1. Jürgen Umbrich Invited talk at eXascale Infolab, Fribourg, June 2016
Multi-level semantic labelling
of numerical values
(ISWC 2016 submission)
Jürgen Umbrich
Invited talk at eXascale Infolab, Fribourg, June 2016

2. ABOUT MYSELF
2012: PhD A Hybrid Framework For Querying Linked Data Dynamically NUI Galway, Ireland (Prof. Decker, Prof. Polleres)
: FUJITSU KI2NA: Exploiting the potential of Linked Data in enterprises
Current: PostDoc at Vienna University of Economics and Business
Linked Traversal querying
Semantic Web Search Engine
Dynamic Linked Data Observatory
SPARQLES Endpoint monitoring
Polyglot backends combined with LD principles
OpenData
Quality assessment
Evolution monitoring
Data freshness

3. INSTITUTE FOR INFORMATION BUSINESS
Vienna University of Economics and Business

4. OPEN DATA PORTAL WATCH
Best Paper @OBD 2015
Scalable quality assessment & monitoring framework for 261+ Open Data Portals
Weekly snapshots
quality assessment
6 dimensions, 19 metrics

5. ADEQUATE (FFG PROJECT)
Quality Improvement of Open Data

6. GRAPHSENSE (FFG PROJECT) INSIGHT INTO DIGITAL CURRENCIES
Exploration and pattern detection in BitCoin

7. HDT: A LIGHTWEIGHT BINARY FORMAT FOR RDF
Highly compact serialization of RDF
Compact RDF store (without prior decompression)
Includes internal indexes to solve basic queries once it is loaded in main memory
Very fast on basic queries (triple patterns), x 1.5 faster than Virtuoso, RDF3x.
C++/Java libraries, standalone application, integration within Jena and Node.js

8. DBPEDIA WAYBACK MACHINE
Best Poster @SEMANTICS 2015
Retrieve historical versions of a DBpedia resource
What was the version of “Donald Trump” on dd/mm/yyyy?
How was its evolution?
Re-apply DBpedia mappings on the Wikipedia revision history

9. ARCHIVING LINKED OPEN DATA
Efficient/Scalable Representations of RDF Archives
Query Languages and Benchmarking
BEAR: Blueprint on benchmarking archives of semantic 2016
Data: Crawl from Linked Data Observatory
Basic queries: Materialize, get Version…
Initial evaluation on archiving policies

10. If you can’t enforce it, contract it: Enforceability in Policy-Driven (Linked) Data Markets
Simon Steyskal, Sabrina Kirrane
Data published in LDM must be accessed and used in a manner, which is compliant with access restrictions, licenses, institutional and community norms, and privacy requirements.
What if compliance cannot be enforced?

11. Multi-level semantic labelling
TODAY’S TALK
Multi-level semantic labelling
of numerical values
ISWC 2016 submission

12. MOTIVATION: OPEN DATA

13. MOTIVATION: OPEN DATA

14. MOTIVATION: OPEN DATA IMPACT
Enterprises using Open Data

15. MOTIVATION: USEFUL DATA
election results
geographical data
streets, kindergarten, trees, parks, first aid stations, …
ambient assistant living
addresses
demographic statistics
health information
many sources are manually curated

16. EXPLOITING THE POTENTIAL transforming CSVs to 5 Star Data
CSV URLs
CSVs link to other CSVs
CSVs link to other resources
(partially) Convert to RDF
allows for better/improved …
search & discovery
integration
processing

17. EXAMPLE dbp:capacity dpo:City dpo:Stadium dpo:Country stadium name
Ernst Happel Stadium
50865
Vienna
Austria
Franz Horr Stadium
13400
Red Bull Arena
32000
Salzburg …

18. EXISTING APPROACHES CSV2RDF approaches primarily focus
semantically label columns and cell values
headers to classes/properties,
values to entities
and perform Ontology alignment
approaches primarily focus
on Web Tables, which are well formed, syntactical structure (<thead>), human readable
textual descriptions for mappings

19. BUT Open Data tables typically contain
a large portion of numerical columns
missing headers
and/or non-textual headers
BableNet
solutions that solely focus on textual “cues” are
only partially applicable for mapping such data sources

20. OUR MISSON
Identifying the most likely semantic label for a bag of numerical values
capacity <a stadium> <country Austria>
stadium name
capacity
city
country
Ernst Happel Stadium
50865
Vienna
Austria
Franz Horr Stadium
13400
Red Bull Arena
32000
Salzburg …

21. OUR MISSON
Identifying the most likely semantic label for a bag of numerical values
capacity <a stadium> <country Austria>
Ernst Happel Stadium
50865
Vienna
Austria
Franz Horr Stadium
13400
Red Bull Arena
32000
Salzburg …

22. OUR MISSON
Identifying the most likely semantic label for a bag of numerical values
capacity <a stadium> <country Austria>
50865
13400
32000 …

23. APPROACH hierarchical clustering over an RDF knowledge base
to build a background knowledge graph
nodes representing typical numerical representatives, annotated with context information, i.e., grouped by properties and their shared domain (subject) pairs
k-nearest neighbours search
aggregation of the results at different levels to find the most likely context of the values, e.g.:
property
type
context

24. EXAMPLE

25. BACKGROUND KNOWLEDGE Construction of the Type hierarchy
represents rdfs:subClassOf relation for all available types
Construction:
Find properties with numerical range
Collect entities and all of their p-o pairs
Materials the OWL class hierarchy
Form a cluster for each type, containing all entity information

26. BACKGROUND KNOWLEDGE Construction of the p-o hierarchy
divisive hierarchical clustering approach
start with one node cluster
build candidates:
constrain property-object: all subject share the same property-object pair
constrain size: candidate nodes are larger than 1% of the parent node and smaller than 99%
sort candidates by their distance
select candidate with largest distance, subsequently select non overlapping candidates

27. K-NEAREST NEIGHBOUR SEARCH GENERAL IDEA
Mapping bags of numerical value to vector space (feature vector)

28. K-NEAREST NEIGHBOUR SEARCH GENERAL IDEA
Mapping bags of numerical value to vector space (feature vector)

29. K-NEAREST NEIGHBOUR SEARCH GENERAL IDEA
Compute & rank k-nearest neighbours for input values
1) input: [ 187, 201, 199, 198, 195, 199, 203, … ]
2) mapping:
3) compute distance to neighbours
4) select K nearest 2 4 3 6 1 5

30. K-NEAREST NEIGHBOUR SEARCH GENERAL IDEA
Compute & rank k-nearest neighbours for input values
1) input: [ 187, 201, 199, 198, 195, 199, 203, … ]
2) mapping:
3) compute distance to neighbours
4) select K nearest
5) rank nodes
diff ranking algorithms
e.g., majority vote: by property > type, p-o pair 2 4 3
Top 1: out of 6
Top 2: out of 6
Top 3: out of 6 6 1 5

31. RESULT AGGREGATION 2 4 3 6 1 5

32. EVALUATION SETUP DISTANCE FUNCTIONS Data
DBPedia 3.9
50 most frequent numerical properties
DISTANCE FUNCTIONS
euclidean distance (min, max, mean, stddev)
distribution similarity (Kolmogorov-Smirnov (KS) distance)

33. EVALUATION SETUP AGGREGATION FUNCTION AGGREGATION LEVELS
majority vote and average distance
AGGREGATION LEVELS
property
exact type
30 GB RAM
3 different knowledge bases
root type
all types
p-o level

34. EVALUATION: TEST/TRAIN DATA
train/test split : 80/20
20% of the subjects for each property as test data
test context graph: similar as background construction, however, without constrains
randomly select leaf nodes

35. EVALUATION: DISTANCE FUNCTION 1787 TEST NODES
Best: Kolmogorov-Smirnov (KS) distance
exact = correct property, type and p-o
prop = correct property
type = correct type
stype = correct super type

36. EVALUATION: LARGE-SCALE 33657 TEST NODES
9% of test nodes are contained 1-1 in knowledge graph !!
aggregation
majority and average vote
different neighbours
majority vote slightly better
more neighbours also better

37. EVALUATION: OPEN DATA TABLES 1170 TABLES
labelling numerical columns
manual inspection of top 100 tables ( based on distance)
Findings
Dealing with timeline data: values for different time points -> not in DBPedia
missing domain knowledge reports about spendings, election results, tourism
Aggregation of column scores: especially for type detection ( majority vote over column types)
Combine with complementary approaches

38. CONCLUSION juergen.umbrich@wu.ac.at
semantic labelling of numerical values
k-nearest neighbour search
hierarchical unsupervised background knowledge(BK)
can assign fine-grained semantic labels if there is enough evidence in BK
99.5% correct properties
96.3% correct parent types
Future work:
find and integrate more background knowledge
inspect Open Data Tables in more detail

39. semantic labelling of numerical values
k-nearest neighbour search
hierarchical unsupervised background knowledge(BK)
we can assign fine-grained semantic labels if there is enough evidence in BK
99.5% correct properties
96.3% correct parent types
Future work:
more background knowledge
Open Data Tables

40. DCAT Quality Dimensions & Metrics (1/3)

41. DCAT Quality Dimensions & Metrics (2/3)

42. DCAT Quality Dimensions & Metrics (2/3)