These slides are from my seminar to the University of Reading Department of Meteorology, November 2013. They contain a (hopefully not very technical) introduction to the concepts of Linked Data and how we are applying them in the CHARMe project (http://www.charme.org.uk). In CHARMe we are using Open Annotation to connect users of climate data with community-generated "commentary information" that helps them to understand a dataset's strengths and weaknesses.
The slide notes contain some helpful context, so you might like to download the PPT file!
The slides are licensed as "Creative Commons Attribution 3.0", meaning that you can do what you like with these slides provided that you credit the University of Reading for their creation. See http://creativecommons.org/licenses/by/3.0/.
DSPy a system for AI to Write Prompts and Do Fine Tuning
In search of lost knowledge: joining the dots with Linked Data
1. In search of lost knowledge: joining
the dots with Linked Data
Jon Blower
j.d.blower@reading.ac.uk
Department of Meteorology
Reading e-Science Centre
National Centre for Earth Observation
University of Reading
With thanks to all CHARMe project partners!
5. How do you fill in the blanks?
• Hope the documentation contains pointers
• … to the right versions of things
• … and using consistent terminology
• In restricted communities, we can gather all
information into a central location and
harmonize it
• OR we hope that a Google search throws up the
right results
9. Lots of useful negative results, which aren’t published
Nearly 1/3 of positive results are false
1000 hypotheses
100 are true
900 are false
Test accuracy 95%
95 true
positives
5 false
negatives
45 false
positives
855 true
negatives
12. “There is a spurious decline in low-cloud fraction in
the ISCCP cloud database due to the viewing angle.
It’s in the literature, but you might not find it if
you’re not specifically looking for it. For example if
you’re using the data for model validation you may
not spot the problem.”
(Claire Barber, paraphrased)
13. Consequences
• We use the data that are most easily
available, not necessarily the best
• Constant re-discovery of the same issues
• Very hard to share information outside
communities
19. Fingers crossed!
• We hope that everything has a document
written about it, or we have nothing to cite
• … and we hope someone writes a new
document when “it” changes
• We hope that we are citing the right
document (i.e. the most authoritative etc)
20. “If … the Web made all the
online documents look like
one huge book, [the
Sir Tim Berners-Lee
Semantic Web] will
make all the data in the
world look like one huge
database.”
Photo by Susan Lesch, from Wikipedia
21. Why is the Semantic Web different?
• Focuses on things, not documents-aboutthings
• Links things together in a meaningful way
• Information is readable by computers
• Here’s how it works…
22. 1. Give things unique identifiers
Must be globally unique and persistent
http://www.reading.ac.uk/people/jon.blower
might represent me
http://nasa.gov/instruments/MODIS
might represent the MODIS instrument
–(these are illustrative, not accurate!)
23. 2. Express relationships between
things
The key concept is the triple
subject predicate object
E.g. “MODIS is carried by the Terra satellite”
all three things need to be unique identifiers
(we don’t use plain English!)
24. Examples of realistic triples
(identifiers are illustrative, not necessarily correct!)
http://www.reading.ac.uk/people/jon.blower
http://xmlns.com/foaf/0.1/knows
http://www.durham.ac.uk/staff/ed.llewellin
http://dx.doi.org/12345.678910
http://purl.org/spar/cito/citesAsDataSource
http://www.badc.ac.uk/datasets/sst
25. Aside: how to use the Semantic Web
to ruin jokes
“A hundred kilopascals go into a bar”
“100000”^^unit:Pascal
owl:sameAs
“1”^^unit:Bar
(Using OWL and QUDT ontologies)
26. 3. Publish your triples
• Can be as simple as putting a document in the
right format on your website
– (triples can be expressed in many formats)
• … or as complex as publishing a multi-billiontriple database
– E.g. Ordnance Survey
• Either way, your data become part of the
Web, not just on the web
27. 4. Profit!
• Everyone can publish their own “view of the world”
– Don’t need a single data structure “to rule them all”
• Use common identifiers and common vocabularies to link between
communities
• Different vocabularies can be mapped to each other
– E.g. map CF standard names to GRIB codes
• Gives a means to record provenance of information
– “Direct line of sight from decisions to data” - NASA
• Reasoning engines can traverse the graph of links and discover new
information
28. Example: SemantEco
Wildlife observations
Ecosystem impacts
Administrative areas
Pollution data and policies
“Find me all the sites where chloride pollution
exceeds the local policy limits”
“Show me the species over time in this region”
