1. The Dream of a Global Network of Knowledge
Martin Doerr
Center for Cultural Informatics
Institute of Computer Science
Foundation for Research and Technology - Hellas
Amsterdam, Netherlands
November 17, 2011

2. Introduction Digital Libraries take on different forms and roles.
Initially collection management systems,
literature collections,
digitized resources
resource libraries (Perseus etc), on-line corpora
In addition, data services
scientific data collections
research systems (e.g., GIS integrated data)
“Metadata” Aggregation Services: a new paradigm using semantic networks
integrate diverse forms of information assets and pointers to them for the support of research and interested public
New grand challenges
Library access paradigm still dominates!

3. Library, Archive, Museum Information
The typical library contents: “The whole stories”, access widely solved!
Primary literature: Fiction.
Categorical: theories and hypotheses
Secondary literature (research results)
Facts brought into causal context
The typical museum information: “Museum objects rarely talk”
Factual documentation of
properties and context per object, references, classification
Highly heterogeneous,
About things taken out of original context, distributed over the world

4. Library, Archive, Museum Information
The typical archive contents: “The needle in the haystack”
Primary sources, “bits and pieces” (letters, legal documents, administration acts, images, scientific records).
factual, kept in the contextual sequence of creation, as by the creator or responsible.
kept due to mandate related to functions.
Similarly, library content itself: “What is in the book?”
parts of book content (citations!) as primary source of investigation
access: not much more than keyword search, if a digital form exists…

5. Epistemology of Integration
exhibit
Libraries
Museums
publish
document
features & context
provide finding aids
illustrate,
exemplify
using
are about
Books
refer to
Objects, Sites
contain narratives
made from
pub lish
refer to
document
manage
provide
finding aids
Archives
SMRs
primary Documents

6. Traditional Information Access
The traditional library task:
Collect and preserve documents and provide finding aids
The job is solved, when the (one, best) document is handed out. “All you want is in this document”.
The digital analogue: implementing “finding aids”:
Assumption: User knows a topic, characterized by a noun, or knows associations of a thing he knows it exists. Associations may be known properties, but not directly correlated to the problem to be solved (e.g. “organic farming” for “host-parasite studies”.)
Semantic interoperability is limited to the aggregation task: Metadata are mainly homogeneous (DC, VRA, etc.), the only challenge discussed is the matching of terminologies (KOS).
…still THE dominant global information integration paradigm

7. Problems
No support to learn from the aggregated sources, to retrieve by contexts,
e.g., Who was the employer of Donald Johanson when he found Lucy?
e.g., Which plant species are documented for the Black Sea coast for BC? (Critical climate hypothesis connected to detecting the Black Sea flood in 5600 BC)
e.g., Which resolution had Galileo’s telescope when he observed...
But understanding lives from relationships. Cultural information has complex relationships. Relationships may be categorical or factual:
Categorical (e.g., “smoking causes cancer”). : Richly exploited by Semantic Web technology. Use and integration limited to research results. Not useful for primary research itself.
Factual associations concatenate information assets to meaningful (“epistemic”) networks (“stories”): support context-based hypothesis building, cross-disciplinary search etc. (e.g. “John smoked with 20”, … ”. “John had lung cancer with 60”)
Knowledge of Factual associations is the “food” of scholarly research

8. What Can IT Do Now?
Access to categorical knowledge is well solved, if hypotheses have names: subject search, keyword search.
content management systems & search engines
Increasing account of structured categorical knowledge built in form of thesauri, ontologies (life sciences!)
access by terms and browsing broader/narrower terms
access by categorical relationships more rarely touched
Access to facts is idiosyncratic to diverse systems and limited to:
structured data services – no general access paradigm
KOS (authors lists, gazetteers)
“surfing and browsing” on the Internet or in Digital Libraries

9. What Can IT Do Now? New promises: Semantic Networks, Semantic Web
RDF Triple Stores
Open World Systems: Billions of facts under any number of schemata in one database
Linked Open Data (LoD): Thousands of triple stores to be accessed
Shift to metadata rich of facts
from Archives, Libraries, Museums, Digital Libraries
from research databases -> difference of data and metadata blurs
A global network of knowledge ? or a perfect intellectual chaos…?

