- Semantic Gap for users (bits, chars, lists -> events/objects, concepts)
- Data and users: triangle
- Semantic gap exists in text too - search engines do little beyond string matching
- Semantic web tools help, much still to do
- Life - events, experiences and multimedia: eventweb
- Lots of multimedia - raises problems
- Multimedia semantics: many approaches
- Content-based: different model layers (data -> concepts)
- MPEG7 - describing multimedia data at different levels, description tools
- COMM: Core Ontology on Multimedia (based on MPEG7 and DULCE)
- LSCOM: Large scale concept ontology for multimedia
- uses MPEG7, around 1000 concepts
- aim to automatically identify these concepts in multimedia
- resulted in segmentation, tagging, annotations approaches
- But is this dealing with semantic gap? Or just refining tools on both sides of semantic gap
- semweb side: ontologies, rdf, ...
- content: concept detection with machine learning to build models
- Current approaches
- semweb effective at high level
- machine learning effective at low level
- Simple example
- what is a dog (dogs are animal most pictured?)
- can we make a model to recognise dogs?
- Contenxt: Content + Context
- represent context and knowledge
- current multimedia information retrieval
- silos - separate stores
- technology worked well with text (humans are the sensors, convert experience into symbols)
- Bridge: unified indexing through events
- objects and events (endurants and perdurants in Dulce)
- Events are good for dealing with dynamic situations and relationships
- Event representation
- information, experiential (text, data), temporal, spatial, causal, structural
- 1 dimensional space, isolated
- Link events together, eventweb
- Start with photos
- taking a photo
- group types of events
- Create upper ontology for events
- Composite events
- Tempero-spatial event model, map ontologies to it (e.g. wedding event ontology)
- Modern cameras
- event capture devices - EXIF
- merge with other (content, LSCOM, calendars etc.)
- Event based query
- mixed content/metadata search
- e.g. Florence outdoor day, Florence indoors
- Parse e-mail
- Audio experience
- Conclusion
- Semantic Multimedia web requires briding the semantic gap
Created
October 27, 2008 10:14
-
-
Save metade/20069 to your computer and use it in GitHub Desktop.
- Data web (linked data, deep data) versus Annotated Web (annotating existing web of documents, shallow web)
- Search Monkey focus on Annotated Web
- Crawling annotated web: rdfa, dataRSS (Atom + RDFa)
- Results: good feedback: more click through, publishers excited
- Data quality issues: complex semweb issues result in strange quirks
- Vocabularies: small coverage, different proposals/versions, not maitained, not designed for the annotated web
- Using metadata for search: how this data might be used to resolve hard queries (images of paris hilton)
- Wikipedia has dense link structure: average of 46 links per page
- Good anchor text on links: clear, brief
- Create a web ontology from Wikipedia by NLP
- Co-reference resolution is the challenge, e.g. Microsoft, Microsoft corporation, ">>it<< develops..."
- Use anchor text to extract synonyms - works nicely, but improvements
- Lead sentence: e.g. "Foo is a bar", >50% of articles have an "is-a" lead sentence
- Important sentence: work out highly related links (e.g. just a general term ("multimedia" versus a specific term "Microsoft Windows")
- Association Thesaurus - e.g. (sports => basketball, football)
- PFIBF: short path between linked page and current page => strong relation between them, e.g. Microsoft => Microsoft Windows (strong), Microsoft => Multimedia (weak) *For best results, use a mix of mehtods: article title, frequent pronoun, synonym detection
- Social tagging: good source of messy social data
- Sense based tagging
- Tag Disambiguation Algorithm (TDA)
- Tagpedia
- I didn't get how tags in delicious are mapped to tagpedia concepts... :(
Guillaume Ereteo: A state of the art on Social Network Analysis and its applications on a semantic web
- Sociograms [Moreno 1993]
- Community detection: hierarchical (agglomerative, divisive), or based on heuristic
- Centrality
- Social Network Analysis typically reduces rich relations to untyped graphs
- Ohloh: Open Source Network
- Making Web 2.0 data available as Linked Data
- Wants to set up W3C working group on this
- Ontology for modeling argumentation on social networking sites
- Could model comments on /music in this way?
- Understanding Advertising
- Crawling forum data using SIOC, then basic NLP on post/comment title(?)
- Provides nice visualizations and query mechanisms
- Full NLP on post/comment content will be investigated in future work
Alexandre Passant and Yves Raimond: Combining Social Music and Semantic Web for music-related recommender systems
- Use Semweb techniques for linking up Web 2.0 services: SIOC, Linking Open Data
- MOAT - Meaning of a Tag
- Music Recommendations:
- Collaborative based filtering (i.e. user) - long tail (biased towards popular artists)
- Content-based analysis (http://mufin.com) - no long tail issue, lack of cultural context
- Hybrid approach works best: use rich linked data
- Example recommendations:
- artists a friend is interested in
- content based query using http://dbtune.org/henry
- mixed query
- Last.fm event + geolocation on DbPedia Mobile
- GNAT + GNARQL
- Music artist recommendations using DBPedia: http://apassante.net/home/2008/10/musicrec
- Future work
- Find a path from a user (web resource) to the recommendation (web resource) that goes nearby other constraints
- interests, personal music collction, listening habits, friends etc
- Interruption management study on mobile phone
- Call filtering strategies
- User policy: e.g. When I am in a meeting, my phone will ring if a friend of mine calls me
- Use semweb to express this
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment