Content uploaded by Martin Taylor
Author content
All content in this area was uploaded by Martin Taylor on Oct 03, 2016
Content may be subject to copyright.
Dialogue
analysis
using
layered
protocols
Martin
Taylor
and
David
Waugh
1
Introduction
For
at
least
130
years,
psychologists
have
believed
that
humans
deal
with
the
world
by
processing
incoming
information
and
acting
on
it
at
successive
lev-
els
of
increasing
abstraction
(e.g.
Craik
and
Lockhart
1972;
Donders
1862;
Norman
1981;
Norman
1984;
Powers
1973).
Language
likewise
is
thought
to
be
produced
and
interpreted
in
a
layered
Way,
the
production
going
from
discourse
goals
to
acoustic
waveforms
(e.g.
Levelt
1989)
or
to
marks
on
pa-
per,
and
the
interpretation
going
back
again
through
the
levels
to
discourse
effects
(e.g.
Taylor
and
Taylor
1983).
It
may
be
argued
from
rst
principles
that
efficient
communication
between
intelligent
partners
requires
layered
coding
if
the
communication
is
to
be
dynamically
stable
(Taylor
1989).
This
chapter
describes
a
layered
approach
to
the
description
of
intelligent
dia-
logue.
We
call
it
the
Layered
Protocol
(LP)
theory
of
communication.
In
the
last
decade
or
two,
layered
systems
have
become
standard
for
com-
munication
in
networks
of
heterogeneous
computers
(Zimmermann
1980;
Tanenbaum
1981).
These
computer-based
standards
operate
on
the
assump-
tion
that
the
transformations
applied
by
one
site
are
well
de ned
and
can
be
inverted
at
the
receiving
end.
They
work
according
to
a
‘coding—decoding’
model
of
communication.
In
human
to
human
or
human
to
computer
dia-
logue,
no
such
inverse
transformation
is
possible.
Layered
computer
com-
munication
may
be
a
good
metaphor
for
those
who
are
so
inclined,
but
it
cannot
describe
interactions
that
involve
a
human,
and
in
which
the
pri-
mary
concern
is
the
satisfaction
of
the
partners
that
the
interaction
has
the
desired
result,
no
matter
how
it
is
achieved.
,
The
‘coding’
transformations
that
relate
the
different
levels
of
abstrac-
tion
in
human
communication
cannot
be
well
defined,
and
indeed
the
proto-
cols
for
a
level
may
be
developed
and
modi ed
as
needed
within
an
ongoing
dialogue.
A
human
involved
in
a
dialogue
must
model
the
partner’s
inten-
190
MARTIN
TAYLOR
84
DAVID
WA
UGH
tions,
if
communication
is
to
be
successful,
and
to
do
so,
must
also
model
the
partner’s
view
of
the
non-dialogue
world.
Following
the
ideas
of
Perceptual
Control
Theory
(PCT;
Powers
1973)
we
take
the
successful
communication
of
a
message
to
have
occurred
in
a
cooperative
dialogue
when
the
originator
of
the
message
believes
it
to
have
had
the
desired
e ect
on
the
recipient,
and
when
the
recipient
believes
that
the
originator
is
satis ed.
The
relation
between
LP
Theory
and
Perceptual
Control
Theory
is
discussed
more
deeply
elsewhere
(
Taylor
and
Waugh
1992;
Taylor
and
Waugh
1999;
and
especially
Farrell,
Gamble,
Hollands,
and
Taylor
1999;
Taylor,
Farrell,
Gamble,
and
Hollands
1999).
Initially,
this
Layered
Protocol
approach
was
developed
as
a
way
of
de-
signing
and
analyzing
computer—human
interaction,
and
it
maintains
its
value
for
this
purpose,
despite
having
subsequently
developed
into
a
general
theory
of
communication
between
intelligent
partners
(Taylor,
McCann,
and
Tuori
1984;
Taylor
1988a;
Taylor
1988b;
Taylor
1989;
Taylor
and
Waugh
1999)
The
approach
is
based
on
two
fundamental
ideas:
the
idea
of
a
pro-
tocol,
and
the
idea
that
interpreting
and
acting
in
the
world
is
done
in
a
series
of
layers
of
abstraction.
A
message
may
be
communicated
at
any
level
of
abstraction,
but
it
is
a
‘virtual’
message,
its
reality
being
implemented
as
an
entire
dialogue
at
a
lower,
supporting,
level,
in
which
the
messages
are
characteristically
different
in
kind.
A
‘command,’
for
example,
may
be
implemented
as
a
verbal
string,
a
gesture,
or
a
mixture
of
the
two,
but
whichever
it
may
be,
words
and
gestures
are
not
themselves
commands.
A
whole
verbal
and
gestural
dialogue
may
be
required
before
the
recipient
of
the
command
is
able
to
interpret
the
intent
of
the
issuer
of
the
command,
and
it
is
only
when
that
intent
is
understood
that
the
command
has
been
fully
communicated.
In
the
literature
on
human-computer
interaction,
a
particular
form
of
layering
has
become
almost
conventional.
A
physical
transmission
layer
supports
a
lexical
layer,
which
in
turn
supports
a
syntactic,
then
a
seman-
tic,
and
then
a
pragmatic
layer
and
possibly
other,
yet
higher
layers
(e.g.
Buxton
1983;
Foley
and
van
Dam
1982;
Moran
1981;
Nielsen
1986).
This
is
not
the
kind
of
layered
protocol
we
are
discussing.
In
LP
Theory,
each
protocol,
at
every
level
of
abstraction,
incorporates
lexical,
syntactic,
and
pragmatic
components.
Semantics
is
seen
as
an
aspect
of
the
interrelations
across
protocol
layers
rather
than
a
property
of
any
layer.
In
this
way,
LP
theory
seems
to
clarify
some
of
the
confusion
that
often
surrounds
these
key
concepts
of
psycholinguistics,
in
particular
the
distinction
between
seman-
tics
and
syntax,
and
between
semantics
and
pragmatics.
As
we
use
the
term,
layered
protocols
refer
to
true
protocols
explicit
LAYERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
191
or
implicit
agreements
as
to
procedures
~
that
can
be
of
almost
any
form.
A
protocol
has
a
meaning
in
diplomacy
that
accords
well
with
our
use
of
the
term.
As
the
diplomat
George
Ignatieff
said:
“I
learned
that
protocol
is
really
a
language,
a
set
of
rules
and
conventions
which
enable
people
of
different
nationalities,
social
backgrounds
and
political
persuasions
to
feel
comfortable
with
each
other”
(Ignatieff
1985).
When
the
partners
are
of
different
‘backgrounds
and
persuasions,’
as
are
any
two
humans,
or
a
human
and
a
computer,
it
is
not
normally
possible
for
the
recipient
of
a
message
to
interpret
it
by
inverting
an
encoding
used
by
the
originator.
However,
by
using
appropriate
protocols
the
partners
can
often
determine
each
others’
intentions
su iciently
accurately
for
the
purpose
at
hand.
The
protocol
of
a
layer
in
Layered
Protocol
Theory
is
the
way
that
the
originator
of
a
message
at
that
layer
can
determine
and
in uence
how
the
recipient
is
interpreting
it,
and
how
the
recipient
can
determine
whether
the
ongoing
process
of
interpretation
is
approaching
the
originator’s
intent.
Individual
protocols
within
a
layer
are
conceptually
executed
by
proces-
sor
structures
called
‘protocol
nodes.’
The
present
chapter
employs
LP,
and
particularly
the
structure
of
and
relationships
among
protocol
nodes,
as
a
tool
for
the
analysis
of
dialogue
between
humans
and
information
systems.
1.1
Dialogues
to
be
analyzed
Much
effort
has
been
expended
on
the
design
and
analysis
of
information
dialogues
such
as
timetable
enquiries,
telephone
number
enquiries,
and
the
like.
The
PLUS
project
is
based
on
such
dialogues
(Black
et
al.
1991).
Information
dialogues
are
well
suited
for
analysis,
since
the
analyst
and
both
partners
agree
initially
on
the
goal
of
the
dialogue,
the
topics
are
(usually)
closely
de ned
and
intelligible
to
the
analyst,
the
stages
toward
reaching
the
goal
are
relatively
clear,
and
the
turn-taking
is
relatively
constrained.
Free
conversation,
on
the
other
hand,
has
no
agreed
overt
goal,
topics
vary
freely,
and
there
are
no
obvious
(to
a
third
party)
stages
in
the
approach
to
the
unstated
goal(s),
if
indeed
such
goals
exist.
Although
LP
theory
applies
to
all
communication
among
intelligent
part-
ners,
in
this
chapter
we
limit
ourselves
to
typed
communication
between
an
information-seeking
‘Client’
(C)
and
a
human
‘Wizard’
who
may
simulate
an
information—providing
computer.
We
consider
two
example
di-
alogues,
one
previously
analysed
by
Bunt
(1989)
and
one
gathered
as
part
of
the
PLUS
project
(Black
et
al.
1991).
The
rst
typed
dialogue,
which
we
will
call
the
‘Alicante
dialogue’
(tran-
192
MARTIN
TAYLOR
63’
DAVID
WA
UGH
scribed
in
Section
1.3.1)
is
between
two
humans,
one
of
whom
is
providing
travel
information
about
Schiphol
(Amsterdam)
Airport.
The
Alicante
dialogue
goes
very
smoothly,
but
presents
a
problem
for
some
methods
of
analysis:
At
one
point,
there
is
the
following
exchange:
C5:
What
is
the
last
train
from
Breda
I
can
take
to
be
in
time
for
ight
IB
885?
W6:
The
train
of
12.06.
C6:
What
is
the
arrival
time
in
Alicante?
W7:
17.00
In
this
exchange,
how
does
the
information
provider
know
that
the
question
is
about
the
ight
rather
than
the
train?
LP
theory
nds
the
answer
in
the
structure
of
the
dialogue,
whereas
it
might
otherwise
be
argued
that
the
answer
must
be
in
the
pragmatic
real—world
knowledge
held
by
the
information
server
about
planes
and
trains
(which
is,
of
course,
useful
even
when
it
is
not
necessary).
The
second,
longer,
dialogue
(see
Appendix)
was
collected
at
LIMSI-
CNRS,
France
as
part
of
the
PLUS
project
(Black
et
al.
1991).
Again
it
is
a
Wizard
of
Oz
dialogue,
in
which
a
human
simulates
a
computer
information
service
that
performs
some
of
the
functions
of
the
Yellow
Pages
of
a
telephone
directory,
with
respect
to
restaurants.
This
dialogue
ows
less
smoothly
than
the
Alicante
dialogue,
partly
because
the
client
deliberately
tries
to
probe
the
limits
of
the
service’s
capability.
1.2
Goals,
messages,
and
protocols
in
LP
Theory
LP
Theory
assumes
that
one
of
the
dialogue
partners
has
a
(possibly
future)
task
that
cannot
be
done
without
help.
The
necessary
cooperation
requires
some
information
to
be
communicated
to
the
partner
so
that
the
recipient
will,
then
or
later,
act
in
a
way
helpful
to
the
originator.
Getting
the
necessary
information
across
to
the
cooperating
partner
is
a
communicative
goal
that
in
most
cases
leads
to
an
intention
to
send
a
message.
The
goal
is
satisfied
when
the
originator
of
the
message
believes
the
recipient
to
have
come
to
the
desired
state
of
knowledge
or
to
be
acting
in
the
desired
way.