10. (archive information?) (archive information?)
Semantic Networks
“…noble simplicity,
silent grandeur…”
(in a library)
Winkelmann’s
death
time
“LAOKOON”
(copy)
(in Vatican museum)
Winkelmann
writes….
Winkelmann
1755
Winkelmann
sees “Laokoon”
unknown Roman
(archive information?)
“LAOKOON”
Winkelmann’s
mother
Winkelmann’s
birth
(archive information?)
unknown Roman
copies “Laokoon”
Published
Inference
(in a library?)
Greece
Rome
Germany
space

11. 3 Grand Challenges
We need a rich, integrating global schema– a core and extensions of any depth
Con: impossible – everybody has his own conceptualization
Pro: CIDOC-FRBR work empirically proves opposite
“Knitting” the network : without co-ref resolution facts/triples do not connect
Con: impossible – automatic means limited, human labor not scalable
Pro or Con?: LoD
Pro: Human labor scales if massively organized
End-users need to query effectively large Triple Stores
Con: impossible to write ad hoc rich SPARQL statements, impossible to memorize hundreds of properties
Pro: use another, simple global schema for querying

12. A Global Schema: The CIDOC CRM
Developed by the CRM Special Interest Group of the International Committee for Documentation (CIDOC) of the International Council of Museums (ICOM)
Is an extensible core ontology of 86 classes and 137 properties describing the underlying semantics of over a hundred database schemata and structures from all museum disciplines, archives and libraries,
Extended by FRBROO, modeling IFLA’s FRBR, and soon FRSAD,FRAD, (RDFS integration with DC, Europeana EDM, ORE exists)
It is result of 15 years interdisciplinary work and agreement.
In essence, it is a generic model of recording of “what has happened” in human scale, i.e. a class of discourse.
By it we can generate huge, meaningful networks of knowledge by a simple abstraction: history as meetings of people, things and information.
An interlingua to transform, transport and merge information from most data structures with clear meaning.

13. Explicit Events, Object Identity, Symmetry
E52 Time-Span
E53 Place
E39 Actor
February 1945
P82 at some time
within
P7 took place at
P11 participated in
E7 Activity
“Crimea Conference”
E38 Image
E39 Actor
P86 falls within
P67 is referred to by
E65 Creation Event *
E39 Actor
E31 Document
“Yalta Agreement”
P14 performed
P81 ongoing throughout
P94 has created
E52 Time-Span

14. Data example (RDF-like form)
Epitaphios GE (entity E22 Man-Made Object)
P30 custody transferred through, P24 changed ownership through
Transfer of Epitaphios GE34604 (entity E10 Transfer of Custody, E8 Acquisition Event)
P28 custody surrendered by
Metropolitan Church of the Greek Community of Ankara (entity E39 Actor)
P23 transferred title from
P29 custody received by
Museum Benaki (entity E39 Actor)
P22 transferred title to
Exchangeable Fund of Refugees (entity E40 Legal Body)
P2 has type
national foundation (entity E55 Type)
P14 carried out by
Exchangeable Fund of Refugees (entity E39 Actor)
P4 has time-span
GE34604_transfer_time (entity E52 Time-Span)
P82 at some time within
1923 – (entity E61 Time Primitive)
P7 took place at
Greece (entity E53 Place)
nation (entity E55 Type)
republic (entity E55 Type)
P89 falls within
Europe (entity E53 Place)
continent (entity E55 Type)
Multiple Instantiation
TGN data

15. CRM Top-level classes useful for integration
E55 Types
refer to / refine
E28 Conceptual Objects
E41 Appellations
E39 Actors
refer to / identify
E18 Physical Thing
participate in
affect or / refer to
location
E2 Temporal Entities
E52 Time-Spans
E53 Places at
within

