The document presents a methodology for natural language processing (NLP) data cleansing based on linguistic ontology constraints, focusing on two main vocabularies: LEMon and NIF. It highlights the challenges and common errors in existing NLP datasets, proposes a test-driven evaluation approach to enhance data quality, and provides examples of successful test cases. The findings indicate a substantial number of errors in current datasets and suggest future work to extend methodologies to additional NLP ontologies.

1. NLP Data Cleansing Based on Linguistic Ontology
Constraints
Dimitris Kontokostas13
Martin Brümmer1
Sebastian Hellmann13
Jens Lehmann1
Lazaros Ioannidis2
1AKSW, University of Leipzig
2Aristotle University of Thessaloniki
3DBpedia Association
2014-05-27
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 1 / 33

2. LOD Cloud (2011)
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 2 / 33

3. LOD Cloud (2011)
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 3 / 33

4. Linguistic Communities
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 4 / 33

5. Linguistic workshops & conferences
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 5 / 33

6. Linguistic workshops & conferences
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 6 / 33

7. Linguistic LOD Cloud (LLOD Cloud)
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 7 / 33

8. Problem denition
Linguistic (related) Data
Purpose-Driven denition
Increasing Data, ontologies vocabularies
New-comers → hard to understand the ontologies / follow updates
Validation is essential
Many dierent pipelines (parsing, annotation, disambiguation, etc)
Errors are propagated
Partially provided by maintainers (incomplete)
Focus on Lemon NIF (proof of concept)
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 8 / 33

9. Lemon - Lexicon Model for Ontologies
Models lexicon and machine-readable
dictionaries
RDF-native form
Linguistically sound structure (LMF)
Separation of the lexicon and
ontology layers
Linking to data categories →
arbitrarily complex linguistic
description
Principle of least power - the less
expressive the language, the more
reusable the data.
http://lemon-model.net/
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 9 / 33

10. Lemon - Example
: l e x i c o n a lemon : Lexicon ;
lemon : entry : Pizza , : T o r t i l l a .
: Pizza a lemon : LexicalEntry ;
lemon : sense [ lemon : r e f e r e n c e
http :// dbpedia . org / resource /Pizza ] .
: T o r t i l l a a lemon : LexicalEntry ;
lemon : sense [ lemon : r e f e r e n c e
http :// dbpedia . org / resource / T o r t i l l a ] .
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 10 / 33

11. Lemon - Example (Correct)
: l e x i c o n a lemon : Lexicon ;
lemon : language en ;
lemon : entry : Pizza , : T o r t i l l a .
: Pizza a lemon : LexicalEntry ;
lemon : canonicalForm [
lemon : writtenRep Pizza @en ] ;
lemon : sense [ lemon : r e f e r e n c e
http :// dbpedia . org / resource /Pizza ].
: T o r t i l l a a lemon : LexicalEntry ;
lemon : canonicalForm [
lemon : writtenRep T o r t i l l a @en ] ;
lemon : sense [ lemon : r e f e r e n c e
http :// dbpedia . org / resource / T o r t i l l a ].
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 11 / 33

12. NIF - NLP Interchange Format
RDF/OWL-based format that aims to achieve interoperability between
Natural Language Processing (NLP) tools, language resources and
annotations
In a nutshell:
Logical formalisation of strings and annotations
Builds on existing standards, e.g. RDF, LAF/GrAF, RFC 5147
Reuse of RDF tool stack
Decreases development cost for integration
Integrated in:
DBpedia Spotlight, Stanford Core NLP, OpenNLP, RDFace, Validator,
ConLL converter , ...
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 12 / 33

13. NIF - Overview
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 13 / 33

14. NIF - Example
http :// abc . com/doc#char=0,17
a n i f : Context ;
a n i f : RFC147String ;
n i f : beginIndex 0 ;
n i f : endIndex 17 ;
n i f : i s S t r i n g My dog l i k e s pizza .
http :// abc . com/doc#char=2,7
a n i f : RFC5147String ;
n i f : anchorOf dog ;
n i f : referenceContext http :// abc . com/doc#char=0,17 .
i t s r d f : taClassRef dbo : Animal ;
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 14 / 33