If
the
dialogue
is
cooperative,
but
not
otherwise,
the
recipient
has
a
goal
to
see
the
originator
as
being
satis ed
with
the
recipient’s
interpretation
of
the
message.
The
dynamically
changing
structure
of
belief
states
of
the
partners
is
at
the
core
of
LPT.
Dialogue
consists
of
an
exchange
of
messages
between
the
partners
at
multiple
levels
of
abstraction.
What
the
originator
wants
would
be
trans-
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
193
Originating
partner
Receiving
partner
Primal
1::
Primal
1\]/gessfgg
gg,
Message
(
ewe
lg.
(Altered
actual
p
ner
state)
Virtual
Messages
(eg.
a
party
invitation)
Virtual
Messages
g.
date
of
party)
-
----------
----
-
--§-------------------4|
6
Messages
690‘
J‘
(e.g.
acoustic;
waveform)
Q7
Figure
1:
A
message
is
communicated
by
a
series
of
transformation
stages,
which
can
be
regarded
as
communicating
a
set
of
virtual
messages
across
the
interface
at
several
levels
of
abstraction.
Each
virtual
message
uses
a
protocol
peculiar
to
its
own
stage
of
coding.
The
message
entering
a
transmitting
(or
leaving
a
receiving
stage)
is
considered
to
be
a
primal
message
for
that
stage.
‘Coding’
and
‘Decoding’
are,
especially
in
the
higher
layers,
conducted
by
negotiation
rather
than
as
algorithmic
transformations.
The
manner
of
negotiation
constitutes
part
of
the
protocol.
mitted
telepathically
if
that
were
possible;
since
that
is
not
possible,
it
must
be
recoded
into
some
physical
form.
The
originator
does
something
the
re-
cipient
can
observe,
so
that
the
recipient
can
determine
the
originator’s
intention.
This
recoding
is
done
in
several
stages
or
layers,
each
of
which
is
represented
by
a
protocol,
a
means
of
transmitting
virtual
messages
of
its
own
kind
across
the
interface
( g.
1).
These
virtual
messages
are
actually
transmitted
by
being
recast
at
the
next
lower
protocol
layer.
All
messages
across
the
gap
between
the
partners
at
a
level
of
abstraction
above
the
phys-
ical
are
known
in
LP
theory
as
virtual
messages,
only
the
physical
effects
being
‘real’
messages.
The
communicative
goal
is
transformed
by
the
originator
into
a
single
set
of
high
level
virtual
messages
that
are
supposed
to
affect
the
recipient
in
a
useful
way.
For
each
protocol,
the
message
it
has
to
communicate
is
a
state
of
the
recipient
that
the
originator
of
that
message
would
like
to
see;
that
194
MARTIN
TAYLOR
6
DAVID
WA
UGH
desired
state
is
called
the
primal
message
for
the
protocol.
In
the
language
of
Perceptual
Control
Theory
(Powers
1973)
the
primal
message
is
a
reference
value
for
the
originator’s
perception
of
the
recipient.
The
virtual
messages
at
one
level
of
abstraction,
which
result
from
the
discrepancy
between
the
desired
and
perceived
states
of
the
partner
at
that
level,
become
the
primal
messages
for
the
next
lower,
supporting,
level.
We
say
that
protocol
A
is
supported
by
protocol
B
if
the
physical
re-
alisation
of
messages
of
protocol
A
involve
the
use
of
protocol
B
at
any
intermediary
stage.
Protocol
A
is
directly
supported
by
protocol
B
if
the
virtual
messages
emitted
by
the
transmitting
node
of
protocol
A
are
them-
selves
the
primal
messages
of
protocol
B.
Another
way
to
interpret
the
concept
of
‘support’
is
that
if
the
originator
is
to
perceive
the
recipient
as
being
in
some
state
specified
by
a
primal
message
of
A,
that
desired
state
can
be
represented
as
a
composite,
or
function,
of
‘supporting’
desired
states
that
are
the
primal
messages
of
lower
levels
communicated
sequentially
or
in
parallel.
For
every
protocol,
there
is
in
each
partner
a
protocol
node
(PN),
a
conceptual
abstraction
that
encapsulates
the
processes
and
information
used
by
that
partner
in
dealing
with
the
particular
protocol.
In
g.
1
the
protocol
nodes
are
represented
by
the
boxes
labelled
T
and
R
for
transmitting
and
receiving
protocol
nodes,
respectively.
The
originator
of
a
message
uses
a
transmitting
PN,
the
recipient
a
receiving
PN.
The
two
kinds
differ
only
in
minor
details,
which
will
be
discussed
later.
We
refer
to
a
transmitting-
receiving
pair
of
protocol
nodes,
together
with
the
link
between
them,
as
a
channel
for
communicating
the
primal
message.
The
originating
partner’s
transmitting
protocol
node
for
a
particular
protocol
receives
as
input
a
primal
message,
which
is
a
state
that
the
origi-
nator
wants
to
perceive
as
existing
in
the
recipient.
It
compares
the
primal
message
with
the
currently
perceived
state
of
the
partner
and
generates
one
or
more
virtual
messages,
which
are
the
means
by
which
the
recipient
is
moved
from
the
currently
perceived
state
toward
the
desired
state.
Encod-
ing
the
primal
message
consists
of
representing
in
a
set
of
simpler
forms
the
difference
between
the
originator’s
model
and
wished—for
states
of
the
recipient.
Each
PN
has
a
Coder
for
generating
the
virtual
messages
and
a
Decoder
for
interpreting
the
virtual
messages
from
the
partner
across
the
link.
A
transmitting
PN
in
the
originator
of
the
message
encodes
the
primal
message
into
zero
or
more
virtual
messages
of
its
special
type.
The
number
will
be
zero
if
the
originator
believes
the
recipient
already
to
be
in
the
state
represented
by
the
primal
message.
A
receiving
PN
converts
the
virtual
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
195
messages
it
receives
into
a
belief
about
what
state
the
originator
wants
the
recipient
to
be
in,
thereby
decoding
the
original
primal
message.
The
recipient
has
two
goals
in
a
cooperative
dialogue.
The
rst
is
to
perceive
the
originator
to
be
satis ed
with
the
recipient’s
interpretation
of
the
primal
message.
This
rst
goal
implies
the
second,
which
is
to
determine
the
state
that
the
originator
wants
the
recipient
to
achieve.
These
are
independent
goals,
and
in
a
non-cooperative
dialogue,
one
or
the
other
may
not
be
true.
For
example,
the
recipient
may
not
care
what
the
originator
wants,
but
still
may
want
the
originator
to
be
satis ed,
or
the
recipient
may
want
to
determine
the
originator’s
intent
secretly.
Either
is
non-cooperative
on
the
recipient’s
part.
Non—cooperation
may
occur
on
the
originator’s
side,
as
well.
The
origi-
nator
may
wish
to
achieve
a
certain
effect
on
the
recipient
while
letting
the
recipient
believe
that
a
different
effect
is
intended
(‘deceit’),
or
even
without
letting
the
recipient
believe
that
a
communication
is
occurring
at
all.
In
Shakespeare’s
Othello
there
is
a
good
example
of
non-cooperative
communication,
in
which
it
is
the
originator
who
is
non-cooperative.
When
Iago
drops
Desdemona’s
handkerchief
for
Othello
to
nd,
part
of
the
point
of
his
deceit
is
that
the
recipient,
Othello,
never
knows
that
the
message
has
been
sent.
For
Iago,
success
consists
of
perceiving
Othello
to
have
reached
a
particular
state
of
belief
about
the
state
of
the
world,
and
that
state
of
belief
must
absolutely
not
include
that
Iago
intends
to
be
satis ed
with
the
message
transmission.
The
two
sample
dialogues
illustrated
in
this
chapter
are
intended
to
be
cooperative,
although
to
call
the
Restaurant
dialogue
‘cooperative’
may
stretch
the
term
a
little.
A
complete
description
of
a
protocol
includes
the
syntax
of
the
virtual
messages
it
can
transmit,
together
with
the
types
of
feedback
messages
that
de ne
the
kinds
of
state
that
the
originator
can
perceive
in
the
recipient.
These
kinds
of
state
determine
the
kinds
of
primal
messages
this
protocol
can
transmit.
Feedback
messages
are
necessary
when
the
effect
of
the
direct
messages
cannot
be
reliably
predicted.
Especially
at
the
higher
levels,
reli-
able
predictions
are
usually
not
possible,
and
so
feedback
from
the
recipient
is
the
main
mechanism
for
assuring
reliable
message
transport.
The
cycle
of
message
between
partners
is
called
a
protocol
loop
( g.
2).
From
the
viewpoint
of
a
third
party
analyst,
a
protocol
is
executed
simultaneously
in
both
communicating
partners;
but
the
originator
and
the
recipient
each
see
it
in
a
different
light.
Each
of
them
is
using
a
model
of
the
other
to
determine
the
interpretation
of
what
the
other
is
saying
and
doing,
interpretations
they
can
test
only
through
feedback.
For
a
third
party,
the
196
MARTIN
TAYLOR
63
DAVID
WA
UGH
Originating
partner
Receiving
partner
Puma]
Primal
.
.
Q,¢_
‘
T
V1rtual‘Messages
R
%\q>°a°%
'04-“?!
T
.
R
399%
%°eJ,
"Real"
so
3
_
i
“
’Ft’écéi\}ih’g?g]
:;.Protoco|
Node
3;:
3;:
Protocol
Node}
Decoder
Coder
Model
"
'rtuaI
Message
Model
I
.‘.
l'I
l
.
'
One
':'Protoc:'o|
Loop
Pnmal
.:'
1:.
1:.
Primal
Message
Coder
;._ ;._
Decoder
_.-_
Message
.
.
Km
.
_:-
Model
vir_tuatMes_sage
"
lmodei
5-_
Figure
2:
A
protocol
loop,
in
context
of
the
ladder
of
coding
layers
(upper
gure),
and
by
itself
(lower
gure).
The
primal
message
is
encoded
using
a
model
of
the
recipient
and
transmitted
as
a
virtual
message
across
the
interface,
where
it
is
decoded
using
a
model
of
the
originator.
If
the
decoding
seems
to
be
correct,
the
message
received
may
be
the
primal
message
that
should
be
delivered
to
the
intended
destination,
and
a
feedback
message
acknowledging
receipt
may
be
sent
back
across
the
interface.
If
not,
a
feedback
message
asking
for
clari cation
is
sent,
and
the
original
sender
may
respond.
The
part
of
the
protocol
loop
belonging
to
either
of
the
partners
is
called
a
protocol
node.
The
message
originator
has
a
‘Transmitting
Node’
and
the
recipient
a
‘Receiving
Node.’
LAYERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
197
primal
messages
are
invisible;
they
are
the
intentions
of
the
parties
at
the
various
levels
of
abstraction.
Each
partner
can
ask
for
clari cation
of
the
intentions
of
the
other,
but
the
observer
can
not.
Analysis
of
a
dialogue
is
therefore
likely
always
to
be
problematic.
The
virtual
messages
that
go
between
the
parties
around
a
protocol
loop
are
collectively
called
a
dialogue.
This
is
a
one-level
View
of
the
dialogue,
as
is
shown
in
the
lower
part
of
g.
2.
Another
view
of
the
same
dialogue
is
that
it
includes
all
the
virtual
messages
in
all
the
protocols
that
support
any
single
primal
message
in
one
protocol,
all
the
way
down
to
the
physical
medium
and
back
up
again,
as
in
the
upper
part
of
g.
