Verb-mediated Composition of Attitude Relations Comprising

Institute of Computational Linguistics
Verb-mediated Composition of
Attitude Relations Comprising
Reader and Writer Perspective
Manfred Klenner
14. April 2017
Simon Clematide
Don Tuggener
Overview
I
Attitude detection: Who is for/against who?
I
Verb-based approach using a lexicon
I
Take event factuality and negation into account
I
Add transitive inference in embedded subclauses
→ Complex interplay between matrix verb and subclause
I
Writer perspective: How are entities depicted in discourse?
I
Reader perspective: How is this depiction of entities related
to the a priori
I
What are the roles of the entities given the two
perspectives?
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
2/15
Verb lexicon
Manually crafted resource. For ∼ 1600 verb senses (∼ 1000
verb lemmas) encode:
I
Attitude relation it induces between entities
I
The polar roles it casts on its arguments/complements
Verb
help
murder
criticize
suffer
survive
Polar roles
source
target
+
bene f actor
bene f iciary+
−
villain
victim−
entity
neg. e f f.−
−
neg. e f f.
+
bene f iciary
-
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Relation
advocate
adversary
adversary
-
Verb-mediated Composition of Attitude Relations,
3/15
Verb lexicon
Manually crafted resource. For ∼ 1600 verb senses (∼ 1000
verb lemmas) encode:
I
Attitude relation it induces between entities
I
The polar roles it casts on its arguments/complements
Verb
help
murder
criticize
suffer
survive
Polar roles
source
target
+
bene f actor
bene f iciary+
−
villain
victim−
entity
neg. e f f.−
−
neg. e f f.
+
bene f iciary
-
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Relation
advocate
adversary
adversary
-
Verb-mediated Composition of Attitude Relations,
3/15
Verb lexicon: Truth commitment and Event factuality
I
I
Verbs cast Truth commitment (Tc ) on embedded subclause
(True, False, None)
Events in subclauses have an event facuality that is
affected by
I
I
Truth commitment (Tc ) of the matrix verb
Affirmative status of the subclause and the matrix verb
Example
He managed to cook
He managed not to lie
He didn’t manage not to lie
She hopes he wins
She hopes he doesn’t wins
Tc (matrix)
True
True
False
None
None
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Fact. emd.
fact.
counter-fact.
counter-fact.
non-fact.
non-fact.
Verb-mediated Composition of Attitude Relations,
4/15
Verb lexicon: Truth commitment and Event factuality
I
I
Verbs cast Truth commitment (Tc ) on embedded subclause
(True, False, None)
Events in subclauses have an event facuality that is
affected by
I
I
Truth commitment (Tc ) of the matrix verb
Affirmative status of the subclause and the matrix verb
Example
He managed to cook
He managed not to lie
He didn’t manage not to lie
She hopes he wins
She hopes he doesn’t wins
Tc (matrix)
True
True
False
None
None
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Fact. emd.
fact.
counter-fact.
counter-fact.
non-fact.
non-fact.
Verb-mediated Composition of Attitude Relations,
4/15
Complex transitive inference with complement clauses
Truth commitment + Factuality + Negation = Relations (Roles)
Example
Tc
X criticizes that Y helps Z
TRUE
Fact.
emb.
fact.
X criticizes that Y doesn’t help Z
TRUE
cfact.
X hopes that Y helps Z
None
nfact.
X doesn’t hope that Y helps Z
None
nfact.
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Relations
X adversary Y
Y advocates Z
X adversary Z
X adversary Y
Y adversary Z
X advocate Z
X advocates Z
Y advocates Z
X adversary Z
Y advocates Z
Verb-mediated Composition of Attitude Relations,
5/15
Complex transitive inference with complement clauses
Truth commitment + Factuality + Negation = Relations (Roles)
Example
Tc
X criticizes that Y helps Z
TRUE
Fact.
emb.
fact.
X criticizes that Y doesn’t help Z
TRUE
cfact.
X hopes that Y helps Z
None
nfact.
X doesn’t hope that Y helps Z
None
nfact.
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Relations
X adversary Y
Y advocates Z
X adversary Z
X adversary Y
Y adversary Z
X advocate Z
X advocates Z
Y advocates Z
X adversary Z
Y advocates Z
Verb-mediated Composition of Attitude Relations,
5/15
Complex transitive inference with complement clauses
Truth commitment + Factuality + Negation = Relations (Roles)
Example
Tc
X criticizes that Y helps Z
TRUE
Fact.
emb.
fact.
X criticizes that Y doesn’t help Z
TRUE
cfact.
X hopes that Y helps Z
None
nfact.
X doesn’t hope that Y helps Z
None
nfact.
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Relations
X adversary Y
Y advocates Z
X adversary Z
X adversary Y
Y adversary Z
X advocate Z
X advocates Z
Y advocates Z
X adversary Z
Y advocates Z
Verb-mediated Composition of Attitude Relations,
5/15
Writer perspective
I
Writer frames entities by her/his way of reporting on them
I
Attitude relations are induced based on verbs
I
Polar roles are inferred based on argument of verb
I
Someone who kills is a villain; someone who recevies help is
a beneficiary etc.
