Abstract:
A growing number of resources are available for enriching documents with semantic annotations. While originally focused on a few standard classes of annotations, the ecosystem of annotators is now becoming increasingly diverse. Although annotators often have very different vocabularies, with both high-level and specialist
concepts, they also have many semantic interconnections. We will show that both the overlap and the diversity in annotator vocabularies motivate the need for semantic annotation integration: middleware that produces a unified annotation on top of diverse semantic annotators. On the one hand, the diversity of vocabulary allows applications
to benefit from the much richer vocabulary available in
an integrated vocabulary. On the other hand, we present evidence that the most widely-used annotators on the web suffer from serious accuracy deficiencies: the overlap in vocabularies from individual annotators allows an integrated annotator to boost accuracy by exploiting inter-annotator agreement and disagreement.
The integration of semantic annotations leads to new challenges, both compared to usual data integration scenarios and to standard aggregation of machine learning tools. We overview an approach to these challenges that performs ontology-aware aggregation. We
introduce an approach that requires no training data, making use of ideas from database repair. We experimentally compare this with a supervised approach, which adapts maximal entropy Markov models to the setting of ontology-based annotations. We further experimentally compare both these approaches with respect to ontology-unaware
supervised approaches, and to individual annotators.
Bajaj Allianz Life Insurance Company - Insurer Innovation Award 2024
ROSeAnn: Reconciling Opinions of Semantic Annotators Poster
1. ROSeAnn:Reconciling Opinions of
Semantic Annotations The Need for Integration Conflicting Opinions
Supervised Aggregation (MEMM)
Another brilliant goal by England midfielder David
Beckham earned Manchester United a point from a 1-1
draw at Stamford Bridge on Saturday after Gianfranco Zola
had put Chelsea ahead.
United stayed top of the English league standings as
Liverpool could only draw 0-0 at home to Blackburn
despite dominating throughout. Newcastle moved into third
as a Les Ferdinand goal was enough to win 1-0 at bottom
club Middlesbrough.
Ian Marshall scored a hat-trick in the first 27 minutes to
help Leicester beat Derby 4-2 while Sheffield Wednesday
came from two down to win 3-2 at Southampton.
Leeds won 1-0 at Sunderland and Coventry against Everton
and Nottingham Forest against Aston Villa ended goalless.
Semantic annotators label text snippets
as referring to certain entities, e.g.,
Barack Obama, London, or as instances
of particular entity types, e.g., actors,
governmental organisations, countries.
Semantic Annotators
A growing number of freely available
online services can enrich documents
with semantic annotations.
Unfortunately, their opinions about an entity
often disagree as they might be based on very
diverse background knowledge such as training
corpora, knowledge bases, contextual
information, POS tags, and crowds.
AlchemyAPI:Person
DBPediaSpotlight:SoccerClub
Lupedia:Settlement
Wikimeta:Organisation
StanfordNER:Organisation
IllinoisNER:Organisation
Extractiv:City
AlchemyAPI:Facility
AlchemyAPI:GeographicFeature
DBPediaSpotlight:SportsTeam
Zemanta:SportsTeam
OpenCalais:GeographicFeature
Luying Chen, Stefano Ortona, Giorgio Orsi, and Michael Benedikt
University of Oxford - Department of Computer Science
(name.surname@cs.ox.ac.uk)
Unsupervised Weighted Repair (WR)
http://diadem.cs.ox.ac.uk/roseann
Each annotator comes with a vocabulary
of semantically-related entity types that
often overlap on common-sense entities,
such as places, persons, and companies.
However, each annotator covers only a
fraction of a much larger universe of
concepts. By relating such vocabularies
to each other via mappings we can
achieve much better coverage.
Museum
Accuracy of individual annotators
varies greatly, redundancy and
logical relationships can be used to
gain confidence about an entity.
Each annotator contributes some
original types. None of them can
be dropped without losing recall.
