Joint Posterior Revision of NLP Annotations via Ontological ......Eric Clapton 3. General...
1
Tools: Gold Corpus (NERC): AIDA CoNLL-YAGO (train) Datasets: AIDA CoNLL-YAGO (test-b) MEANTIME TAC-KBP Research question: Does the posterior joint revision of the annotations from Stanford CoreNLP (NERC) and DBpedia Spotlight (EL), via YAGO, improve their performances? Measures: NERC / EL / NERC+EL Deterministically computable from the classes of the entity (possibly leveraging alignments between EL Knowledge Base and ) Joint Posterior Revision of NLP Annotations via Ontological Knowledge Marco Rospocher ([email protected]), Francesco Corcoglioniti ([email protected]) Mr. Washington was runner-up at Wimbledon in 1996. 1. Problem: Incoherent Mention-level NLP Annotations The GW Bridge is a suspension bridge over the Hudson Washington (state) George Washington Bridge 2. Solution: Coherence via Ontology Eric Clapton is one of the greatest guitar players. PER Eric Clapton 3. General Probabilistic Model Variables Conditional Independence Assumptions 4. Model Instantiated on NERC + EL Ontological Background Knowledge confidence score learned from data Estimating NERC Estimating EL 6,016,695 entities Taxonomy of 568,255 classes (only ingoing links) #4479 5. Evaluation ① ② ③ ① ②③ M CP CP M * CP Prior (popularity based on entity ingoing links) Computed from gold corpus (NERC + class annotations) Consider only class sets restricted to popular classes (seen at least n * times in the gold corpus) Restricted to gold mentions. Similar improvements also considering all mentions, and macro-averaging by document or by NERC type. PER ORG NERC EL Named Entity Recognition and Classification Entity Linking ① ② ③
Transcript of Joint Posterior Revision of NLP Annotations via Ontological ......Eric Clapton 3. General...
Tools:
Gold Corpus (NERC): AIDA CoNLL-YAGO (train)
Datasets: AIDA CoNLL-YAGO (test-b)MEANTIME TAC-KBP
Research question: Does the posterior joint revision of the annotations from Stanford CoreNLP (NERC) and DBpedia Spotlight (EL),via YAGO, improve their performances?
Measures: NERC / EL / NERC+EL
Deterministically computable from the classes of the entity (possibly leveraging alignments between EL Knowledge Base and )
Joint Posterior Revision of NLP Annotationsvia Ontological KnowledgeMarco Rospocher ([email protected]), Francesco Corcoglioniti ([email protected])
Mr. Washington was runner-up at Wimbledon in 1996.
1. Problem: Incoherent Mention-level NLP Annotations
The GW Bridge is a suspension bridge over the Hudson
Washington (state)
George Washington Bridge
2. Solution: Coherence via Ontology
Eric Clapton is one of the greatest guitar players.PER
Eric Clapton
3. General Probabilistic Model
Variables Conditional Independence Assumptions
4. Model Instantiated on NERC + EL
Ontological Background Knowledge
confidence score learned from data
Estimating NERC
Estimating EL
6,016,695 entitiesTaxonomy of 568,255 classes
(only ingoing links)
#4479
5. Evaluation
①②③
①
②③
M
CP
CP
M*
CP
Prior (popularity basedon entity ingoing links)
Computed from gold corpus(NERC + class annotations)
Consider only class sets restricted to popular classes(seen at least n* times in the gold corpus)
Restricted to gold mentions. Similar improvements also considering all mentions, and macro-averaging by document or by NERC type.
PER
ORG
NERC
EL
Named Entity Recognition and Classification
Entity Linking
①② ③