I
Induction is constrained by factuality/negation
Moral
Non-moral
I
source
benefactor, villain
pos./neg. actor
target
beneficiary, victim
pos./neg. affected
Distinction moral/non-moral depends on verb (kill vs.
criticize)
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
6/15
Example: Transitive inference (Writer perspective)
Fac (4) = fact
Fac (8) = fact
Tc (4) = T
Tc (8) = T
Sig (4) = T
Aff (4) = aff
Aff (8) = aff
The right-wing politician3 criticized4 that5 the EU7 helps8 the refugees10 .
Wp (source, 4) = entity
Wp (target− , 4) = neg_affected
Prel (4) = adversary
Wp
(source+ , 8)
P+rel (3, (4, 8), 10) = adversary
Wp (target+ , 8) = beneficiary
= benefactor
Prel (8) = advocate
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
7/15
Example: Transitive inference (Writer perspective)
Fac (4) = fact
Fac (8) = fact
Tc (4) = T
Tc (8) = T
Sig (4) = T
Aff (4) = aff
Aff (8) = aff
The right-wing politician3 criticized4 that5 the EU7 helps8 the refugees10 .
Wp (source, 4) = entity
Wp (target− , 4) = neg_affected
Prel (4) = adversary
Wp
(source+ , 8)
P+rel (3, (4, 8), 10) = adversary
Wp (target+ , 8) = beneficiary
= benefactor
Prel (8) = advocate
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
7/15
Example: Transitive inference (Writer perspective)
Fac (4) = fact
Fac (8) = fact
Tc (4) = T
Tc (8) = T
Sig (4) = T
Aff (4) = aff
Aff (8) = aff
The right-wing politician3 criticized4 that5 the EU7 helps8 the refugees10 .
Wp (source, 4) = entity
Wp (target− , 4) = neg_affected
Prel (4) = adversary
Wp
(source+ , 8)
P+rel (3, (4, 8), 10) = adversary
Wp (target+ , 8) = beneficiary
= benefactor
Prel (8) = advocate
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
7/15
Example: Transitive inference (Writer perspective)
Fac (4) = fact
Fac (8) = fact
Tc (4) = T
Tc (8) = T
Sig (4) = T
Aff (4) = aff
Aff (8) = aff
The right-wing politician3 criticized4 that5 the EU7 helps8 the refugees10 .
Wp (source, 4) = entity
Wp (target− , 4) = neg_affected
Prel (4) = adversary
Wp
(source+ , 8)
P+rel (3, (4, 8), 10) = adversary
Wp (target+ , 8) = beneficiary
= benefactor
Prel (8) = advocate
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
7/15
Example: Transitive inference (Writer perspective)
Fac (4) = fact
Fac (8) = fact
Tc (4) = T
Tc (8) = T
Sig (4) = T
Aff (4) = aff
Aff (8) = aff
The right-wing politician3 criticized4 that5 the EU7 helps8 the refugees10 .
Wp (source, 4) = entity
Wp (target− , 4) = neg_affected
Prel (4) = adversary
Wp
(source+ , 8)
P+rel (3, (4, 8), 10) = adversary
Wp (target+ , 8) = beneficiary
= benefactor
Prel (8) = advocate
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
7/15
Example: Transitive inference (Writer perspective)
Fac (4) = fact
Fac (8) = fact
Tc (4) = T
Tc (8) = T
Sig (4) = T
Aff (4) = aff
Aff (8) = aff
The right-wing politician3 criticized4 that5 the EU7 helps8 the refugees10 .
Wp (source, 4) = entity
Wp (target− , 4) = neg_affected
Prel (4) = adversary
Wp
(source+ , 8)
P+rel (3, (4, 8), 10) = adversary
Wp (target+ , 8) = beneficiary
= benefactor
Prel (8) = advocate
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
7/15
Reader perspective
I
Reader has a priori view of the world: Moral values, political
views
I
Independent of writer perspective/verb-based role
assignment in writer prespective
Political
Moral
actors
MyProponent, MyOpponent
MyValueConfirmer/Contemner
concepts
MyValues, MyAversions
I
Proponent/Opponent (list): Donald Trump, Angela Merkel,
EU
I
Values/Aversions (sentiment lexicon): crime, love
I
ValueConfirmer/Contemner (sentiment lexicon +
composition): corrupt minister, honest politician
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
8/15
Reader perspective
I
Reader has a priori view of the world: Moral values, political
views
I
Independent of writer perspective/verb-based role
assignment in writer prespective
Political
Moral
actors
MyProponent, MyOpponent
MyValueConfirmer/Contemner
concepts
MyValues, MyAversions
I
Proponent/Opponent (list): Donald Trump, Angela Merkel,
EU
I
Values/Aversions (sentiment lexicon): crime, love
I
ValueConfirmer/Contemner (sentiment lexicon +
composition): corrupt minister, honest politician
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
8/15
Reader perspective
I
Reader has a priori view of the world: Moral values, political
views
I
Independent of writer perspective/verb-based role
assignment in writer prespective
Political
Moral
actors
MyProponent, MyOpponent
MyValueConfirmer/Contemner
concepts
MyValues, MyAversions
I
Proponent/Opponent (list): Donald Trump, Angela Merkel,
EU
I
Values/Aversions (sentiment lexicon): crime, love
I
ValueConfirmer/Contemner (sentiment lexicon +
composition): corrupt minister, honest politician
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
8/15
Example: Transitive inference
"The right-wing politician criticized that the EU helps the
refugees."