2.
Together,
the
supporting
protocols
in
the
down—up
view
de ne
a
Thread
for
the
top—level
primal
message.
The
term
‘dialogue’
therefore
can
be
used
interchangeably
for
what
goes
on
in
one
protocol
or
for
everything
that
occurs
in
the
Thread
that
supports
that
protocol.
Inasmuch
as
there
is
a
support
hierarchy
of
protocols,
there
is
an
inclusion
hierarchy
of
dialogues.
If
protocol
B
supports
protocol
A,
a
dialogue
conducted
within
protocol
B
is
part
of
the
dialogue
in
the
Thread
that
supports
protocol
A.
\.
D
—————>
D
Figure
3:
Multiplexed
and
diviplexed
channels.
Channel
D
supports
messages
from
both
B
and
C
protocols,
and
they
are
said
to
be
multiplexed
onto
channel
D.
Channel
A
is
said
to
be
diviplexed
onto.»channe1s
B
and
C,
because
its
messages
are
split
across
those
two
supporting
channels.
Protocols
may
sometimes
be
required
to
convey
primal
messages
from
more
than
one
source,
in
which
case
the
messages
are
said
to
be
multiplexed
onto
the
protocol.
Multiplexing
is
almost
universal
in
spoken
communi-
198
MARTIN
TAYLOR
5
DAVID
WA
cation;
a
talker
uses
voice
modulation
to
communicate
emotional
affect
at
the
same
time
as
the
modulated
words
communicate
the
substance.
The
opposite
also
is
common;
one
primal
message
may
be
supported
simultane-
ously
by
more
than
one
lower
protocol
(perhaps
by
voice
and
gesture,
for
example),
in
which
case
it
is
said
to
be
divipleared
as
shown
in
g.
1
(Taylor
1989;
Taylor
and
Waugh
1999).
1.3
Mutual
models
and
belief
structures
To say
that
the
originator
of
a
message
has
to
know
both
the
current
and
a
desired
state
of
the
recipient
is
to
say
that
the
originator
must
model
the
recipient.
To
sustain
an
accurate
model
of
the
changing
state
of
the
partner
requires
feedback,
and
more
so
the
more
complex
and
dynamic
the
aspect
that
is
modelled.
To
see
why,
we
must
consider
the
process
of
communication
from
the
viewpoint
of
the
communicating
partners.
At
a
low
level,
a
talker
may
be
fairly
sure
that
a
listener
will
identify
the
spoken
words
correctly,
at
least
if
they
are
clearly
enunciated.
This
assurance
is
possible
only
if
both
talker
and
listener
speak
the
same
language
(or
better,
dialect,
if
the
language
is
English!).
The
talker
can
judge
whether
they
do
share
a
language
only
by
some
prior
probing
or
from
externally
supplied
knowledge,
but
once
that
initial
phase
is
complete,
the
language
competence
of
the
listener
is
part
of
the
talker’s
model,
and
it
is
not
expected
to
change
rapidly.
At
this
level
the
model
determines
to
a
large
extent
a
talker’s
choice
of
words
and
even
of
vocal
clarity.
If
a
listener
has
a
poor
command
of
the
talker’s
language,
the
talker
will
use
simple
words
spoken
clearly,
but
if
the
listener
is
uent,
and
especially
if
the
listener
shares
the
same
social
circle
and
technical
knowledge,
the
talker
may
use
short
forms,
jargon,
and
rapid,
unclear
enunciation.
Either
way,
the
talker
will
be
able
to
believe
that
the
listener
will
correctly
identify
most
words,
without
feedback
for
each
word,
even
if
the
listener
does
not
understand
the
import
of
the
Words
at
higher
protocol
levels.
The
originator’s
model
of
the
recipient
is
central
to
how
the
originator
encodes
each
particular
primal
message
into
virtual
messages.
The
recip-
ient’s
model
of.the
originator
is
equally
central
in
the
decoding
process,
as
well
as
to
the
encoding
of
feedback
messages.
In
the
following
discus-
sion,
we
will
personalize
the
originator
and
recipient
as
Oliver
and
Rachel
respectively.
Oliver,
the
originator,
has
a
model
of
Rachel,
the
recipient.
His
model
is
a
set
of
beliefs,
some
held
strongly,
some
weakly.
It
includes
beliefs
about
what
Rachel
believes,
and
about
what
she
may
do
as
a
consequence
of
those
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
199
beliefs.
In
particular,
Oliver’s
model
of
Rachel
includes
beliefs
about
what
she
believes
about
him
and
his
beliefs.
All
these
beliefs
change
rapidly
during
the
communication
of
a
message.
When
Oliver
believes
that
Rachel’s
beliefs
(or
her
actions)
are
not
as
he
would
wish,
he
opens
communication
by
sending
a
virtual
message.
The
result
of
a
successful
communication
is
that
Oliver
comes
to
believe
that
Ra-
chel’s
beliefs
(or
actions)
have
changed
in
the
desired
way.
Oliver’s
model
of
Rachel
is
changed,
and
the
change
usually
is
con rmed
by
Rachel’s
reac-
tions
to
his
message.
The
virtual
message
that
a
protocol
must
encode
and
convey
is
simply
a
pattern
of
differences
between
Oliver’s
existing
beliefs
about
Rachel’s
beliefs
and
the
primal
message
(the
beliefs
Oliver
would
like
to
have
about
her
beliefs).1
If
Rachel’s
actual
beliefs
differ
from
Oliver’s
model
of
them,
then
Rachel
will
not
interpret
his
virtual
message
correctly
as
an
indication
as
to
how
her
beliefs
should
change.
Since
Rachel
cannot
know
for
sure
whether
Oliver’s
model
of
her
beliefs
is
exact,
Rachel
can
never
decode
a
virtual
message
by
inverting
Oliver’s
encoding.
But
Rachel
has
a
model
of
Oliver’s
beliefs
about
her,
and
she
can
use
it
to
judge
how
Oliver
intended
the
message
to
change
her
beliefs,
and
thereby
to
approximate
what
beliefs
Oliver
intended
her
to
arrive
at.
The
message
is
interpreted
in
stages,
as
her
approximation
improves
through
the
use
of
feedback,
thereby
achieving
Oliver’s
intentions
without
inverting
Oliver’s
encoding.
The
upshot
is
that
no
outside
observer,
no
dialogue
analyst,
can
really
ever
determine
what
messages
are
passing
between
two
people,
especially
if
those
people
know
each
other
well
and
their
common
goals
are
not
well
de ned.
If,
however,
they
do
not
know
each
other,
then
they
probably
use
default
models
of
each
other,
and
to
the
extent
that
the
outside
observer
knows
the
task
context
and
shares
the
same
defaults
(comes
from
the
same
cultural
milieu,
for
example),
the
observer
can
match
the
recipient’s
inter-
pretation
of
the
message,
or
determine
from
the
recipient’s
responses
the
manner
in
which
the
recipient
has
interpreted
the
message
differently.
This
is
the
usual
situation
in
the
analysis
of
information—seeking
dialogues.
1Notice
carefully
that
the
difference
is
not
between
Rachel’s
beliefs
and
Oliver’s
goal
for
what
Rachel’s
beliefs
should
be.
Neither
Oliver
nor
Rachel
has
the
information
that
would
allow
such
a.
comparison.
Oliver
can
model
Rachel’s
beliefs,
but
he
cannot
know
them.
Only
an
omniscient
external
observer
could
assuredly
judge
the
success
of
any
particular
communication.
200
MARTIN
TAYLOR
55
DAVID
WA
UGH
2
The
structure
of
a
protocol
node
Before
continuing
with
the
analysis
of
changing
beliefs
in
a
dialogue,
we
must
consider
protocols
and
the
structure
of
a
protocol
node.
We
de ne
the
protocol
used
by
a
protocol
node
as
a
set
of
conventionalized
forms
that
are
part
of
the
model
used
by
each
partner.
These
forms
assist
a
virtual
message
to
pass
effectively
between
the
communicating
parties.
They
include
both
the
internal
syntax
of
a
virtual
message
and
the
back
and
forth
flow
of
feedback
virtual
messages
that
allow
the
partners
to
determine
whether
the
primal
message
is
being
properly
received.
The
internal
syntax
of
a
single
Virtual
message
allows
the
recipient
of
the
message
to
assess
the
probability
that
it
has
been
correctly
received,
whereas
the
feedback
loop
allows
each
to
determine
how
well
the
other
is
satis ed
with
the
recipient’s
interpretation.
Sometimes
we
also
use
the
term
‘protocol’
more
loosely,
to
refer
to
the
processes
or
even
to
the
protocol
node
through
which
the
forms
of
the
protocol
are
implemented.
2.1
Information
sharing
within
and
among
protocols
We
earlier
divided
into
a
series
of
protocol
layers
the
process
by
which
the
physical
form
of
a
message
is
generated
from
an
original
communicative
goal
of
the
sender,
and
by
which
the
physical
form
is
reinterpreted
by
the
receiver
to
determine
the
originator’s
intent.
Each
layer
consists
of
a
transmitting
and
a
receiving
protocol
node
that
interchange
virtual
messages,
the
whole
loop
consituting
a
channel
through
which
the
originator’s
primal
message
is
communicated.
As
described
above,
each
protocol
node
has
a
Coder
and
Decoder,
each
of
which
uses
a
model
of
the
state
of
the
partner,
the
two
models
having
much
the
same
content.
For
conceptual
convenience,
we
de ne
a
separate
Model
entity
that
contains
those
beliefs
about
the
world,
the
partner,
the
task,
and
the
dialogue
that
are
used
by
either
Coder
or
Decoder,
and
relegate
to
the
Coder
and
Decoder
elements
the
processes
and
data
that
are
private
to
that
function.
Coder,
Decoder
and
Model
are
then
seen
as
three
distinct
components
in
any
protocol
node.
Each
protocol
node
may
be
considered
at
three
different
time
scales:
data
and
processes
that
are
concerned
with
its
permanent
capability,
those
that
are
concerned
with
the
current
dialogue
thread,
and
those
related
only
to
the
current
primal
message.
We
call
those
time
scales,
‘Capability’,
‘Thread’,
and
‘Active’
time
scales,
respectively.
In
graphic
depictions
of
protocol
nodes,
we
often
show
three
boxes,
each
box
containing
the
Coder,
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
201
Primal
Message
from
supported
layer
C
-
l
c
51
-
G
1-
Virtual
message
L9}
0
Ii...
ii?
Capability
Thread
ctive
Permanent
Duration
of
D
ion
of
one
dialogue
ssage
thread
Primal
message
to
supporting
layer
Figure
4:
Three
elements
of
a
protocol
node
divided
according
to
three
time
scales.
The
‘Capability’
time
scale
is
associated
with
the
permanent
abilities
and
knowledge
associated
with
the
node;
the
‘Thread’
contains
data
that
are
associated
with
the
current
thread
of
dialogue
(the
duration
of
the
longest
lasting
message
supported
by
this
node);
and
the
‘Active’
component
contains
only
those
data
concerned
with
the
virtual
message
currently
being
transmitted.
A
Transmitting
Node
is
shown,
but
a
Receiving
Node
looks
the
same
except
for
the
direction
of
the
messages.
Decoder
and
Model
for
one
of
these
three
time
scales,
as
shown
in
g.
