3rd European Semantic Web Conference (ESWC 2006)
Author: Raphael, Véronique, Lora
CWI participants: Raphael, Lora
# participants: around 300
A nice conference, with more attendants than the previous year (around 300 this time). What I will name a "Core European Semantic Community" was here, and I personally feel at home with them :-) 181 papers have been submitted, for a final 26% acceptance rate.
All the workshops, poster/demo proceedings are available electronically here.
Workshop URL:
http://gate.ac.uk/conferences/eswc2006/multimedia-tutorial/
Participants: #20 (CWI: Raphael, Lora)
Personal opinion: Overal disapoiting!
The signal analysis guy just enumerate all that they can extract ... but it is so
far away from what the user expect ! The SW guy talk only about text ! Michael did a nice talk, where
we can begin to see some integration between multimedia and SW technologies. But its talk was too
short. He will give us an extended talk during his visit in CWI July 13th-14th.
Globally, once again, we see perfectly how difficult it is to find people who has a vision and original
ideas of how to combine the user needs, the computer analysis results, and the SW technologies! The gap
is not only semantic ! It is between the communities, they don't talk to each other.
Stefan Rüger,
Imperial College, London, UK.
Multimedia information retrieval background. He is in the Scientific Board of K-Space.
Content-based search: query by example. Goals: querying with a sound, image, keywords, to retrieve
text, video, images, sounds, etc.
Automatic analysis: allows to automatically annotate images, with a bag of words. No interpretation,
the Semantic Gap! Ontologies can define these terms and turn them into semantically defined concepts.
He doesn't say a word how that can be done!
Related projects for bridging the Semantic Gap (K-Space, MMKM,
Mesh)
Polysemy: an image = 1000 words! Use approximation and some relevance feedback to narrow the
interpretation. Towards a recommendation system, very much like what Lloyd is doing in CHIP.
Quick state of the art on feature analysis, RGB colour model: what we can extract from images
(colour histograms, structure maps, texture descriptors, text index, localisation feature, shape
analysis, etc.).
Summary: many different low-level features can be computed, why did they do this?
Valentin Tablan,
University of Sheffield, UK.
Semantic Web side, but only text-oriented:-( Talk only about the
SEKT project
Semantic annotation, why ? Enable SW technologies to aggregate search results, summarize, inference ...
better retrieve through conceptual search!
Extracting Semantics from Text: based on Information Extraction (named entity recognition, relation
extraction, event detection, etc.). Not only bag of words but look for the structure in the text.
HLT tools could be used for segmenting multimedia content (split video into segments). Use of the ASR
transcript: not precise. Correct the ASR with the Web (good quality entities)!
Michael Hausenblas,
Joanneum Research, Graz, Austria.
The EU NM2 project:
creation of technologies for non linear interactive narrative-based movie production. Demo on the web
site.
Programme:
http://www.kbs.uni-hannover.de/~henze/swp06/swp06programm.html
Proceedings:
http://www.kbs.uni-hannover.de/~henze/swp06/swp.pdf
Lora has attended to this workshop so I went to the Industry Forum.
Programme:
http://www.eswc2006.org/industry_programme.html
Participants: #50
KeyNote speaker, from Microsoft Cambridge.
Driving force for digitization:
Tools for Health care: SNOMED resource. SNOMED-CT: can be represented in OWL. The DL reasoner and
classification procedure are useful to people who use SNOMED. The talk shows that they need the
expressive power of OWL in the health care domain!
Challenges: SW tools have difficulty with the scale of SNOMED-CT (around 400 000 classes!). They
are willing to adopt RDF, only if the tools prove to be a mature technology, scalable for their problem,
otherwise, they would go for traditional relational data model.
Creation of a drilling ontology in the field of drilling for finding OIL. Ontology-enabled active
search (normal keyword search + query expansion tactics using the concept specialization and
generalization, or the relations). OWL Full is used since they do not need the reasoning part. They
just want to have a common and shared vocabulary usable for annotating documents.
What they need ? Semi-automated annotation, using domain ontologies and visualization of annotated
knowledge. It seems that some collaborative work involving oil companies, suppliers, research and
supporting undustries are being setting up (Semantic Web Collaboration project).
DoCoMo: biggest mobile operator in Japan, 3rd operator in the world.
Voice and data market are saturated. Mobile operators need to offer new services to the users
(multimedia content, ubiquitous / contextual / personalized services). The machine should take care
of interoperability issues! Semantic Web is a key technology for them.
Innovative applications: ContextWatcher. There is a context ontology in OWL DL.
Example scenario: a very funny video, where we see a big line up for buying transport tickets in front
of an automatic machine. A guy has just 50 euros and the machine just refuses this amount of money. A
smart guy arrives with his phone, put his phone in front of the metro map (camera embedded) where he is,
click, put the cell phone on the destination, click, and the payment is done, the ticket goes out of
the machine or is written on the phone ! This is what they want to do ...
Conclusion: contextual intelligence, strong ties with the SW.
The time for discovery and development of a new molecule entity (NME) has gone from approximately 4 years in the 60's to 14 years now !
Knowledge Web talk ...
Use of Semantic Web technologies at AFP.
A News Agency is a Content Provider. It could be for general news, specialized, national perspective,
international coverage, different media, etc. AFP is one of the 3 worldwide news agencies. Work in
different languages (French, English, German, Spanish, Portuguese, and Arabic). Work daily, 1200
journalists, 200 photo-reporters, and 2000 stringers in 165 countries. Provide text, photos and videos.
Facts: we are all under a huge flow of information, more difficult for professionals, a jungle for
the layman (archives = 7M news stories per managed languages).
IPTC: founded in 1965. Standards are: NITF, IIM, NewsML, SportsML, NewsCodes
Link IPTC / W3C ? Is it us ?
Use of IPTC latest standards for our multimedia standards but: THIS IS NOT ENOUGH! We need new tools to
automate this enrichment in a new environment by implementing automatic taxonomy, ontology and
pre-indexing.
Challenge: A News Agency's buisnes has changed today to: adding value to content and
deliver to customers personalized and contextualized content.
Objectives: search (active mode), alerts (passive mode), browse (interactive mode) !
Cluster text-photo for a specific event, improve access to archives, etc.
Current problems: full text search = too noisy; sort/rating = inefficient; multi-criteria: not
mastered; keyword based. They would like to have Semantic Navigation ! Ex: how to get the story and the
associated images/illustrations/videos associated for a particular event ? Rely on the caption is not
enough!
They have developped a small ontology between the 5 top concepts: person, location, organization,
event, work.
Conclusion: traditional "push content" from the news provider to the customers is replaced by
interactive push/pull services: customers select and retrieve what they need. Have to implement
those our features on our different Media, but requirements for multi-media news make things difficult!
Semantics is a maze, we need HELP ! Don't forget, journalists have to work to produce these news
services AND the key will be in the links between NewsItems of the same nature or of different nature
leading to Multimedia content management as early as possible in the prodcution chain.
Question: what tools do you use ? Internal developped tools, that allow to manually annotate the news stories used by journalists.
Mondeca presentation. Mention of their 3 PhD thesis (and Laurence too :-)
Presentations of the Wolters Kluwer use case scenario.
My abstract is really ambitious, don't read it !
Which Semantic Web ?
See the conclusion here.
Goal: build a bridge between lightweight classification system and ontologies. 4 steps process:
1/ disambiguating labels (convert natural language labels to propositional DL labels, use WordNet);
2/ disambiguating edges;
3/ understanding classification alternatives;
4/ making classification choices.
Motivation: Many thesauri existing. SKOS: a standard RDF Schema.
Research question: can step-wise method be developed to assist conversion to SKOS?
All results, conversion tools available at:
http://thesauri.cs.vu.nl/eswc06/
Methods steps: 1/ Thesaurus analysis; 2/ Mapping to SKOS (table); 3/ Conversion program
(develop the code). This is more or less what George has done for the NewsML IPTC codes.
