This document discusses named entity recognition (NER) tasks and benchmarks for evaluating NER tools. It provides a brief history of NER benchmarks including CoNLL 2003/2005, ACE 2004-2007, TAC 2009, and ETAPE 2012. It also summarizes several standalone and web-based NER tools. The document outlines two human-annotated NER benchmarks, WEKEX 2011 and ISWC 2011, that were used to evaluate various NER tools and measure inter-annotator agreement. Finally, it introduces the NERD framework which aims to standardize and improve NER by developing an ontology, REST API, and linking NER extractions to Linked Open Data.
2. What is a Named Entity recognition task?
A task that aims to locate and classify the name of a
person or an organization, a location, a brand, a
product, a numeric expression including time, date,
money and percent in a textual document
12 March 2012 Seminar @ Ecole Centrale, Paris 2/21
3. History of NER benchmarks
CoNLL 2003 and CoNLL 2005
schema (4 types): person, organization, location and miscellaneous
language independent task
ACE 2004, ACE 2005 and ACE 2007
schema (7 types): person, organization, location, facility, weapon,
vehicle and geo-political entity
entity recognition, not just name (e.g. description, pronoun)
find relationships among entities extracted
TAC 2009 (Knowledge Base Track)
schema (3 types): person, organization and location
create a knowledge base from the named entities extracted
ETAPE 2012 (Named Entity Task)
schema: Quaero (7 main types, 32 sub-types)
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale -3
4. NER Tools
Standalone software
GATE
Stanford CoreNLP
Temis
Web APIs
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale -4
5. Factual comparison of 10 Web NER tools
Alchemy DBpe Evri Extr Lup Calais Saplo WM Yahoo Zemanta
dia
Granularity OEN OEN OED OEN OEN OEN OED OEN OEN OED
Language EN EN EN EN EN EN EN EN EN EN
FR GR* IT FR FR SW FR
GR PT* IT SP SP
IT SP*
PT
RU
SP
SW
Quota 30000 unl 3000 3000 unl 50000 1333 unl 5000 10000
(calls/day)
Sample C/C++ Java AS Java N/A Java Java Java JS C#
Clients C# JS Java JS Perl PHP Java
Java PHP PHP PHP JS
Perl Python Perl
PHP5 PHP
Python Python
Ruby Ruby
Content 150KB 452KB 8KB 32KB 20KB 8KB 26KB 80KB 7769KB 970KB
chunk
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale -5
6. Alchemy DBpedia Evri Extr Lup Calais Saplo WM Yahoo Zemanta
Response JSON Factual comparison (II)
HTML HTM HTML HTML JSON JSON JSON JSON XML
Format MicroF JSON L JSON JSON MicroF XML XML JSON
XML RDF JSO RDF RDFa RDF
RDF XML N XML XML
RDF
Entity 324 320 5 34 319 95 5 7 13 81
type
number
Entity N/A char N/A word range of char N/A POS range N/A
position offset offset chars offset offset of
chars
Classif. Alchemy DBpedia Evri DBpe DBpedia OpenC N/A ESTER Yahoo FreeBase
Ontologies FreeBase dia LinkedM alais
Scema.org DB
Defer. DBpeda DBpedia Evri DBpe DBpedia OpenC N/A DBpedia Wikipe Wikipedia
Vocabulari FreeBase dia LinkedM alais Geonam dia IMDB
es USCensus DB es MusicBrai
UMBEL CIAFact nz
OpenCyc book Amazon
YAGO Wikicom YouTube
MusicBrainz panies TechCrun
CrunchBase ch
...
12 March 2012 Seminar @ Ecole Centrale, Paris 6/21
7. Human made benchmarks
We performed two evaluation experiments:
WEKEX 2011
ISWC 2011
t = (entity, type, URI, relevant)
Each field has been rated by a Boolean value: true if
correct, false otherwise
Rizzo G., Troncy R. (2011), NERD: A Framework for Evaluating Named Entity Recognition Tools in the Web of Data.
In: International Semantic Web Conference 2011 (ISWC'11), Bonn, Germany.
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale -7
8. WEKEX 2011 Benchmark
Controlled experiment
4 human raters
10 English news articles (5 from BBC and 5 from The
New York Times)
Each rater evaluated each article for 5 extractors
200 total evaluations
Fleiss's kappa score
moderate agreement among raters
Rizzo G., Troncy R. (2011), NERD: Evaluating Named Entity Recognition Tools in the Web of Data.
