Dear Maximilian,
I had been a bit shorthand.
I referred to a knowledge graph theoretically, as CRM instance, without
a quad feature, or looking at it as being a Named Graph. Of course
current systems have these features built in.
The reason being, that only separating the Graph from the data about the
Graph, be it quad or Named Graph or on paper, we can understand the
logical problem.
If we look at the "metagraph" that is about a graph, it is again a
simple graph. It does not have a provenance. Its provenance
could be another graph about this graph, and we are still at the same
point.
This means, that ultimately, we have to rely on curators of information,
that mediate trust in it. By "trusted source",
I do not mean a logical "TRUE", I mean that someone gives me information
he himself trusts because of further knowledge,
and I trust that it is his best knowledge, regardless whether it turns
out to be wrong later.
We may talk about degrees of trust, and plausibility, but that are only
variant of trust. Trust is not a statistical probability.
If I get an image of a Duerer, I have no bl.. idea if it is a fake or
not. I would not assume that it is a fake, if I can trace its
provenance to a museum. I would check the Website, if it suspicious or not.
The epistemological chain will go back to primary evidence. I assume
that the museum knows how to connect the image to the original paper,
that the paper has a credible chain of owners back to Duerer, or has
been examined by analytical methods to be genuine. Arguing about my
image, I imply all this good practice was applied. Even if this is all
described in the document, without the connection to the human curator
there is no knowledge in it.
So, point b) means, the solution to the descriptive chain is that at the
end there is a trusted source, i.e. a source I have no good reason to
question, which may make errors, but follow a good practice of knowledge
creation.
Your point c) below looks only at one level of metadata. If, e.g., I
have an old physical book from 15th century in a library, I would rely
on the author information and date in it. But I I would simultaneously
rely on the physical book exhibiting features of that are compatible
with the age and I would rely on the library curators not having
smuggled in fakes over time. I would compare with other copies etc. Even
if a fake is detected, I would rely on this human/material provenance to
hypothesize how a fake could have come in. It will still be a form of
(less;-))trusted source.
So, my point is, the trusted source can be directly responsible for the
content. It may only be responsible for the metadata, or the
metametadata, etc.
Each indirection is expected to rely on good practice and hence provide
trust in the level below, i.e.,
the metametadescription about the metadescription, the metadescription
about the description. Each level should describe where its confidence
comes from.
At the end, there is always a (living) human being connecting data with
the real world. In our information systems, we must keep that curation
chain. No other information can save us from fake news, once they can be
spread and multiplied without limits.
OR not?:-)
Best,
martin
On 3/24/2018 6:15 PM, Maximilian Schich wrote:
Dear Martin,
My "recommendation" was just putting into question an aspect of
Florian's suggestion, and not meant to replace it in a final way.
Regarding your points: The practical cases I am familiar with would
use the E13 on the whole triple, i.e. the link/property-type including
a specific source node and a particular target node. This means either
the triple is stored as a quad, or the triple carries an ID or
address, so one can refer to it. TEI standoff markup would be another
practical example.
As an art historian/archaeologist and hopeless class-conceptualist, I
do not believe in trusted sources. Everything comes with a probability. :)
a) Self-description is of course never perfect, yet depends on the
density of information: A signature, as in "Martin performed
Attr.Ass.512" or "A[lbrecht] D[ürer] fecit" is only one form of
(self-)descriptive information, which is as good or bad as anything
else, internal or external. Of course, it is better to see Dürer in
detail or to hear Anne-Sophie Mutter actually play, rather than
relying on a verbal statement of attribution.
b) I don't understand: Any graph-like description of a graph
constitutes a forest of graphs with the original graph, i.e. a
disconnected graph that contains the description of itself. If we
generalize that statement to symbolic representation, you are in
essence saying description is impossible.
c) I think in most cases "description within a set of information
about its provenance" is the only thing we have. There is no default
up the next source of source. Evolutionary biologists, material
scientists, art historians working on renaissance drawings, and
scholars of ancient manuscripts all rely on hysteresis, i.e. history
of the object contained within the object. There never was a
comprehensive DNS for organisms, manuscript fragments, or paintings,
and there never will be. For the same reason we need to embed
provenance in our data sets. Probably we should even block-chain it in
with enough information, so we don't have to rely on simple signatures.
*/To make my case much more simple and short: "All of Wikipedia
includes the full edit-history". /*This is how it is produced, and how
it should be analyzed. The same standard should apply to any cultural
heritage data set. Any other practice would be like citing monographs
without pagination. This is why E13 is really central, particularly in
multi-authored data sets.
On 2018-03-24 15:01, Martin Doerr wrote:
Dear Maximilian,
This makes sense to me, but I do not agree with your recommendation
as a general rule.
There is a fundamental epistemological problem, which has nothing to
do with quantitative evidence. The latter,
by the way, cannot detect an endless recursion anyhow, because people
would break it.
The ramifications of this breaking are huge, as can be seen by your
answer.
Let us start with a more fundamental construct, a simple
CRM-compatible "knowledge graph" with one attribute:
"Martin" has residence "Heraklion".
Using an E13,
"Martin" performed "Attr.Ass.512". has type: "has residence"
assigned: "Heraklion"
assigned to: "Martin"
now reading it, I know the knowledge graph wants to make me believe
who said "has residence", but I do not know, who introduced these
three more attributes.
So, I reify the three new attributes with 9 more, and I am still not
wiser, nor will I be with any other iteration of it.
If I know that the knowledge graph *was produced by Martin as a
trusted source as a whole*, I do not need the E13 in it.
Then, I can add metadata to the whole knowledge graph, e.g., as a
Named Graph or "context" or on paper etc. , but I am
still in the same situation: who produced these metadata, are they
trusted?