15. NIF - Example (Correct)
http :// abc . com/doc#char=0,18
a n i f : Context ;
a n i f :RFC5147 String ;
n i f : beginIndex 0^^xsd : nonNegativeInteger ;
n i f : endIndex 18^^xsd : nonNegativeInteger ;
n i f : i s S t r i n g My dog l i k e s pizza ^^xsd : s t r i n g .
http :// abc . com/doc#char=2,7
a n i f : RFC5147String ;
n i f : beginIndex 2^^xsd : nonNegativeInteger ;
n i f : endIndex 7^^xsd : nonNegativeInteger ;
n i f : anchorOf dog ^^xsd : s t r i n g ;
n i f : referenceContext http :// abc . com/doc#char=0,27 .
i t s r d f : taClassRef dbo : Animal ;
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 15 / 33

16. Maintainer validation
Lemon
Python script
24 tests for structural criteria
too slow on big datasets
not good reporting
NIF
SPARQL queries
11 tests for common errors
not complete
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 16 / 33

17. Built on previous work
Test-driven evaluation of linked data quality. Dimitris Kontokostas, Patrick
Westphal, Sören Auer, Sebastian Hellmann, Jens Lehmann, Roland
Cornelissen, and Amrapali J. Zaveri in WWW 2014.
Horizontal, multi-domain data quality assessment
Massive detection of errors for ve large-scale LOD data sets
291 vocabularies, independent of their domain or purpose
New contributions:
Relation to OWL reasoners
Test Driven Data Engineering Ontology
Domain-specic validation
Quickly improving existing validation options provided by maintainers
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 17 / 33

18. Test-Driven Data Development Methodology
Test case: a data constraint that involves one or more triples
Test suite: a set of test cases for testing a dataset
Status: Success, Fail, Timeout (complexity) or Error (e.g. network)
Fail: Error, warning or notice
RDF: basis for both data and schema
Unied model facilitates automatic test case generation
SPARQL serves as the test case denition language
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 18 / 33

19. Example test case
A nif:RFC5147String should never have a nif:beginIndex greater than
nif:endIndex
Test cases are written in SPARQL
SELECT ? s WHERE {
? s n i f : beginIndex ?v1 .
? s n i f : endIndex ?v2 .
FILTER ( ?v1 ?v2 ) }
We query for errors
Success: Query returns empty result set
Fail: Query returns results
Every result we get is a violation instance
Timeout / Error: needs further investigation on SPARQL Engine
capabilities, query syntax or query complexity
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 19 / 33

20. Patterns Bindings
Data Quality Test Patterns (DQTP)
abstract patterns, which can be further rened into concrete data quality
test cases using test pattern bindings
Existing library of 20 patterns
SELECT ? s WHERE {
? s %%P1%% ?v1 .
? s %%P2%% ?v2 .
FILTER ( ?v1 %%OP%% ?v2 ) }
Bindings
mapping of variables to valid pattern replacement
P1 = n i f : beginIndex | SELECT ? s WHERE {
P2 = n i f : endIndex | ? s n i f : beginIndex ?v1 .
OP = | ? s n i f : endIndex ?v2 .
| FILTER ( ?v1 ?v2 ) }
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 20 / 33

21. Test Auto Generators (TAGs)
RDF(s) OWL (partial) support
Query schema for supported axioms
SELECT DISTINCT ?T1 ?T2 WHERE {
?T1 owl : d i s j o i n t W i t h ?T2 . }
For every result a binding to a pattern is generated a test case
instantiated
Supported axioms at the moment:
RDFS: domain range
OWL: minCardinality, maxCardinality, cardinality, functionalProperty,
InverseFunctionalProperty, disjointClass, propertyDisjointWith,
AsymmetricProperty and deprecated
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 21 / 33

22. Test Case Elicitation Workow
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 22 / 33

23. TD(D)D vs Reasoners
SPARQL test cases detect a subset of validation errors detectable by
an OWL reasoner. Limited by
SPARQL endpoint reasoning support
limitations of the OWL-to-SPARQL translation.
SPARQL test cases detect validation errors not expressible in OWL
OWL reasoning is often not feasible on large datasets.
Datasets are already deployed and accessible via SPARQL endpoints
Pattern library more user friendly approach for building validation rules
compared to modelling OWL axioms.
requires familiarity
non-common validations require manual SPARQL test cases
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 23 / 33

24. Data Engineering Ontology
Input / Output entirely in RDF
Model the methodology in OWL
test suites, test cases, patterns, auto generators
Strict to serve as a validation layer
Four dierent levels of error reporting
simple test case report (success, fail) / enriched with counts
violation instance reporting / enriched with annotations
Reuse dcterms, prov, spin, rlog
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 24 / 33