4.
The
Model
element
of
the
node
can
be
taken
as
a
view
on a
larger,
global
Model
accessible
from
all
protocol
nodes,
rather
than
as
a
private
database.
Elements
of
this
global
Model
that
are
used
by
two
or
more
protocol
nodes
can
be
seen
as
shared
data,
rather
than
necessarily
being
communicated
by
overt
messages
between
the
nodes.
Seen
in
this
way,
the
Model
in
a
PN
is
rather
like
a
view
onto
a
blackboard.
Sharing
between
a
supported
(higher)
Active
Model
and
its
supporting
(lower)
Thread
Model
is
particularly
important.
At
any
moment
in
the
dialogue,
the
current
state
of
the
Active
Model
represents
a
state
of
belief
about
the
partner,
and
the
message
to
be
encoded
is
an
attempt
to
bring
this
state
to
a
desired
condition.
The
desired
state
of
belief
is,
by
de nition,
the
primal
message
for
any
supporting
protocol
that
might
help
to
bring
204
MARTIN
TAYLOR
8’
DAVID
WA
UGH
Primal
Normal
Primal
E-feedback
Feedbck
ls
Null
Abon
primary
Commit
End
OK?
dit
Acoepted
Resolution
continuing
problem
5
Figure
6:
A
greatly
simpli ed
sketch
of
the
General
Protocol
Grammar,
showing
the
major
arcs.
Circles
represent
stages
at
which
the
originator
of
the
primary
message
must
do
something,
squares
stages
at
which
the
recipient
of
the
primary
must
do
something.
state
grammar,
to
do
so
is
only
a
didactic
simpli cation,
and
the
reader
should
keep
this
fact
in
mind.
We
deal
with
it
more
realistically
later
in
the
chapter,
taking
into
account
the
fact
that
the
two
partners
often
overlap
their
turns,
and
that
the
so—called
state
transitions
are
fuzzy
rather
than
discrete.
The
construction
and
interpretation
of
any
message
at
any
level
incor-
porates
lexical,
syntactic,
semantic,
and
pragmatic
aspects.
Every
protocol
involves
a
lexicon
that
consists
of
the
range
of
message
elements
that
the
two
decoders
can
identify.
These
lexical
elements
are
more
likely
to
be
combined
in
some
ways
than
in
others.
These
differential
probabilities
de-
termine
a
syntax,
and
relate
pragmatically
to
the
states
of
the
Models
in
the
nodes.
Semantics
is
the
set
of
interrelationships
between
a
supporting
and
a
supported
level,
particularly
the
different
uses
to
which
the
different
lexical
items
may
be
put
in
different
contexts
(their
meanings
at
the
higher
level,
or
the
ways
a
given
meaning
can
be
executed
at
the
lower
level).
Anaphora
(in
two
forms)
and
ellipsis
are
seen
to
be
inherent
in
most
protocols,
rather
than
being
treated
as
anomalies
to
be
explained
away.
Ellipsis
is
the
omission
of
information
that
would
be
needed
for
a
com-
plete
presentation
of
the
message.
Anaphora
is
the
substitution
of
a
generic
LAYERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
205
pointer
to
information
that
the
recipient
is
expected
to
nd
elsewhere
than
in
the
realization
of
the
message.
Both
ellipsis
and
anaphora
occur
at
all
levels
of
abstraction.
Typically,
the
omitted
information
is
in
the
Model
of
the
partner’s
current
state.
When
the
primal
message
state
and
the
cur-
rently
modelled
state
have
elements
in
common,
those
elements
should
not
occur
in
the
virtual
message.
Anything
in
the
virtual
message
is
likely
to
be
interpreted
as
a
desired
change
in,
rather
than
a
con rmation
of
the
recipient’s
state.
An
informal
way
of
stating
this
rule
is:
If
it
ain’t
broke,
don’t
x
it.
The
ellipsis
/
anaphora
rule
is
really
a
restatement
of
something
akin
to
the
Gricean
maxims
of
quantity
and
quality
(Grice
1975).
Looking
at
com-
munication
this
way,
the
necessity
for
anaphora
and
ellipsis
becomes
almost
self-evident.
Both
require
that
the
originator
of
the
message
believe
the
required
information
to
be
immediately
available
to
the
recipient,
and
the
recipient
to
be
aware
of
this
belief
held
by
the
originator.
The
relation-
ship
between
anaphora
and
ellipsis
on
the
one
hand,
and
diviplexing
on
the
other,
is
discussed
in
more
detail
by
Taylor
and
Waugh
(1999).
2.3
The
GPG
as
a
dynamic
belief
structure
We
now
approach
the
same
GPG
from
an
entirely
different
viewpoint.
Rather
than
seeing
it
as
a
nite—state
grammar
that
represents
discrete
transitions
among
labelled
states,
we
consider
it
as
representing
continu-
ously
changing
beliefs
that
may
be
held
by
either
partner
about
three
well-
defined
propositions
relating
to
the
transmission
of
a
primal
message:
P1:
The
recipient
has
made
an
interpretation
of
the
primal
message.
~
P2:
Whenever
P1
is
true,
then
the
recipient’s
interpretation
will
be
or
is
adequate.
P3:
It
is
not
worth
continuing
to
try
to
reach
the
state
(P1
&
P2).
These
three
propositions
are
statements
of
fact.
When
both
parties
be-
lieve
P3
to
be
true,
the
communication
stops.
P3
could
be
true
for
several
different
reasons,
bothgood
and
bad
in
their
implications:
(P1
&
P2)
might
already
be
true
so
no
further
effort
need be
expended
(success),
external
factors
such
as
successful
completion
of
a
supported
message
might
have
made
this
message
irrelevant
(possible
success),
or
there
might
be
so
much
4
confusion
that
to
reach
(P1
&
P2)
would
be
a
considerable
effort
(aborting
the
communication;
i.e.,
failure).
Beliefs
about
these
three
propositions
can
208
MARTIN
TAYLOR
8
DAVID
WA
UGH
R1
E
SR(P1
&
P2
&P3))))
If,
at
any
time,
Oliver
comes
to
believe
-IR1,
that
Rachel
does
not
have
a
goal
to
complete
the
message
transmission
correctly,
communication
has
ceased
to
be
cooperative,
no
matter
what
Rachel’s
actual
goal
state
may
be.
Before
starting
the
message,
Oliver’s
state
must
also
include
at
least
02
E
O(‘IP3
S
R(‘\P3)
2
Oliver
at
least
weakly
believes
that
it
is
worth
trying
to
pass
the
message
and
that
Rachel
does
not
disbelieve
it
is
worth
trying,
and
at
the
same
time
Oliver
does
not
believe
both
that
Rachel
has
interpreted
the
message
(he
hasn’t
started
sending
it
yet)
and
that
she
believes
she
has
the
information
that
it
would
transmit.
The
reason
for
the
message
transmission
is
to
change
this
particular
belief
state
into
the
desired
state.
Sometimes
the
beliefs
that
form
part
of
02
are
called
‘felicity
conditions’
for
Oliver
to
send
the
message.
On
Rachel’s
part
the
felicity
conditions
are
simpler,
being
restricted
to
N
3
R(P3);
she
is
willing
to
listen.
In
Layered
Protocol
terms,
felicity
conditions
can
be
restricted
to
two:
there
is
a
discrepancy
between
one
partner’s
current
and
desired
beliefs
about
the
state
of
the
other,
and
there
is
a
means
for
reducing
this
discrepancy.
Even
in
non—cooperative
dialogue,
felicity
conditions
apply,
but
they
are
different.
When
Iago,
in
Shakespeare’s
Othello,
drops
Desdemona’s
handkerchief
for
Othello
to
nd,
his
felicity
condition
does
not
include
W
3
I
ag0(W
3
Othello( P3)),
since
part
of
the
point
of
his
deceit
is
that
Othello
never
knows
that
the
message
has
been
sent.
The
verbal
interpretation
of
N
3
R(-wP3)
is
“Rachel
at
least
weakly
believes
that
it
is
worth
continuing
the
communication
or
has
no
relevant
belief.”
P3
says
that
it
is
not
worth
continuing,
so
if
Rachel
has
an
opinion
on
the
matter,
she
believes
it
is
worth
continuing.
Before
Oliver
begins
communicating,
it
is
most
probable
that
Rachel
does
not
even
know
that
Oliver
has
a
message
awaiting
transmission.
If
she
does
not
know
it
exists,
she
could
hardly
be
expected
to
believe
that
it
is
worth
trying
to
interpret
it.
Only
after
Oliver
has
given
her
reason
to
believe
that
he
wants
to
communicate
could
she
be
expected
to
have
a
belief
about
whether
she
should
put
some
effort
into
cooperating
with
him.
Even
that
minimal
belief
depends
on
her
model
of
Oliver.
Are
his
communications
likely
to
be
worth
worrying
about?
One
might
ask
why
the
recursion
of
belief
is
carried
only
to
three
levels.
The
answer
is
purely
pragmatic.
In
the
cases
we
have
examined,
further
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
209
levels
of
recursion
do
not
cause
differences
in
behaviour.
In nite
recursion
is
possible
in
principle,
and
further
study
might
show
cases
in
which
a
deeper
recursion
than
three
levels
is
needed.
However,
one
might
hazard
a
guess
that
there
is
a
limit
imposed
by
the
uncertainty
of
beliefs
about
a
belief
held
by
another
person.
Perhaps,
beyond
three
levels
there
is
so
much
uncertainty
that
to
behave
differently
for
different
strengths
of
fourth-level
recursive
beliefs
would
usually
be
unwarranted.
The
three
propositions
say
nothing
about
the
content
of
the
message
that
is
being
transmitted.
But
the
message
is
sent
because
Oliver
believes
Rachel
to
have
a
belief
structure
that
must
be
changed
if
she
is
to
act
on
the
world
in
a
way
that
Oliver
wants.
The
discrepancy
between
what
Oliver
now
believes
she
believes
and
what
he
Wants
to
believe
she
believes
determines
the
message
content
(e.g.
Beun
1991).
If,
on
some
matter,
Oliver
wants
Rachel
to
believe
:1:,
and
believes
Rachel
to
believe
y
(2:
and
y
being
structured
propositions),
then
the
discrepancy
can
be
described
as
the
proposition
Pym.
Pya:
is
the
content
of
the
virtual
message
that
Oliver
will
use
to
encode
his
primal
message
:2.
If
Rachel
actually
does
believe
y,
then
after
she
receives
and
believes
Pym,
she
will
believe
an
if
Oliver’s
model
of
her
is
correct.
Using
this
de nition
of
Pym,
another
way
of
describing
the
starting
state
ofthe
communication
is
(N
Z
O(W
3
R(Py:c)))
&
SO(g,SO(S'R(Pya:))).
Oliver
does
not
believe
that
Rachel
believes
Pym,
and
has
a
goal
that
she
should
(which
would
bring
her
to
believe
Successful
completion,
as
de ned
earlier,
implies
SO(P1
&
P2).
The
truth
of
(P1
&
P2)
implies
SR(Pyz),
so
successful
completion
implies
that
the
goal
S
O(
g,
S
O(SR(Pyw))
has
been
satis ed.
Now
let
us
look,
by
way
of
example,
at
some
of
the
nodes
and
arcs
of
the
GPG
as
states
and
changes
of
states
of
belief.