Use case: MeSH, 23000 descriptors, polyhierarchy
Lessons learned: things are easy to convert; others are not in SKOS (qualifiers, compound concepts,
etc.). They provide feedback to SKOS for adding support for compounds, for considering skos:Term
(multilingual, terms should have an identity)
Scientific problem: Ontologies often need to be built in a decentralized way, ontologies must be given
to a community in a way such that individuals have partial autonomy over them (need of local control
and local updates) and ontologies have a life cycle that involves an iteration back and forth
between construction/modification and use. While recently there have been some initial proposals to
consider these issues, they lack the appropriate rigor of mature approaches. i.e. these recent
proposals lack the appropriate depth of methodological description, which makes the methodology
usable, and they lack a proof of concept by a long-lived case study.
Proposal: distributed ontology modelling (former publication by Pinto et al.), tested
their methodology in real-life application, in the travelling domain.
Usecase: ontology to categorize documents, sharing and retrieving knowledge.
Social Network Analysis (SNA) emerged in the late 70's. Presentation of the benefits of SNA for
analyzing ontologies (Semantic Network Analysis, SemNA).
Eigensystem analysis for structural analysis of a graph.
Presentation of the TANGRAM project: http://iis.fon.bg.ac.yu/TANGRAM/home.html a learning environment for the domain of Intelligent Information Systems, and the TANGRAM application: http://ariadne.fon.bg.ac.yu/TANGRAM/app/
Two relationships between SW and personalization: personalizations technologies to enhance usability
of SW applications; SW technologies to enhance user-adaptive applications.
E-learning domain.
Different levels of semantic: what is semantic, how to represent it, the interrelations between different semantic levels/units, where to get the semantics from?
Basic idea is that different users have different needs and viewpoints about one domain, applicaion
dependent viewpoint. There are some big ontologies that gather different viewpoints, his idea is to
use the logic definition of concepts in the "big" ontology to create local (and smaller) consistent
ontologies. Multi-viewpoint reasoning, automatic selection of the features of the definition
to be selected to create the sub-ontology (features that are common to different concept definitions).
New logic notation needed.
Problem: a global ontology has to exist to generate the different viewpoints, as the method creates
the new subsumption relations from a global ontology gathering different viewpoints. Check
in the paper the evaluation and the usecases.
Linked to OntoRank
Conditional preferences to express statements, conditional preference base. Ranging the statements:
Yuppies express a 70% prference for red cars (not a yes/no statement).
Goal : visual tools for extending /reparing ontologies. Information Visualisation tool to visually
clean an ontology.
Method: spatial reasoning to clean ontologies.
They make a visual interpretation of DL definition of concepts, and do the cleaning using
mereotopological axioms: visual representation of the KB instances and checking consistency with
mereotopological axioms (?? the presentation was not very clear...).
Method to represent graphically an ontology, including cases where membership of an instance to a
class is uncertain.
Background of the speaker: PhD in Dublin, automatic annotation of Web Services. Uses two sets of features for mapping 2 ontologie's concepts:
DRAGO tool: distributed DL as a framework, distributed reasoning architecture for the Semantic Web
(see former paper).
Mapping concepts is often taken into account, but these are not the only "things" in ontologies
(more and more people are interested in relations!!!). Wedding can be a concept/a relation -> has to
be mapped! -> extend languages for ontology mappings to map ELEMENTS of ontologies, not only concepts.
Presented different usecases of KR that leads to heterogeneous ontology elements mapping (triple to a
concept, concept to a triple).
Defined the constructs: more general than/more specific than (equivalent is described by both
relations), applied to ontology elements. They modeled rules for concept to concept, role into role,
concept into role and role into concept mapping. Some heterogeneous mapping are not taken into account
in the DDL. Different ways of mapping heterogeneous elements taken into account in the paper.
More and more structured data on the Web: Swoogle 1.5M SW documents. Need and new possibilities
to map different sources of information.
oMAP ontology alignment tool, combines different classifiers, including one based on OWL definitions
of entities.
Inspiration: GLUE system: combine several specialized components for finding the best set of data.
Terminological classifier: based on distance metrics computed from WordNet (use WN synonyms).
... the rest was too fast for Véronique :-)