29. So what’s “Linked Data” then?
• The Web is the “web of documents”
• The Semantic Web is the “web of data”
• “Linked Data” is really a set of principles for
using the Semantic Web
– http://www.w3.org/DesignIssues/LinkedData.html
• (Many people use “Linked Data” and
“Semantic Web” somewhat interchangeably)
34. From observations to decisions
Observations
Satellites
(obs. sometimes
used for model
initialization)
Climate
Record
Creation
and
Curation
Climate Data
Records
Reports
Analysis
and
applications
Decision
making
Decisions,
Policies
Model results
Climate
Modelling
(Adapted from Dowell et al., 2013),
“Strategy Towards an Architecture for
Climate Monitoring from Space”)
35. Where can users go for help?
• Scientific literature
– Huge, verbose and inaccessible to some communities
– Not well linked to source data
• Technical reports and conference proceedings
– Hard to find, scattered or inaccessible
• Data centres
– increasingly strong at providing some important metadata, but don’t
usually include community feedback
– Not all countries and communities have data centres!
• Websites and blogs
– From CEOS Handbook to a scientist’s blog
– Increasingly useful, but scattered
36. How can climate data users decide whether a
dataset is fit for their purpose?
(N.B. We consider that “data quality” and “fitness for purpose”
are the same thing)
Not specific to climate data!
38. Examples of commentary metadata
• Post-fact annotations, e.g. citations, ad-hoc comments and notes;
• Results of assessments, e.g. validation campaigns,
intercomparisons with models or other observations, reanalysis;
• Provenance, e.g. dependencies on other datasets, processing
algorithms and chain, data source;
• Properties of data distribution, e.g. data policy and licensing,
timeliness (is the data delivered in real time?), reliability;
• External events that may affect the data, e.g. volcanic eruptions, ElNino index, satellite or instrument failure, operational changes to
the orbit calculations.
General rule: information originates from users or external entities,
not original data providers
– However, sometimes information is not available from the data
provider!
39. How will this be done?
• CHARMe will create
connected repositories of
commentary information
Data provider
website
rd
33rdparty
party
system
system
– Stored as triples in
“CHARMe nodes”,
• Information can be read
and entered through
websites or Web Services
• Using principles of Open
Linked Data and the
Semantic Web
CHARMe
node
CHARMe
node
CHARMe
node
40. Open Annotation
•
We are using Open Annotation for
recording commentary
– World Wide Web Consortium standard
•
Associates a body with a target
•
E.g. a publication could be the body,
an EO dataset could be the target
•
Can record the motivation behind an
annotation
– Bookmarking, classifying, commenting,
describing, editing, highlighting,
questioning, replying… (lots more)
– Covers a lot of CHARMe use cases!
•
An annotation can have multiple
targets
– A key requirement from users
41. Where does CHARMe fit in?
Observations
Satellites
(obs. sometimes
used for model
initialization)
Climate
Record
Creation
and
Curation
Climate
Modelling
(e.g. CMIP5)
Climate Data
Records
Reports
Analysis
Analysis
and
and
applications
applications
Decision
making
Decisions,
Policies
Model results
Supports analysts and scientists in
production of information for
decision-makers
43. Challenges
• Ensuring community adoption:
– We are “injecting” CHARMe capabilities into existing
websites used in the community
– We will “seed” the CHARMe system with information
to attract users (e.g. links between publications and
datasets)
• Ensuring quality of commentary metadata
– Moderation is a strong user requirement
– We will provide guides to creators of commentary –
what makes a comment helpful?
44. What CHARMe will enable
(some examples)
Users:
- “Find me all the documents that have been written about this dataset”
- “… in both peer-reviewed journals and the grey literature”
- “… and specifically about precipitation in Africa”
- “… in both NEODC’s and Astrium’s archives”
- “What factors might affect the quality of this dataset?”
e.g. upstream datasets, external events
- “What have other users already discovered about this dataset?”
- “I want to find information related to the dataset I’m looking at”
Data providers:
- “Who is using my dataset and what are they saying about it?”
- “Let me subscribe to new user comments and reply to them”
45. What CHARMe will not enable
• “Give me the best dataset on sea surface temperature”
– The “best” dataset depends on the application
• CHARMe will not provide a new “quality stamp” for
datasets
– But will be able to link to such things if other people publish
them, e.g. Bates maturity index, QA4EO certification
• CHARMe will not provide access to actual data
– (but will enable discovery of data)
• Not planning to create (another) “one-stop shop” for
information
– We want the information to appear where users are already
looking
47. How will you use CHARMe?
• Search for datasets on existing data provider website
• Use the “CHARMe button” to view or record
commentary on a particular dataset
48. “Significant events” viewer
•
•
•
•
Google Finance (above) matches stock prices with news events
CHARMe tool will match climate reanalysis data with “significant events”
– Algorithm changes, instrument failures, new data sources
Will allow user annotations on the data and events
Will be available on the Web
49. Fine-grained commentary
• Will allow creation and discovery about specific parts of
datasets
• E.g. variables, geographic locations, time ranges
• cf http://maphub.github.io
50. Jon Blower
Debbie Clifford
Tristan Quaife
Sam Williams
•
•
Using Linked Data to combine EO and socioeconomic data in 8 application areas
16-partner EU FP7 project, just started!