16. The CIDOC CRM The types of relationships
Identification of real world items by real world names
Observation and Classification of real world items
Part-decomposition and structural properties of Conceptual & Physical Objects, Periods, Actors, Places and Times
Participation of persistent items in temporal entities
creates a notion of history: “world-lines” meeting in space-time
Location of periods in space-time and physical objects in space
Influence of objects on activities and products and vice-versa
Reference of information objects to any real-world item

17. The Hierarchy of Participation Properties
P33 used specific technique (was used by)
P16 used specific object (was used for)
P142 used constituent (was used in)
P146 separated from (lost member by)
P29 custody received by (received custody through)
P96 by mother (gave birth)
P25 moved (moved by)
P22 transferred title to (acquired title through)
P14 carried out by (performed)
P23 transferred title from (surrendered title through)
P143 joined (was joined by)
P28 custody surrendered by (surrendered custody through)
P145 separated (left by)
P11 had participant (participated in)
P144 joined with (gained member by)
P99 dissolved (was dissolved by)
P12 occurred in the presence of
(was present at)
P13 destroyed (was destroyed by)
P93 took out of existence (was taken out of existence by)
P124 transformed (was transformed by)
P100 was death of (died in)
P112 diminished (was diminished by)
P31 has modified (was modified by)
P110 augmented (was augmented by)
P108 has produced (was produced by)
P123 resulted in (resulted from)
P95 has formed (was formed by)
Generalization
P92 brought into existence (was brought into existence by)
P98 brought into life (was born)
P94 has created (was created by)
P135 created type (was created by)

18. Schema Integration by Property Generalization
CIDOC
Conceptual Reference Model (CRM)
Access all data from any level
by CRM property generalization
Dublin Core
CDWA
MIDAS
Data
Few concepts,
high recall
Thing
Actor
was present at
Event
happened at
Special concepts,
high precision
Acquisition
used object
automatic
data export

19. Knitting the Network: Extracted Relations & Co-reference
Linking documents
via co-reference, not
hyperlinks!
Time-
Span
Primary link
extracted from
one document
Thing
Actor
Event
CRM:
global classification
of relationships
Deductions
Place
Fact
Integration
Johanson's Expedition
Discovery of Lucy
AL 288-1
Donald Johanson
Lucy
Cleveland Museum
of Natural History
Ethiopia
Hadar
Fact
Extraction
Documents,
Data,
Metadata 19 19

20. Co-reference Knowledge and Reality
symbolic level
(“vocabulary”)
same as
(data comparison)
same as
not same as
(direct negotiation)
(direct negotiation)
interpretion
(“speakers”)
real world
(“objects”)
M.Smith
born
M.Smith
born

21. Theory of Co-reference
A group of “speakers”(a database)” shares unique identifiers for a set of things. Another group “matches” their identifiers to mean the “same as”.
The transitive closure of “same as” – “not same as” exhibits “impossible worlds”, the only indication of false knowledge at the data level.
Ultimate knowledge is what the author meant by “her/him/it” – a part-of-speech, a database key, an occurrence of a name or URI.
Co-reference is primary knowledge, true research, not a “cleaning” issue.
Co-reference is more fundamental than schema integration: Supports integration without schema. Schema integration can be seen as co-reference problem.
Co-reference is more fundamental than Reference KOS: No description elements are needed. Reference KOS can help co-reference. Co-reference can be distributed!
Automatic “duplicate detection” is based on/ improved by co-reference,
“Negotiation with the speakers” is the ultimate confirmation = scholarly research.

22. Query “Friends of a Friend”
Co-reference Problem
Query “Friends of a Friend”
1. query
Content
has friend
Read output:
find “Kostas”,
guess
“Κώστας”
“Kostas”
input: “Martin”
2. query
Source 1
Content
has friend
input: “Κώστας”
“Κώστας”
output: “George”
Source 2

23. Join across sources by transitivity
Co-reference via Authority
Join across sources by transitivity
of co-reference
first match
local ids
query
Content
Authority service
input: “Martin” .
resulting link
ids L i n k t a b l e
“Κώστας” /
“Kostas”
match
Source 1
friend-of-a-friend
local ids .
Content .
second match
output: “George”
Source 2