It
would
be
tedious
to
examine
the
whole
of
the
GPG
here,
since
we
have
discovered
some
47
plausible
instantiations
of
23
arcs
connecting
11
nodes
in
the
full
GPG.
But
as
an
example,
consider
the
node
labelled
‘Is
it
what
I
want?’
in
g.6.
this
point,
the
originator
has
already
sent
a
primary
message
and
the
recipient
has
provided
some
kind
of
feedback
to
indicate
W
3
R(P1
&
P2).
Possibly
there
has
been
some
interchange
in
respect
of
a
temporary
problem,
but
whenever
this
node
of
the
GPG
is
occupied,
Oliver
believes
that
Rachel
at
least
weakly
believes
the
message
to
have
been
adequately
interpreted.
That
is
what
it
means
for
Oliver
to
be
at
this
node
rather
than
at
the
one
marked
‘Fix
Problem’
in
g.
6.
So,
at
the
node
‘Is
it
what
I
want?’,
we
have
at
least
W
3
O(W
g
R(P1
&
P2
&
-wP3)).
It
is
Oliver’s
turn
to
do
something,
but
what
he
does
depends
on
the
rest
212
MARTIN
TAYLOR
63
DAVID
WA
UGH
providing
information
that
helps
them
to
achieve
their
own
goals
for
the
communication.
In
human
communication,
it
is
important
that
higher—level
protocols
be
given
even
partial
information
from
lower—level
ones
as
soon
as
it
is
available.
Likewise,
the
expectations
of
lower-level
protocols
are
affected
by
ongoing
interpretations
at
higher
levels.
Sometimes
these
expectations
can
become
so
strong
that
the
recipient
has
a
very
strong
belief
about
Pym
before
the
originator
has
transmitted
the
message:
“Don’t
bother...I
know
what
you
are
going
to
say!”
In
such
a
case,
SR(P3
&
W
3
O( P3))
and
the
feedback
message
is
attempting
to
bring
about
SR(.S'O(P3)),
which
would
abort
the
message
transmission.
Sometimes,
such
abort
messages
fail,
because
the
originator
does
not
come
to
believe
that
the
recipient
has
a
correct
interpretation,
and
therefore
N
Z
O(P2)
which
leads
to
W
3
O(R(P2))
and
SO(-»P3).
The
originator
then
has
the
job
of
changing
the
recipient’s
belief
structure,
and
the
communicative
cooperation
may
have
a
period
of
con ict,
as
each
partner
tries
to
change
the
other’s
opinion.
Such
issues
are
beyond
the
scope
of
this
chapter.
3
Analysis
of
information
dialogues
We
are
nally
in
a
position
to
sketch
an
approach
to
the
analysis
of
the
two
information
dialogues,
though
we
will
look
only
at
selected
interchanges.
Both
dialogues
are
requests
for
information
from
a
client
to
a
domain
expert.
In
analysing
these
and
other
information
dialogues,
we
have
found
the
simple
structure
of
g.
7
to
be
both
necessary
and
sufficient,
taking
for
granted
the
protocols
at
lower
levels
of
the
Threads.
(More
complex
structures
are
required
for
other
kinds
of
dialogue
and
for
multi-person
conversation).
Figure
7
shows
the
information
service’s
(wizard’s)
Threads.
The
client’s
Threads
would
be
the
same,
apart
from
the
interchange
of
transmitting
protocols
and
receiving
protocols.
There
are
two
main
Threads,
one
to
transmit
queries
from
the
client
to
the
information
service
(the
wizard),
and
one
to
transmit
answers
in
the
other
direction.
Parts
of
these
Threads
carry
the
feedback,
or
protocol,
messages.
Two
levels
of
protocol
are
considered
for
each
direction
(there
are
many
other
levels
such
as
‘word’
or
‘phrase’
below
these
two,
but
they
are
not
of
speci c
interest
in
this
analysis).
In
the
receiving
Thread
the
two
protocols
are:
one
to
perceive
proposi-
tions
(Proposition
In)
and
one
to
perceive
queries
based
on
those
proposi-
tions
(Query
Assembler).
The
Query
Assembler
collates
propositions,
and
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
213
when
it
has
a
query
that
it
believes
to
represent
the
intent
of
the
questioner
(i.e.
S
=
Wz'zard(P1
&
P2)),
it
passes
the
query
to
the
information
sys-
tem.
We
call
the
information
required
to
complete
a
syntactically
complete
query
a
query
frame,
although
it
may
not
correspond
exactly
to
the
notion
of
“frame”
as
it
is
used
technically
in
some
forms
of
Arti cial
Intelligence.
A
query
frame
contains
data
for
the
vehicle
mode
(air
or
train),
departure
and
arrival
time,
identi cation
number,
and
the
like.
If
one
or
more
of
these
is
marked
as
unknown
and
wanted
information,
the
frame
contains
a
query.
Otherwise
it
contains
a
response.
In
the
transmitting
Thread,
the
upper
layer
is
a
Response
Formulator,
supported
by
Proposition
Out.
The
Response
Formulator
is
the
inverse
of
the
Query
Assembler,
in
that
it
accepts
whatever
information
the
system
provides,
and
selects
items
that
the
Thread
Model
of
the
Response
Formula-
tor
showed
as
having
been
unknown
in
the
original
Query
frame.
Response
Formulator
and
Query
Assembler
share
much
of
their
Thread
Model
data.
Proposition
In
and
Proposition
Out
are
multiplexed
protocols,
each
sup-
porting
feedback
messages
in
addition
to
the
queries
and
responses
that
form
the
main
dialogue
content.
3.1
The
Alicante
dialogue
The
Alicante
dialogue
was
originally
in
Dutch.
The
following
English
trans-
lation
was
provided
by
Bunt
(1989).
W1:
Schiphol
information.
C1a:
I
have
booked
for
flight
IB
885,
next
Saturday,
to
Alicante.
C1b:
What
time
should
I
report
at
Schiphol?
W2:
You
should
check
in
half
an
hour
before
departure
at
the
latest.
C2:
So
between
what
time
and
what
time?
W3:
Between
twelve
and
twelve-thirty.
C3:
Do
you
also
have
information
about
departure
and
arrival
times
of
trains?
W4:
In
Holland?
C4:
Yes.
W5:
I
do
C5:
What
is
the
last
train
from
Breda
I
can
take
to
be
in
time
for
ight
IB
885'?
W6:
The
train
of
12.06.
214
MARTIN
TAYLOR
65
DAVID
WA
UGH
Information
System
(IS)
Capability
Thread
‘
'
Capability
Thread
Active
"
l
l:l
A!’
Response
Formulator
(RF)
Query
Assembler
(Q
Capability
Thread
Active
Capabili
Thread
Active
El
gh ‘
E
l1_l|:|
l3.||:||:l
Proposition
ln
(Pl)
Proposition
Out
(PO)
Figure
7:
The
wizard’s
Thread
network
supporting
the
simple
information
di-
alogues
analyzed
in
this
chapter.
The
Thread
supporting
the
Query
Assembler
is
marked
by
small
circles,
and
the
Thread
supporting
the
Response
Formulator
by
small
crosses.
The
two
Threads
share
common
pathways
at
and
below
the
Proposition
level.
C6:
What
is
the
arrival
time
in
Alicante
W7;
17.00.
C7:
What
is
the
duration
of
the
bus
travel
A1icante—Benidorm?
W8:
We
don’t
have
information
about
that.
C8:
Thank‘you.
W9:
You’re
welcome.
Bunt
points
out
that
voice
interchanges
between
humans
are
typically
more
verbose
than
this,
but
this
is
not
necessarily
so
when
the
human
identi es
the
partner
as
a
computer
(Morel
1989)
and
it
may
not
be
so
for
typed
interchanges
in
general.
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
215
Let
us
consider
the
development
of
the
rst
query
in
the
Alicante
dia-
logue
(up
to
utterance
W3).
The
ringing
of
the
telephone
leads
the
informa-
tion
system
to
expect
that
a
caller
will
pose
a
time—table
query.
But
rst,
the
caller
is
assumed
to
have
an
unspoken
preliminary
query:
Did
I
dial
the
right
number?
which
is
answered
by
the
greeting
W1:
Schiphol
information.
W1
is
E-feedback
(Expectation
feedback,
see
Engel
and
Haakma
1993)
that
lets
the
caller
know
that
the
answerer
is
prepared
to
entertain
timetable
queries
(ensuring
W
3
C(W
_<_
Wizard(-vP3))).
This
greeting
is
initiated
as
a
consequence
of
a
default
client
model
accessed
by
the
Response
For-
mulator
(RF)
protocol,
not
as
a
consequence
of
anything
in
the
timetable
database.
In
C1a:I
have
booked
for
ight
IB
885,
next
Saturday,
to
Alicante
the
client
does
not
ask
for
information,
but
provides
three
propositions.
The
client’s
model
of
the
information
system
includes
that
a
query
requires
this
kind
of
information,
and
that
the
information
system
does
not
yet
have
it.
These
propositions
can
be
used
by
the
Query
Assembler
(QA)
in
at
least
three
different
ways.
Firstly,
the
QA
Active
Decoder
is
building
a
query
from
one
of
a
small
number
of
permissible
query
frames,
and
these
propositions
can
fill
some
of
the
slots
in
one
or
more
of
those
frames
during
the
development
of
the
rst
query.
Secondly,
the
propositions
may
be
stored
in
the
QA
Thread
Decoder
for
use
as
anaphora
in
later
queries
in
the
same
dialogue.
Thirdly,
and
importantly
for
the
proper
interpretation
of
C6,
the
effect
of
the
propositions
is
stored
in
the
Thread
Model,
which
builds
up
a
representation
of
the
pragmatic
situation
in
a
way
accessible
not
only
to
the
QA
but
also
to
other
protocols,
notably
the
Response
Formulator.
In
Clb:
What
time
should
I
report
at
Schiphol?,
the
client
asks
a
speci c
question,
but
the
answer
given
in
W2:
You
should
check
in
half
an
hour
before
departures
at
the
latest.
is
shown
to
be
inadequate
(N
2
C(P2))
by
the
client’s
response
in
C2
So
between
what
time
and
what
time
.9.
Two
possible
analyses
could
account
for
the
inadequacy
of
the
W2
answer,
and
its
later
elaboration.
One
is
that
the
QA
decoder
did
not
use
the
information
provided
at
Cla,
and
passed
to
the
information
system
a
default
query
that
would
be
suited
to
any
traveller
“How
long
before
take—off
should
a
passenger
report”,
for
which
the
information
system
provided
W2
as
a
default
answer.
The
reaction
(N
2
C(P2))
at
C2
then
cues
the
QA
to
use
the
information
in
its
Thread
Coder
in
constructing
a
more
speci c
query,
to
which
the
information
system
provides
a
speci c
answer.
The
other
analysis
is
more
interesting,
and
more
human.
It
is
that
the
QA
did
produce
a
speci c
query,
to
which
the
information
system
provided
a
speci c
answer,
but
having
no
speci c
model
of
the
user
the
RF
used
216
MARTIN
TAYLOR
8
DAVID
WA
UGH
its
default
model
and
provided
only
the
information
it
assumed
would
be
unknown
to
the
client.
The
RF
would
have
in
its
Thread
Model
that
the
user
knew
the
ight
to
be
IB885
on
Saturday,
and
would
have
in
its
Capability
Model
the
default
belief
that
clients
who
know
ight
numbers
also
know
ight
times.