Writer perspective
I Right-wing politician adversary EU
I
I
EU advocate refugee
I
I
I
neg.e f f.− EU
bene f actor+ EU
bene f iciary+ refugee
Right-wing politician adversary refugee
Reader perspective (assuming left/liberal)
I
Right-wing politician: MyValueContemner
I
EU: MyProponent
I
Refugee: MyProponent
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
9/15
Example: Transitive inference
"The right-wing politician criticized that the EU helps the
refugees."
Writer perspective
I Right-wing politician adversary EU
I
I
EU advocate refugee
I
I
I
neg.e f f.− EU
bene f actor+ EU
bene f iciary+ refugee
Right-wing politician adversary refugee
Reader perspective (assuming left/liberal)
I
Right-wing politician: MyValueContemner
I
EU: MyProponent
I
Refugee: MyProponent
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
9/15
Combining Reader/Writer perspective
Five tuple:
hReader(source).Writer(source).Relation.Reader(target).Writer(target)i
e.g. The right-wing politician criticizes the EU
hMyValueContemner.Entity.Adversary.MyProponent.NegA f f ectedi
”A MyValueContemner as an entity is an adversary of
MyProponent that is negatively affected”
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
10/15
Charged relation tuples
Interesting interaction between Reader and Writer perspective
Relation Tuple
hsome, entity, is, adversary, of, myAversions, neg_affectedi
e.g. someone condemns terror (→ NewProp?)
hmyProp, benefactor, advocate, myValContemner, beneficiaryi
e.g. US supports dictator (→ NoLongerProp?)
hsome, villain, adversary, myValues, neg_affected i
e.g. someone ridicules human behaviour (→ NewOpponent?)
Cf. our LSDSem/EACL 2017 paper: ”Stance Detection in
Facebook Posts of a German Right-wing Party”
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
11/15
Evaluation: Precision and error analysis
I
Precision evaluation including relations (adversary,
advocate) and roles (victim, myAversions . . . )
I
Output over 3.5 Mio. sentences from ZEIT and Spiegel
(German periodicals)
I
Sample two instances of each of roughly 200 salient tuples
(e.g. MyProponent.Villain.Adversary.MyProponent.Victim),
i.e. 400 instances
I
Two annotators annotate precision errors (90% agreement),
no false negatives
I
Guideline: Relations and roles converted to natural
language entailments, do you intuitively agree with them
given the sentence?
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
12/15
Reader
Writer
Evaluation: Precision and error analysis
Role
advocate
adversary
neg_affected
pos_affected
villain
victim
beneficiary
benefactor
myValues
myAversions
myValueConfirmer
myValueContemner
Prec
0.73
0.74
0.70
0.58
0.70
0.65
0.43
0.35
0.95
0.91
0.71
0.86
I
Reader roles are easier
(lookup + composition)
I
Writer roles prec. varies
I
Errors: 50% preprocessing
(parsing, PAS extraction),
30% semantics (verb
polysemy, negation), 20%
factuality detection
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
13/15
Evaluation: Lexicon ablation and Recall
I
80 manually crafted sentences with multiple subclauses, 80
real sentences (German periodicals Spiegel/ZEIT)
I
Test coverage and impact of lexicon and verb signatures
(effect on factuality of embedded clauses)
I
Evaluate adversary / advocate relations
I
Set factuality of subclauses to either True/Non-factual/False
if matrix verb is negated (True if not negated)
Subclause Fact.
False
Non-fact
True
Lexicon
R
69.36
71.62
75.23
75.20
P
69.06
75.36
74.88
83.50
F1
69.21
73.44
75.06
79.10
→ Lexicon with verb signatures has a large impact on precision
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
14/15
Conclusion
I
Verb-based attitude detection model, purely lexicon-based
I
Model truth commitment and event factuality
I
Transitive inference of attitudes through subclauses
I
Model Reader and Writer perspectives and their interplay
Thanks!
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
15/15
Conclusion
I
Verb-based attitude detection model, purely lexicon-based
I
Model truth commitment and event factuality
I
Transitive inference of attitudes through subclauses
I
Model Reader and Writer perspectives and their interplay
Thanks!
14. April 2017
University of Zurich, Institute of Computational Linguistics,
Manfred Klenner, Simon Clematide, Don Tuggener
Verb-mediated Composition of Attitude Relations,
15/15