Hence, I conclude three things:
a) There is no completely self-descriptive information. The trusted
source ("sender of the message" in Claude Shannon's sense) lies
outside the information unit. It must always be the default. In order
to characterize the default, we need semantics different from E13.
b) It makes no sense to describe the default in the graph itself.
c) Any description within a set of information about its provenance
pushes the level where the default applies up to the next source of
source. Hence, if a team decides to register actions of their
members, the team as a whole pushes the default up to the trust in
the registration, rather than in the primarily registered. I see all
you examples as practices of this kind. There may be many reasons to
do this, but in other cases also not to do it.
Such a rule cannot replace understanding the basic epistemology,
which is always the same.
Does that make sense:-)?
All the best,
Martin
On 3/24/2018 12:10 PM, Maximilian Schich wrote:
Dear Florian and all,
Based on quantitative evidence, I'd object to the following to part of your
suggestion:
"This fact must not individually be registered for all instances of properties
provided by the maintaining team, because it*/would result in an endless recursion/* of
whose opinion was the description of an opinion."
=> This would only be correct if the maintaining team would add additional E13 Attribute Assignments to their own E13 statements. Otherwise,*/in practice, the data would (a) more or less double, plus (b) a
non-exploding truncated tail of additional E13 correction statements/**/, where the maintaining team corrects itself./*
=> Example for (a): In large data sets such as the "Census of Antique Works of Art and
Architecture" the "record history" approximately doubles the data set as a whole. Note: The
Census "record history" is the place where the maintaining team records their own E13-like/attribute
//assertions /(aka/assertions of database record authorship/). It is important to point out that the record
history, where an internal database curator implicitly claims authorship for say an artist attribution in the
Census, is conceptually in no way different from an external author providing a differing opinion (both usually
have PhDs in art history). Ergo there are two default cases: (1) The internal database curator claims authorship
for a*/direct assertion/* via a single E13 Attribute assignment in the record history; (2) The internal
database curator claims authorship for a*/cited assertion/* via an E13 attribute assignment in the record
history on top of the*/original assertion/* that connects the stated opinion to its external source via another
E13 attribute assignment.
=> Example for (b): In large data sets where the multiplicity of opinion is
recorded, the number of competing assertions including both record history and
external opinions, is usually characterized by a tailed frequency distribution*.
This usually means in practice that the data set stays in the same order of
magnitude relative to the case where the maintaining team decides to follow one of
the alternative assertions.**
* The frequency distributions would look similar to Schich 2010 "Revealing
Matrices" Fig. 14-8. Indeed, my pre-publication version of this figure had a column
for the record history, not included in the article, as the networks were too large for
the preceding figure.
** Yes, we should expect some "assertion cascades" to be exceedingly large, but we can
also expect the median cascade length being very short, between 1 and 2 in cultural heritage
databases based on personal experience, and still short in very large scale cases, such as
spreading rumors on the Web (cf. Friggeri et al. 2014 "Rumour cascades" Fig. 5).
=> The recommendation, in my opinion, should be:*/By default, the maintaining team should establish authorship by
adding an E13 Attribute Assignment to each assertion in the data
set. Yet, the maintaining team should _only_ add an E13 Attribute
Assignment to their own E13 Attribute Assignments in the case of
discernible modifications, updates, or corrections. To avoid comment
cascades, such alternative E13 statements should be done in /**/*/parallel(!) not recursively.***/* This recommended procedure
establishes a record history and granular ability to cite data set
contributions by author, yet also avoids a recursive explosion of
E13 statements./*
*** Parallel, means E13 statements in the internal record history should never
be about statements in the record history itself. This can easily be maintained
with users being logged in or recorded via IP and timestamp. Working example:
The Wikipedia edit history.
Hope this makes sense.
Best, Max
--
--------------------------------------------------------------
Dr. Martin Doerr | Vox:+30(2810)391625 |
Research Director | Fax:+30(2810)391638 |
| Email:mar...@ics.forth.gr |
|
Center for Cultural Informatics |
Information Systems Laboratory |
Institute of Computer Science |
Foundation for Research and Technology - Hellas (FORTH) |
|
N.Plastira 100, Vassilika Vouton, |
GR70013 Heraklion,Crete,Greece |
|
Web-site:http://www.ics.forth.gr/isl |
--------------------------------------------------------------
_______________________________________________
Crm-sig mailing list
Crm-sig@ics.forth.gr
http://lists.ics.forth.gr/mailman/listinfo/crm-sig
--
*Dr. Maximilian Schich*
Associate Professor, Arts & Technology
Founding member, The Edith O'Donnell Institute of Art History
*/The University of Texas at Dallas/*
800 West Campbell Road, AT10
Richardson, Texas 75080 – USA
US phone: +1-214-673-3051
EU phone: +49-179-667-8041
www.utdallas.edu/atec/schich/ <http://www.utdallas.edu/atec/schich/>
www.schich.info <http://www.schich.info/>
www.cultsci.net <http://www.cultsci.net/>
Current location: Dallas, Texas
_______________________________________________
Crm-sig mailing list
Crm-sig@ics.forth.gr
http://lists.ics.forth.gr/mailman/listinfo/crm-sig
--
--------------------------------------------------------------
Dr. Martin Doerr | Vox:+30(2810)391625 |
Research Director | Fax:+30(2810)391638 |
| Email: mar...@ics.forth.gr |
|
Center for Cultural Informatics |
Information Systems Laboratory |
Institute of Computer Science |
Foundation for Research and Technology - Hellas (FORTH) |
|
N.Plastira 100, Vassilika Vouton, |
GR70013 Heraklion,Crete,Greece |
|
Web-site: http://www.ics.forth.gr/isl |
--------------------------------------------------------------