Semi-automated metadata extraction in the long-term

GRANT AGREEMENT: 601138 | SCHEME FP7 ICT 2011.4.3
Promoting and Enhancing Reuse of Information throughout the Content Lifecycle taking account of Evolving Semantics
[Digital Preservation]
“This project has received funding from the European Union’s Seventh
Framework Programme for research, technological development and
demonstration under grant agreement no601138”.
Semi-automated metadata extraction
in the long term
Emma Tonkin, King’s College London
DPC Workshop, Belfast, Dec 2015

Structure of presentation
2
 Introduction to Pericles
 Layers of Metadata
 Sources of Metadata
 Time, space and data
 Semi-automated metadata as mitigating factor

Introduction to PERICLES
4
 Four-year Integrated Project (2013-2017) funded by the European Union
under its Seventh Framework Programme
 Promoting and Enhancing Reuse of Information throughout the Content
Lifecycle taking account of Evolving Semantics
 Two (or three) domains:
− Digital artworks, such as interactive software-based installations, and
other digital media from Tate's collections
− Material from Tate's archives
− Experimental scientific data originating from the European Space
Agency and International Space Station.

Model-driven approach
5
 Essentially all archives are based around some conceptual model of the
material held
 PERICLES applies formal models to describe
− Objects
− Entities associated with objects
− Broader community
 These models support processes such as appraisal and QA, and
consequentially functionality such as maintenance and actions taken for
sustainability
 A broad variety of models are under consideration: semantic (ontological)
models to formally describe objects; social network graphs to describe
community; statistical models to describe technology obsolescence...

Open Archival Information System
7
 OAIS reference model
− “conceptual framework for an archival system dedicated to preserving
and maintaining access to digital information over the long term“
-Lavoie, B. (2000). Meeting the challenges of digital preservation: The OAIS reference model
 OAIS-compliance
− adherence to ISO 14721:2003 or (now) ISO 14721:2012
− Specifies conceptual framework, functional model, information model

Descriptive metadata
9
 Supporting humans and machines
 Goal: interpreting data object
 Not always possible to automatically interpret data objects on any level
(some are fully opaque)
 Consider:
− 'Unstructured' natural-language texts, such as letters, books, articles
− Images of artworks
− Images of letters
− Recordings of audiovisual presentations
− Complex data files

Sources of descriptive metadata

Automated metadata extraction
11
 Popular view on indexing metadata:
− “the more, the merrier”
 Risks of low-quality metadata:
− Low accuracy on search and browse tasks; occasionally embarrassing
misinterpretations
 Benefits:
− Additional metadata can improve search indexing

How good is automated metadata
extraction?
12
 Varies significantly depending on the precise task and source material
 Automated metadata extraction tends to apply probabilistic (machine
learning) or heuristic approaches
 Machine-eye view:
− describe what is present
− Infer what is not based on:
 Knowledge base
 Comparison with other items
 Learning from training examples ('supervised learning')

Crowdsourcing metadata
13
 The 'phone a friend' approach to metadata generation
− Make material available to public
− Encourage them to annotate (example: social tagging)
− Examine the result
 The likely result:
− Some material extensively annotated; some descriptive annotations;
some formally structured; some personal ('cryptic')
− Some/most material receives no notice and is not annotated at all
 Mitigation: engineer more consistent coverage through, for example,
gamification (see Galaxy Zoo)
 Identify incentives that encourage public to contribute

Capturing 'live' metadata
14
 If the environment is accessible at the time of creation:
− Technical 'live' metadata may be captured
− Within Pericles, this is referred to as 'significant environment
information'
− Example: steps in creation, time of creation, contextual relevance of
other files…
 Another sort of 'live' metadata emerges from observation of behaviour of
those engaging with the data
− Interaction with search/browse interfaces (cf. information scent)
− Satisfaction with results
− Patterns of sharing and reuse (information diffusion on social networks,
for example)

Theoretical reach of information
16

Theoretical reach of information
17
Image source: S Korotkiy

 Receiving the signal is only the start
 Can we decode the signal?
− Technical decoding
− Practical comprehension
 Confounding factors in decoding metadata:
− Language
− Dialect
− Prerequisite knowledge
Practical reach of information
18

Language change: Time travel
20
 Language may be viewed as a complex adaptive system (Beckner et al,
2007)
− Made up of many tiny parts - people talking, writing, gesturing
− Adaptive, because we change our behaviour based on past
interactions
− Many factors influence its development: biology of perception; social
structure; experience
 Probabilistic processes underlie language change: collective experience
and eventual consensus

Example: Photogram (Getty Art &
Architecture Thesaurus)
21

The challenges of decreasing
accessibility

 Unfamiliar data
− Technical encoding – well-understood problems
− Challenges of internationalisation
 Unfamiliar texts
− Conventions and best practices change over time
− Coherence degrades long before it fails entirely (slower to read: takes
more effort: machines trained on modern texts are likely to encounter
issues with texts outside that timeframe)
 Challenges of unfamiliar artefacts
− There are many more questions that may be asked about an object: for
example, in the case of artworks, “artist's intent” may be significant
− Once lost, these are very difficult to infer
Understanding unfamiliar material
23

 Understanding unfamiliar material, though hard, is easier than finding it
 Separate processes:
− Recognising a term
− Identifying (generating) a term
 Recognition is faster and more reliable
 Why:
− Recognising a term: connecting term to concept
− Generating terms: search around a concept looking through large pool
of candidate terms for the one that might work best here
− Think yourself into the curator's shoes: what terms might they have
used for the concept that interests you, and why?
Term recognition vs generation
24

Semi-automated metadata as
mitigating factor

 Peirce: semiotic triad, relating symbol, object and interpreter
− Software agents: machine-level features (machine perception) –
words found in documents, colours, shapes or patterns found in
images…
− Human agents: perception; comprehension; application of relevant
knowledge; interpretation into a set of concepts; encoding
observations into terms
 Observing the behaviour of human agents throughout the lifecycle of the
digital object allows us to study change in manual interpretation and
encoding
 This permits us to characterise these patterns of change
 It also permits software agents to be brought into line with changing norms
Relating concept, feature, agent and
term
26

 PERICLES combines
− model-led approaches to data management
− data-led approaches to modelling and characterising the changing
environment and context(s) of reuse
 Approach acknowledges dynamical nature of system in which reuse occurs
 Downside: such an approach requires ongoing availability of material
(ethically) gleaned from observational data
− Consequentially, a closed archive or an archive that excites little
interest remains difficult to sustain, unless data is sourced elsewhere
 In conclusion, therefore, data-led approaches gain from joint infrastructure
and open data
Conclusion

Semi-automated metadata extraction in the long-term

More Related Content

What's hot (10)

Viewers also liked (6)

Similar to Semi-automated metadata extraction in the long-term (20)

More from PERICLES_FP7 (20)

Recently uploaded (20)

Semi-automated metadata extraction in the long-term