From
these
it
could
conclude
(in
this
case
wrongly)
that
the
only
item
unknown
to
the
client
was
the
relation
between
the
take—off
time
and
the
check-in
time,
which
is
valid
for
more
ights
than
the
speci c
one
in
question.
The
primal
message
for
RF
was
that
the
client
should
know
the
check—in
time;
the
difference
between
this
and
the
wizard’s
current
belief
about
the
client
was
that
the
client
did
not
know
the
relation
between
departure
time
and
check-in
time.
In
this
analysis
the
client’s
reaction
C2
is
received
by
the
RF
Active
Decoder
from
the
Proposition
In
as
an
ambiguity
correction
request
leading
to
the
‘Fix
Problem’
node
of
the
GPG
(see
g.
6),
which
the
RF
can
immediately
satisfy
from
information
already
in
its
Thread
Model.
After
some
questions
and
answers
about
train
times
(C3
to
W6),
the
client
returns
to
the
question
of
the
Alicante
ight.
But
the
question
at
C6:
What
is
the
arrival
time
in
Alicante?
does
not
mention
the
ight,
and
the
immediately
preceding
discussion
has
been
about
a
train
from
Breda.
If
C6
were
to
be
interpreted
using
syntactic
anaphora
using
the
Thread
Decoder,
which
contains
the
last
used
lexical
item
of
any
type,
in
this
case
‘vehicle’,
the
interpretation
might
be
What
time
does
the
12:06
train
from
Breda
arrive
in
Alicante?
It
is
at
this
point
that
the
difference
between
the
functions
of
Thread
Decoder
and
Thread
Model
becomes
important.
In
the
Thread
Decoder
there
is
no
reference
to
pragmatic
reality,
but
only
to
the
results
of
past
decoding.
In
it
the
most
recently
mentioned
vehicle
is
the
train.
On
the
other
hand,
in
the
Thread
Model,
there
is
a
frame
in
which
Alicante
and
ight
IB885
on
Saturday
co—occur.
The
reference
to
Alicante
evokes
this
frame,
in
which
the
arrival
time
slot
is
un lled,
and
thus
evokes
the
ight
and
would
do
so
even
if
there
existed
a
train
from
Breda
to
Alicante
leaving
Breda
at
12:06.
No
reference
to
the
real-world
possibilities
or
the
content
of
the
information
system’s
database
is
required
for
a
proper
interpretation
of
this
delayed,
pragmatic
anaphora.
Bunt
(1989)
distinguishes
dialogue
control
acts
from
factually
informa-
tive
acts,
and
treats
W1,
Cla,
C3
to
W5,
and
W8
to
W9
as
dialogue
control
acts,
or
even
‘dialogue
sugar’.
Within
the
Layered
Protocol
framework,
we
make
no
such
distinction2,
seeing
all
the
utterances
as
having
functions
that
2LP
theory
does
have
a
category
of
messages
we
call
‘control
messages’
but
they
are
not
at
all
the
same
as
Bunt’s
dialogue
control
acts.
In
LP
theory,
control
messages
control
LAYERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
217
directly
or
indirectly
provide
the
conditions
that
allow
the
primal
message
to
be
transmitted.
Even
direct
requests
for
information
cannot
have
their
intended
impact
unless
the
Partner
Models
correctly
assert
that
the
partner
is
in
a
state
that
will
allow
the
query
to
be
correctly
interpreted.
Informa-
tive
responses
cannot
be
correctly
formulated
without
reference
to
what
the
client
knows
and
needs
to
know.
The
messages
that
Bunt
calls
dialogue
control
are
to
us
more
reasonably
interpreted
as
messages
that
set
and
test
the
states
of
the
various
models
used
in
message
coding
and
decoding.
3.2
The
restaurant
dialogue
The
restaurant
dialogue
(see
Appendix)
does
not
ring
true
as
the
kind
of
dialogue
that
a
real
client
seeking
information
would
have
with
a
Yellow
Pages
system.
It
may
be
characterized
as
only
partly
cooperative,
as
the
client
seems
more
to
be
probing
for
weaknesses
in
the
information
system
than
to
be
seeking
wanted
information.
The
information
system
models
the
client,
by
default,
as
having
a
primary
goal
of
nding
a
suitable
restaurant,
and
nothing
in
the
dialogue
suggests
that
this
default
belief
is
ever
modi ed.
A
human
information
provider
confronted
with
such
a
client
probably
would
realize
that
the
client
is
more
interested
in
the
information
provider
than
in
the
information
provided.
If
the
wizard
were
not
pretending
to
be
a
computerized
information
system,
and
acted
in
a
more
human
way,
the
dialogue
might
well
have
included
some
probing
by
the
wizard
to
determine
the
real
goal
being
pursued
by
the
client.
Such
a
mismatch
between
the
goal
held
by
one
partner
and
the
model
of
that
goal
held
by
the
other
partner
can
lead
to
devastating
misunderstand-
ings.
In
this
case,
it
does
not,
perhaps
for
two
reasons:
(1)
the
information
system
can
cooperate
with
no
goal
other
than
the
default
assumption
that
the
client
wants
to
nd
a
suitable
restaurant,
and
(2)
the
client’s
testing
is
restricted
(until
near
the
end)
to
determining
which
variants
of
the
default
goal
could
be
satis ed
by
the
information
system.
W1:
Welcome
to
the
yellow
pages
access
system.
Hit
the
<
esc
>
key
to
send
your
text
to
the
system.
W2:
Hello.
Please
formulate
your
query.
The
rst,
welcome,
message
has
the
same
function
as
“Schiphol
information”
the
Thread
structure
of
an
interface,
and
they
are
treated
by
the
protocols
exactly
as
are
any
other
messages.
The
only
difference
is
that
their
sources
and
targets
are
processes
concerned
with
the
interrelationships
of
protocols
(the
Thread
network).
We
are
not
concerned
with
such
messages
in
analyzing
these
information
dialogues.
220
MARTIN
TAYLOR
59'
DAVID
WA
UGH
a
limitation
to
Paris,
but
the
Query
Assembler
cannot
accept
the
query
about
Japanese
restaurants
as
being
fully
speci ed.
Pragmatically,
Paris
is
so
big
that
the
wizard’s
default
model
of
the
client
includes
that
clients
normally
wish
to
restrict
the
search
to
a
speci c
part
of
Paris.
Within
the
Query
Assembler
(QA)
GPG,
the
wizard’s
Client
Model
included
the
belief
that
the
client
believed
that
a
query
had
been
made
(S
C(P1
&
P2)),
but
in
the
wizard’s
view
it
was
incomplete:
SWizard(P1
&
-1P2
&
S'C(P1
&
P2)).
There
was
a
mismatch
between
S'Wizard(-=P2)
and
SWizard(S'C(P2)).
Therefore
the
wizard
moved
to
the
node
‘Fix
Problem’,
using
an
instantiation
of
the
problem
are
that
provided
the
client
with
in-
formation
as
to
the
nature
of
the
problem.
The
frame
for
a
fully
speci ed
query
seems
to
require
at
least
an
indication
as
to
whether
there
is
a
pre-
ferred
district.
Japanese
Restaurant
and
In
Paris
have
at
this
point
become
part
of
the
Query
Assembler’s
Thread
Model,
as
well
as
being
prepared
for
use
as
anaphoric
or
elliptic
reference
in
the
QA
Thread
Decoder,
but
they
do
not
suffice
to
specify
a
query.
0m:06s
C4:
No.
0m:32s
W5:
Wait
a
moment.
I’m
looking...
4m:43s’C5:
In
the
6th
District.
Om:38s
W6:
Here
is
the
list:
at
which
point
the
wizard
starts
providing
a
list
of
Japanese
restaurants,
none
of
which
is
in
the
6th
district.
The
client’s
No
was
suf cient
to
complete
a
query
frame,
when
put
together
with
the
information
that
a
Japanese
restaurant
was
required.
The
client’s
No
changed
the
wizard’s
SWizard( P2)
to
S'Wizard(P2),
satisfying
the
communicative
goal,
since
P1
&
P2
implies
P3.
In
the
QA
GPG,
the
client
had
provided
a
‘Resolve
Problem’
message
that
led
to
a
set
of
beliefs
sufficient
for
the
wizard
to
commit
the
message
using
only
a
direct
acknowledgment
that
admitted
of
no
further
elaboration
or
message
editing
on
the
part
of
the
client.
The
wizard
bypassed
the
‘Is
it
what
I
want?’
node,
using
a
‘Commit’
arc
not
shown
in
the
simpli ed
GPG
of
g.
4,
which
parallels
the
‘Abort’
are
from
the
‘Is
it
xed?’
node.
More
than
four
minutes
later,
after
waiting
for
the
information
system
to
search
a
presumably
large
database,
the
client
revises
the
model
of
the
information
system’s
capability,
now
deciding
that
an
‘anywhere’
question
is
legal
but
impractical.
As
a
consequence
of
this
revision,
the
client
asks
a
new
question
at
C5.
This
new
question,
according
to
LP
theory,
is
not
an
amendment
of
the
previous
one,
inasmuch
as
the
previous
question
has
been
committed
and
is
being
acted
upon.
But
to
ll
in
the
ellipses,
it
uses
the
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
221
content
of
the
wizard’s
QA
Thread
Decoder,
which
is
presumed
to
be
much
the
same
as
that
of
the
client’s
transmitting
QA
Thread
Coder.
Both
have
“Japanese”
as
the
most
recent
lexical
entries
for
restaurant
type,
and
“any-
where”
as
the
most
recent
for
district.
Given
that
queries
about
restaurants
are
the
only
possible
communication
that
the
client
can
initiate
with
the
wizard,
“In
the
6th
district”
can
only
be
a
lexical
entry
for
the
“district”
item
in
a
query
frame.
Inasmuch
as
the
rest
of
the
frame
exists
identically
in
both
the
Thread
Model
and
the
Thread
Decoder
of
the
wizard’s
QA
node,
the
phrase
constitutes
a
legal
query.
But
the
wizard
does
not
abort
the
actions
that
it
started
after
the
earlier
query.
From
the
wizard’s
point
of
view
it
is
ambiguous,
at
first,
whether
the
new
query
derived
from
C5
should
result
in
aborting
the
old,
since
the
client
may
actually
want
a
wider
range
and
be
restricting
the
query
only
as
a
con-
sequence
of
changing
the
model
of
the
wizard
to
include
that
the
wizard
cannot
do
wide—ranging
searches
in
a
reasonable
time.
Considerable
intelli-
gence
is
at
work
here,
in
that
the
wizard
may
continue
the
search
operation
as
a
consequence
of
modelling
the
client
to
have
wrongly
remodelled
the
capability
of
the
information
system.
If
the
wizard
knows
that
the
search
will
end
soon,
a
completed
answer
to
the
earlier
question
may
both
satisfy
the
client
(the
wizard’s
model
of
the
client
is
that
the
client
wants
the
wide
ranging
answer),
and
allow
the
client
better
to
model
the
capabilities
of
the
information
system.
On
the
other
hand,
continuing
the
search
may
be
the
consequence
of
a
much
less
subtle
inability
of
the
system
simulated
by
the
wizard
to
abort
a
search
once
started.
At
this
point,
neither
the
client
nor
the
analyst
can
tell
why
the
wizard
provides
the
list
of
restaurants
outside
the
6th
district,
but
whatever
the
reason,
the
client
now
can
make
a
more
precise
model
of
the
capabilities
of
the
information
system,
and
refrains
from
making
any
more
“all
district”
enquiries
until
late
in
the
probing
part
of
the
dialogue.
