1. Seman&c
Analysis
in
Language
Technology
http://stp.lingfil.uu.se/~santinim/sais/2014/sais_2014.htm
Seman&c
Role
Labelling
/
Predicate-‐Argument
Structure
Marina
San&ni
san$nim@stp.lingfil.uu.se
Department
of
Linguis&cs
and
Philology
Uppsala
University,
Uppsala,
Sweden
Autumn
2014
Lecture 3: SRL/PAS
1
3. The
seman&cs
of
events
• Predicates
in
FOL
have
fixed
arity:
they
take
a
fixed
number
of
arguments
–
predicates
have
a
fixed
arity
Lecture 3: SRL/PAS
3
4. event
variables
à
(neo)
Davidsonian
event
representa&on
• No
need
to
specify
a
fixed
number
of
arguments
• The
event
itself
is
a
single
argument.
• Everything
else
is
captured
by
addi&onal
predica&on
!
Ǝe eating(e) ∧ eater(e, speaker)∧ eaten(e,turkey sandwich) ∧
meal(e,lunch) ∧ location(e,desk)∧time(e,tuesday)!
Lecture 3: SRL/PAS
4
5. What
is
the
seman&c
similarity
here?
• John
broke
the
window
Ǝe x,y, breaking(e) ∧ breaker(e, x) ∧ john(e,x)brokenThing(e,y) ∧ window(e,y)!
• Mary
opend
the
door
Ǝe x,y, opening(e) ∧ opener(e, x) ∧ mary(e,x) ∧ openThing(e,y) ∧ door(e,y)#
Deep roles = agents
Lecture 3: SRL/PAS
5
6. Examples:
Thema&c
Roles
• Thema&c
roles
refer
to
a
par&cular
model
of
seman&c
roles
• Them
roles
try
to
capture
the
seman&c
commonality
betw
breaker
and
eater
à
agents
à
voli&onal
causa&on
• brokenThing
and
openedThing
are
inanimate
objects
that
are
affected
by
te
ac&on
à
themes
Lecture 3: SRL/PAS
6
7. 2
seman&c
constraints
on
the
arguments
of
event
predicate
1. Seman&c
Roles
2. Selec&onal
Constraints
Lecture 3: SRL/PAS
7
8. I.
Seman&c
Roles
• Express
the
seman&c
of
the
arguments
and
its
rela&on
to
predicate
Lecture 3: SRL/PAS
8
10. Why
are
they
useful?
• Help
generalize
over
different
surface
realiza&ons
of
predicate
arguments.
• Ex:
Diathesis
Lecture 3: SRL/PAS
10
11. Problems
• No
standard
set
of
roles
• Some&mes,
many
fine-‐grained
roles
• Difficult
to
formalize
• Solu&on?
– Generalized
seman&c
roles
• PROTO-‐AGENT,
PROTO-‐PATIENT,
etc.
…
the
more
an
argument
displays
agent-‐like
proper&es
(voli&on,
inten&onality
etc),
the
greater
the
possibility
that
the
argument
can
be
labelled
a
proto-‐agent…
Lecture 3: SRL/PAS
11
12. Predicate-‐Argument
Structure
The
argument
structure
of
a
verb
is
the
lexical
informa&on
about
the
arguments
of
a
predicate
and
their
seman&c
and
syntac&c
proper&es.
Argument
structure
is
generally
seen
as
intermediate
between
seman&c-‐role
structure
and
syntac&c-‐func&on
structure.
See:
h^p://www.glo^opedia.org/index.php/Argument_structure
Lecture 3: SRL/PAS
12
13. Ex
Argument
structure
is
what
makes
a
lexical
head
induce
argument
posi&ons
in
syntac&c
structure
is
called
its
argument
structure.
Example:
the
head
open
has
an
argument
structure
which
induces
obligatorily
one
argument
posi&on
(Theme),
and
op&onally
two
more
(Agent
and
Instrument).
Lecture 3: SRL/PAS
13
14. PropBank
• Resource
of
sentences
annotated
with
seman&c
roles.
– The
English
PropBank:
sentences
from
the
PennTreeBank.
• Each
sense
of
each
verb
has
a
specific
set
of
roles:
– Arg0
=
proto-‐agent
– Arg1
=
proto-‐pa&ent
– The
seman&c
of
the
other
roles
is
specific
to
each
verb
sense…
Lecture 3: SRL/PAS
14
15. Ex
• Same
role,
despite
the
differing
surface
forms:
increase
and
Arg1
Lecture 3: SRL/PAS
15
16. FrameNet
• Project
that
a^empts
to
generalize
seman&c
roles
on
different
verbs
and
also
betw
verbs
and
nouns
Lecture 3: SRL/PAS
16
17. Frame
• A
structure
with
seman&c
roles
includes
frame
elements:
– Core
roles
– Non-‐core
roles
Lecture 3: SRL/PAS
17
18. Each
word
evoke
a
frame
• Ex:
change_posi&on_on_a_scale
Lecture 3: SRL/PAS
18
19. II.
Selec&onal
Restric&ons
• Seman&c
constraints
on
arguments
• Constraints
that
the
verb
imposes
on
the
concepts
that
are
allowed
to
fill
its
arguments
roles.
– I
want
to
eat
home
– I
want
to
eat
French
food
How
do
we
know
that
”home”
is
not
a
argument
of
eat?
Seman&cally,
we
can
say
that
the
theme
of
”eat”
is
edible.
edible
becomes
a
selec&onal
restric&on
of
the
theme
of
eat.
Lecture 3: SRL/PAS
19
20. Selec&onal
Restric&ons
and
FOL
• neo-‐Davidsonian
representa&on
of
events:
Lecture 3: SRL/PAS
20
• Drawbacks
(p.
662)
– Using
FOL
for
a
simple
task
like
this
is
overkill.
Far
too
computa&onally
expensive
– We
would
need
a
KB
of
facts
and
concepts
that
is
very
large…
21. A
more
prac&cal
approach
• State
selec&onal
restric&ons
in
terms
f
WordNet
synsets
rather
than
as
logical
concepts.
• Each
predicate
simply
specifies
a
WordNet
synset
as
the
selec&onal
restrictons
on
each
of
its
arguments.
ex:
eat
(food,
nutrient)
Selec&onal
restric&on
o
the
theme
role
of
eat
to
the
sysets
àfood,
nutrient
Lecture 3: SRL/PAS
21
22. Seman&c
Role
Labelling
• Synonyms:
– Thema&c
role
labelling
– Case
role
assignment
– Shallow
seman&c
parsing
• What
is
it?
– The
task
of
automa&cally
finding
the
appropriate
role
for
each
predicate
in
a
sentence
Lecture 3: SRL/PAS
22
23. Current
Approaches
• Based
on
supervised
machine
learning
– Adequate
amounts
of
training
and
testng
sets.
– FrameNet
and
PropBank
have
been
used
for
this
purpose.
Lecture 3: SRL/PAS
23