– Coordinated from Reading
51. Summary
•
Linked Data can join up information
from all over the Web
•
Makes disparate data sources
discoverable and processable
•
CHARMe is using Linked Data
techniques to help users of climate
data to connect with all the
experience in the community
– Techniques are not specific to
climate data
•
MELODIES project will combine
environmental data and
socioeconomic data to develop
real-world services using Linked
Data
Here’s a simplified representation of the things we are interested in in Earth Observation
And here’s a more complex version. The instrument produces “Level 0” data, which are then processed using a series of algorithms to produce something that the user community finds easier to use.
Documents may of course be written about any of these steps. All these documents will be written by different people at different times, and possibly about different versions of the same thing
Now let’s say you only know about a few of these things – a few publications, the algorithms you know already and the latest versions of a few datasets. Where do you go from here?
This person is trying to make a good decision based on climate data from satellites and from computer simulations. She also needs to know about other factors such as population growth and changes in urbanization. It’s a pretty hard job to bring together information from all these communities and correlate them. Everyone will use different standards, formats and terminology
If you are in a data-rich field that relies a lot on statistics, you may well be in this situation where the false positives outweigh the true ones. (95% test accuracy does not mean that there is a 95% chance that your hypothesis is true!)
Obviously, you can play with the numbers here, but in general this will be significant if you’re in the case where most hypotheses are false.
This may contribute to findings that results in the peer-reviewed literature can’t be reproduced – not necessarily due to poor experiments or tests.
Sites such as FigShare allow publication of stuff that would not make it into the peer-reviewed literature, but that are still useful.
Who here uses NASA instead of ESA because the data are easier to get hold of? NCEP instead of Met Office?
Just a reminder of what our simplified situation in EO looks like
This is how the Web represents our case: just a set of documents with simple links. The Web records no information about what the documents represent, or why they are linked.
Lots of links here, which is good! However, you need context (and natural language understanding) to determine what the links mean – are they affiliations, projects Richard works on, links to “institutional” (not personal) stuff, or what? A computer can’t reason based on this information (although link density gives some indication of relevance and popularity).
Here’s part of the citation list from one of my recent papers. Some of these are papers, some are standards, some are simple URLs. There is no information about why I’m citing them, unless you trawl the text and divine my meaning.
These are some of the problems of using documents-about-things to talk about stuff
These identifiers may look like web addresses (Uniform Resource Locators) but in fact are identifiers – just strings. You don’t expect to download me when you put my identifier in your web browser but you might get a representation of me (e.g. my home page)
The bottom one is a paper citing a dataset as a data source. Note that we can say *why* we’re citing the data
Your homework is to apply the same technique to the joke about the Dalai Lama going to a hot-dog stall and saying “make me one with everything”
The point about being part of the web is important, but hard to grasp. Your data can be found by semantically-aware web crawlers, which might be able to do something interesting with it, rather than just finding a document that requires a human to read.
So the Web of Data can more accurately record our situation, and can describe the nature of the links between things.
Lots of data, particularly public sector data, is being released as Open Linked Data
Met Office
Ordnance Survey
Office of National Statistics
Data.gov.uk
People need to publish not only their data, but the terms that they use and the relationships between them
A close-up of some of the geospatial part. Note the Met Office.
So what are we doing with Linked Data in the climate field? Here is a European project, about half way through,
Simple “value chain” for climate data observed from space.
See next slide for explanation of this
This is basically the same as the previous slide in words instead of pictures! The last paragraph is crucial – we’re looking at stuff the data provider doesn’t already know.
We’re not trying to turn the whole of climate data into Linked Data, but focusing on user-provided annotations
Important to note that decision-makers will probably not use CHARMe directly
So this person might be able to use CHARMe to find interesting information from all these communities, and find out what the experts in those communities think about the data. It will be easier to discover the common pitfalls.
CHARMe is often compared with Amazon, Tripadvisor etc, but the user base will be smaller and more specialist, therefore it is important for the information to be accurate. Can’t rely on the wisdom of crowds!
This is the most basic use case for CHARMe. In the next couple of slides we’ll look at more advanced uses. The fact that the annotations (i.e. the commentary information) will be machine-readable means that it’s possible to build all kinds of applications.
Here’s an “advanced application” for using CHARMe information, which will be correlated with climate reanalyses and external events.
Here’s another project that is just starting, which will use Linked Open Data to create new services and products. Many of the participants are small and medium-sized enterprises who will base business ideas on Open Data.