At
the
end
of
the
list
of
Japanese
restaurants
in
the
6th
district,
an
interesting
colloquy
occurs.
During
the
wait,
the
client
has
asked
for
a
restaurant
(modelled
as
Japanese)
on
a
speci c
street,
Rue
M.
LePrince
in
the
5th
district:
end
of
W7:
PALAIS
IMPERIAL
(1)
46
33
07
31
39
R
MAZARINE
6E
I
will
carry
on
looking
for
Rue
M.
LePrince.
0m:38s
C7:
The
Palais
Imperial
is
Chinese
1m:13s
C7a:
Armenian
Restaurant
0m:34s
C7b:
Stop
looking
on
Rue
M.
LePrince
222
MARTIN
TAYLOR
63'
DAVID
WA
UGH
0m:36s
C7c:
Look
for
an
Armenian
Restaurant
1m:50s
W8:
I
do
not
know
any
Rue
M.
LePrince
in
the
5th
District,
but
I
know
a
Rue
M.
LePrince
in
the
6th.
I
am
stopping
the
search.
As
for
the
Palais
Imperial,
I
have
it
down
as
a
Japanese
Restaurant.
I
cannot
look
for
“an
Armenian
Restaurant,”
I
do
not
have
sufficient
information
in
my
database.
The
information
system
now
has
in
its
Response
Formulator
(and
thus
the
Query
Assembler)
Thread
Model
the
fact
that
the
client
has
been
in-
formed
that
the
Palais
Imperial
is
a
Japanese
restaurant.
The
client
also
has
this
fact
in
the
corresponding
Thread
Models,
and
therefore
the
client’s
model
of
the
information
system
is
that
the
information
system
has
in
its
world
model
that
the
Palais
Imperial
is
a
Japanese
restaurant.
C7
seems
to
encode
a
totally
new
primal
message,
not
a
simple
part
of
the
preceding
dialogue.
This
comment
could
have
been
delayed
until
the
end
of
the
whole
dialogue
without
affecting
any
of
the
other
interchanges.
C7
is
not
even
a
query,
thus
violating
the
fundamental
assumption
that
only
queries
about
restaurants
are
suitable
messages
for
the
client
to
initiate.
It
represents
the
client’s
belief
about
a
state
of
the
information
system’s
data-
base
(The
Palais
Imperial
being
Japanese)
that
differs
from
the
state
of
the
database
that
the
client
would
like
to
see
(The
Palais
Imperal
being
Chi-
nese).
No
matter
which
is
in
fact
the
case,
neither
state
of
the
information
system’s
database
affects
the
client’s
ability
to
nd
a
suitable
restaurant,
though
it
might
affect
the
answer
the
information
system
would
give
to
a
similar
query
on
a
later
occasion.
At
C7a,
the
client
attempts
to
abort
an
ongoing
search,
using
the
fact
that
the
Thread
model
of
the
QA
must
include
the
fact
that
a
search
was
requested
for
a
Japanese
restaurant
on
Rue
M.
LePrince
in
the
5th
district.
The
wizard
makes
no
response,
and
the
client
may
not
believe
that
the
message
was
received
(W
>
C(P1)).
Previously,
at
C5,
an
attempt
to
abort
an
ongoing
search
had
also
re-
sulted
in
no
response
and
no
abortion
of
the
search,
so
another
possibility
is
that
the
client’s
model
of
the
information
system’s
capability
may
now
include
that
searches
cannot
be
aborted
and
that
feedback
does
not
occur
during
a
search.
Since
all
other
messages
seem
to
be
correctly
received,
at
least
at
the
Proposition
In
protocol,
we
would
have
S
=
C(P1
&
P2)
at
that
level.
However,
the
client
tries
again
to
abort
the
search,
by
requesting
a
new
search,
so
this
alternative
is
at
least
not
strongly
believed
by
the
client.
Once
again,
the
client
uses
the
presumed
contents
of
the
wizard’s
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
223
QA
Thread
Decoder
to
substitute
Armenian
for
Japanese.
The
responses
to
this
sequence
of
attempted
interrupts
by
the
client
await
the
completion
of
the
search
for
Rue
M.
LePrince
in
the
5th
district.
Apparently
the
information
system
used
the
Guyomard
and
Siroux
(1989)
approach
of
relaxing
the
query,
and
nding
no
Rue
M.
LePrince
in
the
5th
district,
searched
in
the
6th,
which
the
system’s
model
of
the
client
had
already
in
focus.
One
of
the
elements
of
the
Thread
Model
that
we
have
not
so
far
dis-
cussed
is
its
Priority
Output
Queue,
or
POQ
(Edwards
and
Mason
1989).
Typically,
a
set
of
messages
to
be
output
are
held
in
some
kind
of
priority
order,
often
time-sequential.
But
since
Rue
M.
LePrince
is
in
focus
in
the
Response
Formulator
Thread
Model,
the
response
to
the
instruction
to
stop
the
search
for
it
is
given
a
higher
priority
in
the
POQ
than
is
the
response
to
the
preceding
comment
about
the
Palais
Imperial
being
Chinese.
One
of
the
criteria
for
priority
on
a
message
in
a
POQ
is
that
it
should
use
so
far
as
possible
those
lexical
items
that
are
still
available
as
anaphoric
or
elliptic
references
in
the
Thread
Coder.
In
this
case,
Rue
M.
LePrince
has
most
claim,
and
by
stating
I
am
stopping
the
search
before
going
on
to
the
Palais
Imperial
proposition,
the
Response
Formulator
has
no
need
to
provide
any
further
identi cation.
If
it
had
delayed,
such
a
response
might
have
been
taken
to
refer
to
stopping
the
search
for
the
Armenian
restaurant,
which
appears
not
to
have
started,
to
judge
from
the
nal
comment
of
W8.
The
nal
item
in
W8
is
a
little
ambiguous,
in
that
it
does
not
enable
the
client
to
determine
whether
it
is
the
query
that
is
inadequately
speci ed
(moving
to
the
‘Fix
Problem’
node
of
the
GPG
in
the
Query
Assembler)
or
the
ability
of
the
information
system
that
is
inadequate
(the
query
having
V
been
accepted
and
silently
committed).
This
ambiguity
is
emphasized
at
C8—W1O
by
the
willingness
of
the
information
system
to
search
for
a
Tibetan
restaurant,
speci ed
no
more
precisely
than
the
Armenian
that
it
refuses
to
seek.
If
there
were
no
Armenian
restaurants
in
the
database,
an
appro-
priate
response
would
have
been
to
say
so,
helping
the
client
to
enhance
the
model
of
the
information
system’s
capabilities.
If
a
nationality
by
itself
were
inadequate
as
a.
speci cation,
why
did
it
accept
the
task
of
seeking
both
a
Japanese
and
a
Tibetan
restaurant
with
no
further
speci cation‘?
A
legitimate
client
would
probably
have
regarded
the
answer
to
the
Armenian
request
as
ambiguous,
and
asked
for
clari cation.
But
to
the
outside
ob-
server
it
is
clear
from
this
point
onward
that
the
client’s
goal
is
more
to
cause
the
information
system
to
fail
than
to
cooperate
with
its
presumed
goal
of
helping
the
client
to
nd
a
restaurant.
224
MARTIN
TAYLOR
63
DAVID
WA
UGH
3.3
Analysis
of
the
dialogues:
nal
comment
In
analyzing
the
Restaurant
dialogue,
there
has
been
little
need
to
refer
to
the
Proposition
level
protocols.
They
provide
no
special
di iculties.
The
Restaurant
dialogue
does
show
perhaps
more
clearly
than
the
Alicante
di-
alogue
the
role
of
the
Thread
Decoder
in
the
use
of
anaphoric
and
elliptic
references
having
content
such
as
“The
desired
location
is
the
6th
district.”
Such
propositions
are
the
lexical
items
of
the
Query
Assembler
and
the
Re-
sponse
Formulator.
The
Restaurant
dialogue
also
demonstrates,
perhaps
more
clearly
than
the
Alicante
dialogue,
the
close
coupling
between
the
Thread
models
in
the
Query
Assembler
and
the
Response
Formulator.
To-
gether,
these
two
dialogues
illustrate
some
fairly
simple
examples
of
the
use
of
the
belief
structures
of
the
GPG,
mainly
in
the
Query
Assembler
protocol.
4
Conclusions
This
chapter
has
presented
an
introduction
to
some
aspects
of
the
Layered
Protocol
theory
of
communication,
and
has
sketched
its
use
in
the
analysis
of
two
simple
information—seeking
dialogues.
Layered
Protocol
theory
evolved
from
a
long
historic
tradition
in
psy-
chology,
that
people
organize
their
world
in
a
hierarchy
of
levels
of
abstrac-
tion.
It
asserts
that
when
people
communicate,
their
primary
intentions
are
to
affect
some
aspect
of
their
partner’s
knowledge
or
action.
The
ac-
tual
physical
communication
cannot,
however,
be
performed
at
the
same
abstraction
level
as
the
intention,
but
must
be
done
by
physical
means
such
as
sound
waves
or
photon
streams.
There
must
be
a
translation
process
that
converts
the
intention
rst
into
a
high-level
(virtual)
message,
and
thence
to
a
physical
representation;
the
physical
representation
is
interpreted
by
the
receiving
partner,
but
not
as
an
inverse
translation.
It
is
just
as
hard
for
the
originator
of
a
message
to
determine
whether
it
has
had
the
desired
effect
as
it
is
to
cause
the
effect
in
the
first
place.
Neither
can
be
done
by
telepathy.
All
the
originator
can
use
as
evidence
for
the
effect
of
a
message
is
the
recipient’s
actions.
If
the
recipient
is
cooperative,
those
actions
will
be
feedback
messages,
possibly
but
not
necessarily
verbal,
that
have
the
intent
of
getting
the
originator
to
perceive
how
the
recipient
is
interpreting
the
primal
message.
When
the
originator
perceives,
through
the
feedback,
that
the
recipient’s
state
has
come
close
enough
to
the
desired
state
that
gave
rise
to
the
original
message,
the
message
is
said
to
have
been
transmitted.
This
process
of
acting
so
that
one’s
perceptions
of
an
external
state
come
closer
to
a
desired
value
than
they
originally
were
is
at
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
225
the
core
of
Perceptual
Control
Theory
(PCT;
Powers
1973).
LP
Theory
can
be
seen
as
a
branch
of
PCT
speci c
to
the
interaction
between
two
people
or
between
a
person
and
a
computer.
The
conflicting
demands
of
stable
but
error-free
communication
can
be
met,
to
some
degree,
by
layering
the
communication
translation
processes,
using
protocols
that
are
more
or
less
independent
in
their
operation.
Each
protocol
performs
some
part
of
the
translation
process,
leaving
supporting
and
supported
protocols
to
complete
the
work.
The
structure
that
does
the
work
is
called
a
‘protocol
node’
which
corresponds
to
the
Elementary
Control
Unit
of
PCT.
The
set
of
protocol
nodes
that
support
one
another
to
complete
the
communication
of
a
single
high—level
message
is
called
a
‘Thread.’
Each
protocol
node
conducts
its
own
independent
dialogue
be-
tween
the
partners.
One
can
say
that
‘Within
one
message
may
be
many
dialogues.’
