Multimedia:Meeting in Paris/Notes/Metadata

• Saturday, November 7, 2009 - room 2
• Moderator: Duesentrieb
• Note-taker: Brion

Structured data!

• metadata about pages/revisions
• metadata about media files
  • from MediaWiki [categories etc]
  • embedded
    • inherent: file size, page count etc
    • add'l data [EXIF, ICTP, XMP]
  • supplied on description page
  • external data
    • For batch uploads... would these come into the page?

What's the authoritative source?

• Wikitext as primary source? [editable]
  • -> cache in index [searchable, queryable]
  • -> can override other sources [exif, file-inherent data]
  • External (read-only) authority databases?
    • (Do we need/want such or just import things per-source when doing batch uploads?)

Structured data use cases

• media file metadata export for re-use
  • api, embedded
    • [Modify EXIF or XMP data to embed extra stuff?] [Recommend pushing that to later -brion]
    • orig artists, repo, institution...
  • Including info from authority files
  • using structured data about things+files for search and research

Let's try to keep this minimal?

• Real use cases?
  • search
    • license, keywords etc
  • reuse
    • -> Generate byline
      • author(s)
      • license
  • Main fields... (many of these can be tied to Dublin Core fields, so import/export for academic collections can hit common denominators)
    • Fixed data formats
      • license
        • license features
        • dc:rights, creative commons extensions
      • geo coordinates
      • file format
        • jpeg vs tiff vs ogg
      • media type
        • drawing vs photo vs map
      • filesize / resolution
      • content language
        • dc:language
      • source collection + identifier
      • source URL
        • dc:source
      • creation date
        • dc:date
    • Freetext [can have multilingual variants]
      • author
        • dc:creator
      • description
        • dc:description
      • source
        • dc:source
      • "other fields" from foreign source
        • keywords
          • (problems with consistency? controlled vocabularies with some sources, but don't match up reliably)
        • people shown in picture
        • etc

How to source/edit the stuff...

• source fields from file inherent/EXIF
• scary templates
  • Nesting is required for extracting triples sanely...
    • {{Information|desc={{en|hello}}{{de|hallo}}}}
      • We _think_ we can break this out
      • "template" "Information@desc/en" "hello"
      • "template" "Information@desc/de" "hallo"

Normalization of property specs/extraction most important

Normalization of property values can be left for later

Metadata extraction spec

Possibility...

• Define for <dc:description> sourcing from:
  • "template" "Information@desc/<lang>"
  • "exif" "Description"
• Gives us these trips:
  • "File:Hallo.jpg" "dc:description@lang=de" "hallo"
  • "File:Hallo.jpg" "dc:description@lang=en" "hello"
  • "File:Hallo.jpg" "dc:description" "Created with The Gimp"
• also other sources of info available:
  • revision data / uploader / editor
  • external authority data [rdf triples given from bulk uploader?]

Possibly spec as a special page or a mediawiki:horrid or config file. ;)

Whatever consider that it may change; so we need to be able to re-run metadata extraction in bulk.

Backend storage & search...

• mysql? (page/field/keyword trips ... or searchindex?)
• lucene? (for each page page: field/keyword pairs)
• specialized triple store? (sparql)
  • potentially really cool searches possible!
  • recommend leaving this to later - third parties cover this area well like dbpedia and freebase
• Look at what Semantic MediaWiki does for its storage
• Minor note: with RDF triples, subject can be tricky (page vs image vs subject of image)
  • SMW -- talk page's subject is the page, so you can put data about a page there ;)
  • This seems simpler

Horrible questions: licensing?

• metadata licensing is a scary world out there...
• Let's pretend the problem doesn't exist for now ;)
  • [but database protection stuff in Europe may need to be looked at... but we think that hits on our side mostly]

Points of action!

• Good news: exif extraction is already here
• Work: template extraction sounds feasible and not too insane
  • [brion: see about a proof of concept to make sure this works like we think :D]
• Work: need to specify an extraction spec language
• Work: core DB storage + query/dump export
• Work: need to add interface to add the fields to search backends
  • Lucene <- key for live
  • MySQL fulltext <- good for third-parties

Other considerations

• sometimes metadata is about *subject of image* instead of *image*. Ughhhhh?
  • at minimum, need to be able to spec for the subject... maybe
• Location: for paintings -- have loc of the painting or the subject of the painting?

Export methods

• raw RDF via API
• XML RDF in special:export and dump XML
  • ^if included in dumpBackup and OAI export, our Lucene indexer can easily ingest that!
• RDFa embedded in HTML pages