25. Data Engineering Ontology - Denition Generation
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 25 / 33

26. Data Engineering Ontology - Result Representation
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 26 / 33

27. Lemon NIF Test case elicitation
RDFUnit Suite implements our methodology
Run on Lemon NIF ontologies
TAGs could not yet handle some complex owl:Restrictions
owl:unionOf, owl:allValuesFrom, owl:someValuesFrom,
owl:hasSelf and some rdfs:subPropertyOf cases
Manual test cases for constraints not captured in OWL.
Total Domain Range Datatype Card. Disj. Func. I. Func. Manual
Lemon 182 40 34 1 29 64 3 1 10
NIF 96 42 24 4 6 10 10
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 27 / 33

28. Example of manual Lemon test case
lemon:narrower denotes that one sense of a word is narrower than the
other and must never be symmetric or contain cycles.
SELECT DISTINCT ? s WHERE {
? s lemon : narrower+ ? narrower .
? narrower lemon : narrower+ ? s . }
lemon:language must not have a language tag (RDF1.1 to the rescue)
SELECT DISTINCT ? s WHERE {
? s lemon : language ?v1 .
FILTER ( lang (? v1 )!=) }
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 28 / 33

29. Example of manual NIF test case
Ensure that nif:beginIndex nif:endIndex index are correct
SELECT DISTINCT ? s WHERE {
? s n i f : anchorOf ? anchorOf ;
n i f : beginIndex ? beginIndex ;
n i f : endIndex ? endIndex ;
n i f : referenceContext
[ n i f : i s S t r i n g ? r e f e r e n c e S t r i n g ] .
BIND (SUBSTR(? r e f e r e n c e S t r i n g ,
? beginIndex ,
(? endIndex − ? beginIndex ) ) AS ? t e s t ) .
FILTER ( s t r (? t e s t ) != s t r (? anchorOf ) ) . }
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 29 / 33

30. Evaluation Datasets
Name Description Ontology Type
lemon datasets
LemonUby Wiktionary EN Conversion of the English Wiktionary into UBY-LMF model lemon,
UBY-LMF
Dictionary
LemonUby Wiktionary DE Conversion of the German Wiktionary into UBY-LMF model lemon,
UBY-LMF
Dictionary
LemonUby Wordnet Conversion of the Princeton WordNet 3.0 into UBY-LMF
model
lemon,
UBY-LMF
WordNet
DBpedia Wiktionary Conversion of the English Wiktionary into lemon lemon Dictionary
QHL Multilingual translation graph from more than 50 lexicons lemon Dictionary
NIF datasets
Wikilinks sample of 60976 randomly selected phrases linked to
Wikipedia articles
NIF NER
DBpedia Spotlight dataset 58 manually NE annotated natural language sentences NIF NER
KORE 50 evaluation
dataset
50 NE annotated natural language sentences from the AIDA
corpus
NIF NER
News-100 100 manually annotated German news articles NIF NER
RSS-500 500 manually annotated sentences from 1,457 RSS feeds NIF NER
Reuters-128 128 news articles manually curated NIF NER
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 30 / 33

31. Evaluation results
Size SC FL TO ER Auto Errors Man Errors MWarn MInfo
WiktDBp 60M 177 5 - - 3.746.103 7.521.791 - 3.582.837
WktEN 8M 168 14 - - 752.018 394.766 - 633.270
WktDE 2M 170 12 - - 273.109 66.268 - 155.598
Wordnet 4M 166 16 - - 257.228 36 - 257.204
QHL 3M 170 11 - 1 433.118 538.933 - 538.016
Wikilinks 0.6M 91 4 - 1 141.528 21.246 - -
News-100 13K 91 2 - 3 3.510 - - -
RSS-500 10K 91 2 - 3 3.000 - - -
Reuters-128 7K 91 2 - 3 2.016 - - -
Spotlight 3K 92 3 - 1 662 68 - -
KORE50 2K 89 6 - 1 301 55 - -
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 31 / 33

32. Conclusion
Extended a previously introduced methodology for test-driven quality
assessment
Data engineering ontology
Devised 277 test cases for NLP datasets using the Lemon and NIF
vocabularies
Revealed a substantial number of errors for Lemon NIF datasets
Future directions
extend the test cases to more NLP ontologies (MARL, NERD, ITSRDF)
automatic dependencies between test cases
wrap RDFUnit for NLP services (integrated in NIF)
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 32 / 33

33. Thank you!
Dimitris Kontokostas
With kind support of
John McCrae (Lemon model)
http://rdfunit.aksw.org
http://github.com/AKSW/RDFUnit
#eswc2014kontokostas
Kontokostas et al. (ESWC2014) NLP Data Cleansing 2014-05-27 33 / 33