the autonomy of participating
> > parties which cannot be advanced in a large-scale federated
> > environment. In this sense, harvesting as implemented in
> > OAI-PMH relieves servers from any other semantic alignment
> > beyond metadata format and thus suffer considerably less from
> > these problems. With distributed computing, and thus with
> > SRW/SRU, each party which contributes its data must also
> > participate of service provision and this amplifies
> > requirements of mutual consistency and the problems these
> > requirements raise in a loosely-coupled environment.
> >
> > regards,
> >
> > fabio simeoni
> >
> > ***************************************
> > Fabio Simeoni
> > Senior Research Fellow
> > Centre for Digital Library Research (CDLR)
> > Computer and Information Sciences Department
> > University of Strathclyde
> > Tel:0044-(0)141-5485855
> >
>