Empirical Evaluation
1
0.8
0.6
0.4
0.2
0
Precision Recall FScore
Person Date Movie
1
0.8
0.6
0.4
0.2
0
Precision Recall Fscore
Location Sport Movie
Thing
Organisation Facility
Place Person
Point of
Interest
Club
Soccer
Club
Location
Settlement Natural
City
Feature
Geographic
= Feature
Organisation(X) Person(X)
Organisation(X) Location(X)
Settlement(X) GeographicFeature(X)
Location(X) PointOfInterest(X)
Place(X) Person(X)
Person(X) Facility(X)
Annotation vocabularies are semantically
related to each other via existing knowledge
bases (DBPedia, Freebase) or via common-sense.
These relations can be used to map
them to a common ontology to check for
logical conflicts or compatibility.
IS-A constraints enable type inference,
thus increasing recall at the expense of
precision. Disjointness constraints induce
logical inconsistencies used to locate
potentially erroneous annotations, thus
increasing precision.
WR computes an ontology-aware repair of the
set of annotations that is logically consistent
and “fair” to the annotators involved.
WR is unsupervised and does not assume any
prior knowledge about the annotators.
If the global ontology only states IS-A and
disjointness constraints, a solution can be
computed efficiently (repair → 2-SAT).
WR is designed for scenarios where training
data is unavailable or sparse.
Conflicts also occur at span-level, i.e.,
annotators agreeing on the entity type but
not on the extension of the span.
Notion of span adapted to support
composite annotations consisting of tokens
carrying logically incompatible types, e.g.,
“[[Subic] [Naval Base]]”.
0.85
0.8
0.75
0.7
0.65
0.6
Politician
Politician
Place
Extractiv WeightedRepair MEMM
μP μR μF1 MP MR MF1
Place ≠ Person
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
Politician
Place
Politician
Zemanta WeightedRepair MEMM
Score(Politician) -> +1
Score(Person) -> +1
Score(Place) -> -1
μP μR μF1 MP MR MF1
0.9
0.8
0.7
0.6
Fox WeightedRepair MEMM
μP μR μF1 MP MR MF1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
NERD WeightedRepair MEMM
μP μR μF1 MP MR MF1
WR and MEMM have been tested on
~670 documents from 4 corpora: MUC7,
Reuters, NETagger, and Fox.
Comparison have been carried out against
individual annotators and competitor
aggregators such as Fox and NERD.
Semantic (micro/macro) precision and
(micro/macro) recall metrics have been
adopted for the comparison.
When training data is available,
supervision can be used to
learn the most probable
sequences of annotations given
those available from gold
standard annotated documents.
MEMM can learn unorthodox
relationships among annotations that do
not necessarily follow standard inference
rules, e.g., it can learn to predict a
subclass C from (a set of) annotations
mentioning a superclass C’ of C.
Person
Politician
Place
¦ u
c
max Score(c) xc
Politician
Person
Politician
WR and MEMM perform in average better than all individual
annotators and aggregators with the exception of OpenCalais.
However, its vocabulary represents some 18% of all types.
MEMM is more accurate than WR, but on sparse datasets WR
shows better performance than MEMM. WR delivers higher
recall than MEMM that, in turn, is more precise than WR.
512
256
128
64
32
16
WR Solution Computation
Reuters MUC7 NETTagger FOX
2 5 8 11
msec
# Annotators
10000
1000
msec
100
10
1
MEMM Prediction
Reuters MUC7 NETTagger FOX
2 5 8 11
# Annotators
Online aggregation of annotation is feasible in practice
(~300ms for WR and ~1s for MEMM). The aggregation time is
orders of magnitude less than the time required to invoke the
online services and collect their answers.
Apart from the entity type and the
source annotator, the feature set for
MEMM includes ontological features
such as IS-A and disjointness. All
features are token based.
Online annotators are often black boxes
characterised by a continuously evolving
vocabulary, where entity types are
added, merged, or removed.
Region
Person Country
Scientist
Planet
Brand
Product
Planet
Ocean
Company
Mansour
WR receives as input
an annotated span
and produces as
output a logically
consistent set of
annotations.
An atomic score is
computed for each
opinion, based on
(inferred) support /
opposition by other
opinions.
Annotations are
inserted or deleted
from the initial
solution to obtain a
consistent set of
annotation that
maximizes the
objective function.
An initial solution
consists of the
possibly inconsistent
union of all entity
types.
Only the most
specific annotations
are retained in the
final solution.
xc 1
xc 1