In: (ISWC'11) Workshop on Web Scale Knowledge Extraction (WEKEX'11), Bonn, Germany.
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale -8
9. Results
different behavior
for different sources
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale -9
10. ISWC 2011 Benchmark
Controlled experiment
10 human raters
2 English news articles from The New York Times
each rater evaluated each article for 6 extractors
120 total evaluations
Fleiss's kappa score
substantial
agreement among
raters
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 10
11. Results
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 11
12. What is NERD?
ontology1 REST API2
UI3 The NERD ontology has been
integrated in the NIF project,
a EU FP7 in the context of the
LOD2: Creating Knowledge
out of Interlinked Data
1 http://nerd.eurecom.fr/ontology
2 http://nerd.eurecom.fr/api/application.wadl
3 http://nerd.eurecom.fr/
12 March 2012 Seminar @ Ecole Centrale, Paris 12/21
13. NERD Ontology
Align the taxonomies used by the extractors
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 13
14. Building the NERD ontology NERD type Occurrence
Person 10
Organization 10
Country 6
Company 6
Location 6
Continent 5
City 5
RadioStation 5
Album 5
Product 5
... ...
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 14
15. NERD REST API
/document
/user GET,
/annotation/{extractor} POST, JSON/RDF*
/extraction PUT,
/evaluation DELETE
“entities” : [{
... “entity”: “Tim Berners-Lee” ,
“type”: “Person” ,
“uri”: "http://dbpedia.org/resource/Tim_berners_lee",
“nerdType”: "http://nerd.eurecom.fr/ontology#Person",
“startChar”: 30,
“endChar”: 45,
“confidence”: 1,
“relevance”: 0.5
}]
Rizzo G., Troncy R. (2012), NERD: A Framework for Unifying Named Entity Recognition and Disambiguation Web Extraction
Tools. In: European chapter of the Association for Computational Linguistics (EACL'12), Avignon, France.
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 15
16. NIF: NLP Interchange Format Framework
Different outputs for the NLP tools
OpenCalais DBpedia Spotlight
"_type": "Organization", "@URI": "http://dbpedia.org/resource/DBpedia",
“name": "North Atlantic Treaty Organization", "@types": "DBpedia:Software,DBpedia:Work”
"organizationtype": "governmental civilian", "@surfaceForm": "dbpedia",
"nationality": "N/A", "@offset": "0",
"_typeReference": "@support": "11",
http://s.opencalais.com/1/type/em/e/Organization", "@similarityScore": "0.2387271374464035",
... …
Manual effort required for integration or reuse
time consuming
need to capture the definition of the attributes used in the
response format
NIF uses RDF for representing NER results as
Linked Data
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 16
17. Named Entities as textual annotations
Let's consider the document:
http://www.w3.org/DesignIssues/LinkedData.html
The Semantic Web isn't just about putting data on the web. It is about
making links, so that a person or machine can explore the web of data.
With linked data, when you have some of it, you can find other, related,
data.….
All the above plus, Use open standards from W3C (RDF and SPARQL) to
identify things, so that people can point at your stuff
...
entities: {
…
[entity: W3C, startChar: 23107, endChar: 23110],
…
}
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 17
18. NERD meets NIF
Model documents through a
set of strings deferencable on
the Web
: offset_23107_ 23110 a str:String ;
str:referenceContext :offset_0_26546 .
Map string to entity
: offset_23107_ 23110 sso:oen dbpedia:W3C.
Classification
dbpedia:W3C rdf:type nerd:Organization .
Rizzo G, Troncy R., Hellmann S. and Bruemmer M. (2012), NERD meets NIF: Lifting NLP Extraction Results to the Linked
Data Cloud. In: (LDOW'12) Linked Data on the Web (WWW'12), Lyon, France.
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 18
19. NERD Demo
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 19
20. NERD Timeline and Future Work
beginning Comparison of named entity extractors
NERD benchmarks
NERD REST API and NERD ontology
Lift NERD output results to the LOD cloud
today
NERD “smart” service: combining the best of
all NER tools
Dashboard for improving the NERD user
experience
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 20
21. http://nerd.eurecom.fr
@giusepperizzo @rtroncy #nerd
http://www.slideshare.net/giusepperizzo
12/03/2012 - Multimedia Semantics and Interaction - Séminaire Ecole Centrale - 21