24. Join across sources by transitivity
Curating Co-reference without Authority
Join across sources by transitivity
of co-reference
local ids
query
Content
make a
co-reference
local ids
input: “Martin” .
“Κώστας” /
“Kostas”
match .
Source 1
friend-of-a-friend
local ids
make a
co-reference
Content .
output: “George”
Source 2

25. Managing Co-reference Clusters
explicit initial “same as” (n-1)
explicit redundant “same as”
New link
connecting
clusters !
implicit link ( n(n-1)/2 )
reference occurrence
What happens ?
“M. Doerr”
“M. Dörr”
Authority files are good “attractors” of co-reference links, but do not solve co-reference !

26. A New Service: Global Co-reference Indices
Co-reference links should be persistent and public. Primary Co-reference links should be curated and preserved in local databases: “co-reference indices”.
Use NER and duplicate-detection algorithms to prepopulate co-reference indices. Use appropriate belief values for generated data.
Automated, global, distributed consistency control services are feasible.
Co-reference indices are much larger than ontologies, but not larger than search engines.
Mobilize general users and domain experts to enhance and verify co-reference information by social tagging to scale-up human labor and precision.
Install global supervision by open consortia setting the rules and doing central services.
Then the network may converge to consistent global knowledge.
Linked Open Data has no co-reference concept so-far. It will lead to a proliferation of URIs.

27. Last Problem: How to query 250 properties?
Humans think consciously in “compressed relations” (G.Fauconnier “The Way We Think”), in particular omitting events:
“What do we have from New Guinea?”
There are a few “Fundamental Categories” that partition our concepts (Ranganathan, “Who, When, Where, What..) and disambiguate most words e.g., a “”museum” is a “who”, a “where” or a “what”
If we implement a simple semantic network with few compressed relationships, we cannot integrate knowledge, because the intermediates are missing, and we cannot manage the immense number of redundant relations
If we implement a CIDOC CRM network, end-users cannot write queries
Solution:
Define a new “datamodel” of “Fundamental Categories” and “Fundamental Relationships” for querying only!
implemented as automated deductions from a CRM-based network

28. How to query with 250 properties?
Fundamental Categories:
Thing, Actor, Time, Place, Event (E2), Type
Fundamental Relationships:
has type /is type of
is similar to or same with
is part of (is member of) / has part (has member)
has met
from (has founder or has parent) / is origin, founder, parent, provider or creator of
had (=owns, keeps) / were owned/kept by at
refers to or is about / is referred by/ is referred to at
Relationships change interpretation depending on category of domain and range.

29. Thing is about Thing Path Expression
Following this schema, we have implemented over a hundred deductions such as:
Thing -> P130F.shows_features_of (0,n) OR P130B.features_are_also_found_on (0,n) -> {
E24.Physical_Man-Made_Thing -> P62F.depicts -> Thing OR
E24.Physical_Man-Made_Thing -> P128F.carries(0,n) -> E73.Information Object
-> P67F.refers_to-> Thing
D1.Digital_Object -> {L11B.was_output_of -> D3.Formal_Derivation -> L10F.had_input -> D1.Digital_Object ->}(0,n) L11B.was_output_of -> { D7.Digital_Machine_Event
-> P9B.forms_part_of(0,n) ->}(0,1) D2.Digitization_Process -> L1F.digitized -> E18.Physical_Thing }
It works!!!

30. Conclusions
After 50 Years of “Artificial Intelligence” research and 15 years “Semantic Web”, the Global Network of Knowledge is still a dream.
Today, we have the chance to lay foundations for global knowledge network(s!) with
a limited consistency,
with a tendency to converge to something more consistent
a limited common language,
a limited way to globally explore deep relationships
For that, we have to
Overcome intellectual barriers in conceptual modelling (“quick & dirty”, W3C “beliefs”, ignoring empirical scientific methods, political thinking, domain blindness)
Organize domain communities to curate collectively data and co-reference by new awarding methods
Invest in technology and methodology for a long data life-cycle by mapping, and transforming data “for ever”, as we do since antiquity…