Within
a
protocol
node,
there
must
be
processes
that
we
label
Coders
and
Decoders,
which
create
or
interpret
‘virtual
messages’
in
ways
that
change
depending
on
the
differences
between
the
goal
state
and
the
per-
ceived
current
state
of
the
partner.
In
support
of
these
processes,
the
pro-
tocol
may
require
access
to
information
about
the
status
of
the
ongoing
dialogue,
about
the
partner,
about
the
world
within
which
the
communi-
cation
occurs,
and
so
forth.
Such
information
is
held
in
an
element
of
the
protocol
node
called
the
Model.
The
state
of
the
partner
in
the
Model
is
continuously
compared
with
the
desired
state,
and
the
difference
is
used
by
the
Coder
to
generate
virtual
messages
that
are
intended
to
reduce
the
dis-
crepancy
between
desired
and
actual
states
of
the
partner.
In
the
case
of
the
originator,
the
desired
state
is
the
‘primal
message’
to
be
communicated.
In
the
case
of
the
recipient
in
a
cooperative
dialogue,
it
is
to
see
the
originator
satis ed
with
the
recipient’s
interpretation
of
the
primal
message.
We
nd
it
convenient
to
separate
the
information
and
processes
within
a
protocol
node
into
three
chunks,
de ned
by
the
duration
of
validity
for
the
information
or
process.
Some
aspects
are
valid
over
the
long
term,
such
as
default
knowledge
of
the
world
and
of
how
to
execute
the
particular
proto-
col.
An
intermediate
chunk
relates
to
the
events
and
process
modi cations
appropriate
to
this
particular
dialogue,
and
these
are
valid
for
the
life
of
a
Thread. The
most
rapidly
changing
information
relates
to
the
passage
of
a
single
message
by
the
protocol
node.
It
includes
the
changing
belief
structures
about
the
current
state
of
message
transmission,
which
are
en-
capsulated
under
the
name
of
the
General
Protocol
Grammar
(GPG).
The
GPG
describes
the
intricate
structure
of
the
relations
among
the
back-and
forth
messages
in
the
feedback
loop.
228
Om
:
36s
0m:58s
1m:15s
1m
:4s
1m:16s
0m:54s
2m:33s
1m:13s
0m:37s
0m:15s
0m:52s
Om:34s
0m:30s
0m:44s
0m:59s
0m:51s
1m:51s
1m:8s
0m:41s
Omzlls
MARTIN
TAYLOR
8
DAVID
WA
UGH
C11:
W12:
C12:
W13:
C13:
W14:
C14:
W15:
C152
C153:
W16:
C16:
W17:
C17:
W18:
C18:
W19:
C19:
W20:
C20:
Evening
opening
times
Sorry,
but
I
have
no
information
about
opening
times.
You
must
call
the
restaurants
that
interest
you.
Capacity
how
many
people
Sorry,
but
I
have
only
the
locations
of
the
restaurants,
and
no
other
information.
For
further
information,
I
advise
you
to
call
directly.
What
would
you
recommend
for
a.
light
meal
I
cannot
answer.
I
do
not
have
enough
information
to
answer.
My
friend
is
someone
who
likes
to
have
a
good
wine
served
with
a
delicious
meal.
What
type
of
restaurant
would
you
advise
us
to
go
to?
If
you
want
to
know
about
restaurants,
I
advise
you
to
consult
the
Gault
and
Millau
Guide.
I
am
only
a
Yellow
Pages
information
service.
Look
for
a
vegetarian
restaurant
And
more
quickly
than
that
Sorry,
but
I
don’t
know
the
details
of
the
menus
And
however
they
I
don’t
know
Look
for
a
restaurant
whose
proprietor
is
Mme
LaBorde
I
don’t
know
the
names
of
restaurant
proprietors,
I
am
sorry
Look
for
restaurants
with
Danube
in
the
name
I
have
three
restaurants
called
“The
Danube”
LE
DANUBE
(1)
42
03
16
13
7
R
CARDUCCI
19E
LE
DANUBE
(1)
42
08
44
09
53
R
FBG
DU
TEMPLE
10E
LE
DANUBE
(1)
48
78
39
59
52
R.
CONDORCET
9E
Look
for
French
restaurants
at
less
than
100
Francs
per
person
Sorry,
but
I
don’t
know
the
prices
at
the
restaurants
Big
worthless
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
229
0m:24s
W21:
I
don’t
understand.
Could
you
reformulate
your
query
0m:38s
C21:
Dirty
big
0m:15s
C213:
My
dirty
big
1m:14s
C21b:
Look
for
panoramic
restaurants
with
handicapped
access
1m:7s
W22:
Wait
a
moment.
I
am
looking
for
panoramic
restaurants
3m:25s
C22:
This
will
be
the
last
request,
so
be
quick
4m:8s
C22a:
We
will
stop
now
ieferences
Beun,
(1991).
A
framework
for
cooperative
dialogues.
Presented
at
the
Second
Venaco
Workshop
on
the
Structure
of
Multimodal
dialogue,
Maratea,
Italy.
Black,
W.,
J.
Allwood,
H.
Bunt,
F.
Dols,
C.
Donzella,
G.
Ferrari,
J.
Gallagher,
R.
Haidan,
B.
Imlah,
K.
Jokinen,
Lancel,
J.
Nivre,
G.
Sabah,
and
Wachtel
(1991).
A
pragmatics-based
language
understanding
system.
In
Information
processing
systems
and
software:
results
and
progress
of
selected
projects,
Brussels.
ESPRIT.
A
g
Bunt,
H.
(1989).
Information
dialogues
as
communicative
action
in
relation
to
partner
modeling
and
information
processing.
In
M.M.
Taylor,
F.
Néel
and
D.
Bouwhuis
(Eds),
The
structure
of
multimodal
dialogue.
Amsterdam:
North—Holland
Elsevier.
Buxton,
W.
(1983).
Lexical
and
pragmatic
considerations
of
input
structures.
Computer
Graphics
17,
31-37.
Craik,
F.
and
R.
Lockhart
(1972).
Levels
of
processing:
a
framework
for
memory
research.
Journal
of
Verbal
Learning
and
Verbal
Behavior
11,
671-684.
Donders,
F.
(1862).
Die
Schnelligkeit
physischer
Processe.
Archiv
Anatomische
Physiologic,
657-681.
Edwards,
J.
and
Mason
(1989).
The
structure
of
intelligence
in
dialogue.
In
M.
Taylor,
F.
Néel,
and
D.
Bouwhuis
(Eds),
The
structure
of
multimodal
dialogue.
Amsterdam:
Elsevier
Science
Publishers.
Engel,
F.
and
R.
Haakma
(1993).
Expectations
and
feedback
in
user—system
communication.
International
Journal
of
Human-Computer
Studies
39,
427-452.
Farrell,
P.,
H.
Gamble,
J.
Hollands,
and
M.
Taylor
(1999).
Perceptual
control
and
layered
protocols
in
interface
design,
I;
International
Journal
of
Human-
Computer
Studies
50,
489-520.
Foley,
J.
and
A.
van
Dam
(1982).
Fundamentals
of
interactive
computer
graph-
ics.
Reading,
MA:
Addison-Wesley.
230
MARTIN
TAYLOR
6’
DAVID
WA
UGH
Grice,
H.
(1975).
Logic
and
conversation.
In
P.
Cole
and
J.
Morgan
(Eds.),
Syntax
and
semantics
3:
speech
acts,
pp.
41-58.
New
York:
Academic
Press.
Guyomard,
M.
and
J.
Siroux
(1989).
Suggestive
and
corrective
answers:
a
single
mechanism.
In
M.
Taylor,
F.
Néel,
and
D.
Bouwhuis
(Eds.),
The
structure
of
multimodal
dialogue.
Amsterdam:
North-Holland.
Ignatieff,
G.
(1985).
The
making
of
a
peacemonger.
Toronto:
University
of
Toronto
Press.
Levelt,
W.
(1989).
Speaking:
from
intention
to
articulation.
Cambridge,
MA:
MIT
Press.
Moran,
T.
(1981).
The
command
language
grammar:
a
representation
for
the
user
interface
of
interactive
computer
systems.
International
Journal
of
Man-Machine
Studies
21,
365-375.
Morel,
M.
(1989).
Computer-human
communication.
In
M.
Taylor,
F.
Néel,
and
D.
Bouwhuis
(Eds.),
The
structure
of
multimodal
dialogue.
Amsterdam:
Elsevier
Science
Publishers.
Nielsen,
J.
(1986).
A
virtual
protocol
model
for
computer-human
interaction.
International
Journal
of
Man-Machine
Studies
24,
301-312.
Norman,
D.
(1981).
Categorization
of
action
slips.
Psychological
Review
88,
1~15.
Norman,
D.
(1984).
Stages
and
levels
in
human-machine
interaction.
Interna-
tional
Journal
of
Man-Machine
Studies
21,
365-375.
Powers,
W.
(1973).
Behavior:
the
control
of
perception.
Chicago:
Aldine.
Tanenbaum,
A.
(1981).
Network
protocols.
Computing
Surveys
13,
453489.
Taylor,
I.
and
M.
Taylor
(1983).
The
psychology
of
reading.
New
York:
Aca-
demic
Press.
Taylor,
M.
(1988a).
Layered
protocols
for
computer-human
dialogue.
I:
Princi-
ples.
International
Journal
of
Man-Machine
Studies
28,
175-218.
Taylor,
M.
(1988b).
Layered
protocols
for
computer-human
dialogue.
II:
Some
practical
issues.
International
Journal
of
Man-Machine
Studies
28,
218-257.
Taylor,
M.
(1989).
Response
timing
in
layered
protocols:
a
cybernetic
view
of
natural
dialogue.
In
M.
Taylor,
F.
Néel,
and
D.
Bouwhuis
(Eds.),
The
structure
of
multimodal
dialogue.
Amsterdam:
Elsevier
Science
Publishers.
Taylor,
M.,
C.
McCann,
and
M.
Tuori
(1984).
The
interactive
spatial
informa-
tion
system.
DCIEM
Technical
Report
84-R-22.
Taylor,
M.
and
D.
Waugh
(1992).
Principles
for
integrating
voice
1/0
in
a
complex
interface.
In
Advanced
Aircraft
Interfaces:
the
Machine
Side
of
the
Man-Machine
Interface.
Neuilly
sur
Seine:
AGARD:
AGARD
Avionics
Panel
Symposium
Publication
AGARD-CP-521.
Taylor,
M.,
P.
Farrell,
H.
Gamble,
and
J.
Hollands
(1999).
Perceptual
con-
trol
and
layered
protocols
in
interface
design,
II.
International
Journal
of
Human-Computer
Studies
50,
521-555.
fr
_,
LA
YERED
PROTOCOL
ANALYSIS
OF
DIALOGUES
231
Taylor,
M.
and
D.
Waugh
(2000).
Multiplexing,
diviplexing
and
the
control
of
multimodal
dialogue.
In
M.
Taylor,
F.
Néel,
and
D.
Bouwhuis
(Eds.),
The
structure
of
multimodal
dialogue
II.
Amsterdam:
John
Benjamins.
Pre-
sented
at
the
Second
Venaco
Workshop
on
the
Structure
of
Multimodal
Dialogue,
1991.
Zimmermann,
H.
(1980).
OSI
reference
model
—
the
ISO
model
of
architec-
ture
for
open
systems
interconnection.
IEEE
Transactions
on
Communica-
tion
COM-28,
425-432.
