Dialogue Systems: Simulations or Interfaces?

About This Presentation

Title:

Dialogue Systems: Simulations or Interfaces?

Description:

Dialogue Systems: Simulations or Interfaces? Staffan Larsson G teborg University Sweden Introduction Basic question What is the goal of formal dialogue research? – PowerPoint PPT presentation

Number of Views:226

Avg rating:3.0/5.0

Slides: 101

Provided by: Staf59

Category:

more less

Transcript and Presenter's Notes

Title: Dialogue Systems: Simulations or Interfaces?

1
Dialogue Systems Simulations or Interfaces?

Staffan Larsson
Göteborg University
Sweden

2
Introduction
3
Basic question

What is the goal of formal dialogue research?
Formal dialogue research
formal research on the semantics and pragmatics
of dialogue

4
Two possible answers

Engineering view the purpose of formal dialogue
research is
interface engineering (services and technologies)
enable building better human-computer interfaces
Simulation view the ultimate goal of formal
dialogue research is
a complete formal and computational
(implementable) theory of human language use and
understanding

5
The convergence assumption

There is an extensive if not complete overlap
between the simulation of human language use and
the engineering of conversational interfaces.

6
Aim of this presentation

Review an argument against the possibility of
human-level natural language understanding in
computers (simulation view)
Explicitly apply this argument to formal dialogue
research, arguing that the covergence assumption
is dubious
Draw out the consequences of this for formal
dialogue research

7
Formal dialogue research and GOFAI
8
The Turing test

Can a machine think? Turing offers an operational
definition of the ability to think
Turings imitation game
Test person A has a dialogue (via a text
terminal) with B.
As goal is to decide whether B is a human or a
machine
If B is a machine and manages to deceive A that B
is a human, B should be regarded as able to think

9
The Turing test and the Simulation view

The Turing Test can be seen as the ultimate test
of a simulation of human language use
The ability to think is operationalised as the
ability to carry out a natural language dialogue
in a way that is indiscernible from that of a
human
The goal of formal dialogue research coincides
with the goal of AI (as originally perceived)

10
GOFAI

Artificial Intelligence
Goal simulate human/intelligent
behaviour/thinking
Weak AIMachines can be made to act as if they
were intelligent
Until the mid-80s, the dominating paradigm of AI
was the idea that thinking is, essentially,
symbol manipulation
The physical symbol hypothesis
All intelligent behaviour can be captured by a
system that reasons logically from a set of
facts and rules that describe the domain
This is sometimes referred to as GOFAI
(Good Old Fashioned AI)

11
Dialogue systems and GOFAI

Since around the mid-80s, GOFAI has been
abandoned by many (but not all) AI researchers
Instead, focus on NEFAI (New-Fangled AI)
connectionism,
embodied interactive automata,
reinforcement learning,
probabilistic methods, etc.
However, a large part of current dialogue systems
research is based on the GOFAI paradigm
Information States, for example
Formal pragmatics is often used as a basis for
the implementation of dialogue managers in
GOFAI-style approaches

12
Formal semantics and GOFAI

GOFAI and formal semantics deals, to a large
extent, with similar problems and use similar
methods
Formal symbolic representations of meaning
Natural Language Understanding as symbol
manipulation
(Even though many early GOFAI researchers appear
oblivious to the existence of formal semantics of
natural language in the style of Montague, Kamp
etc.)
Formal semantics perhaps not originally intended
to be implemented, and not as part of AI
Still, formal semantics shares with GOFAI rests
on the assumption that natural language meaning
can be captured in formal symbol manipulation
systems

13
Why GOFAI?

Why GOFAI in formal semantics and pragmatics?
It seems to be the most workable method for the
complex problems of natural language dialogue
Natural language dialogue appears to be useful
for improving on current human-computer
interfaces
But is GOFAI-based research also a step on the
way towards human-level natural language
understanding in computers, i.e. simulation?

14
Phenomenological arguments against GOFAI
15
Some problems in AI

Frame problem
updating the world model
knowing which aspects of the world are relevant
for a certain action
Computational complexity in real-time
resource-bounded applications
Planning for conjunctive goals
Plan recognition
Incompleteness of general FOL reasoning
not to mention modal logic
Endowing a computer with the common sense of a
4-year-old
AI is still very far from this

Humans dont have problems with these things
Is it possible that all these problems have a
common cause?
They all seem to be related to formal
representations and symbol manipulation

17
Background and language understanding

Dreyfus, Winograd, Weizenbaum
Human behaviour based on our everyday commonsense
background understanding
allows us to experience what is currently
relevant, and deal with tings and people
crucial to understanding language
involves utterance situation, activity,
institution, cultural setting, ...

Dreyfus argues that the background has the form
of dispositions, or informal know-how
Normally, one simply knows what to do
a form of skill rather than propositional
knowing-that
To achieve GOFAI,
this know-how, along with interests, feelings,
motivations, social interests, and bodily
capacities that go to make a human being,...
... would have to be conveyed to the computer as
knowledge in the form of a huge and complex
belief system

19
CYC (Lenat) and natural language

An attempt to formalise common sense
The kind of knowledge we need to understand NL
using general categories that make no reference
to specific uses of the knowledge
Lenats ambitions
its premature to try to give a computer skills
and feelings required for actually coping with
things and people
L. is satisfied if CYC can understand books and
articles and answer questions about them

20
The background cannot be formalised

There are no reasons to think that humans
represent and manipulate the background
explicitly, or that this is possible even in
principle
...understanding requires giving the computer a
background of commons sense that adult humans
have in virtue of having bodies, interacting
skilfully with the material world, and being
trained into a culture
Why does it appear plausible that the background
could be formalised knowing-that?
Breakdowns
Skill acquisition

21
Skills and formal rules

When things go wrong - when we fail there is a
breakdown
In such situations, we need to reflect and
reason, and may have to learn and apply formal
rules
but it is a mistake to
read these rules back into the normal situation
and
appeal to such rules for a causal explanation of
skilful behaviour

22
Dreyfus account of skill acquisition

1. Beginner student Rule-based processing
learning and applying rules for manipulating
context-free elements
There is thus a grain of truth in GOFAI
2. Understanding the domain seeing meaningful
aspects, rather than context-free features
3. Setting goals and looking at the current
situation in terms of what is relevant
4. Seeing a situation as having a certain
significance toward a certain outcome
5. Expert The ability of instantaneously
selecting correct responses (dispositions)

There is no reason to suppose that the beginners
features and rules (or any features and rules)
play any role in expert performance
That we once followed a rule in tying our
shoelaces does not mean we are still following
the same rule unconsciously
Since we needed training wheels when learning
how to ride a bike, we must now be using
invisible training wheels.
Human language use and cognition involves symbol
manipulation, but is not based on it

24
Recap

Language understanding requires access to human
background understanding
This background cannot be formalised
Since GOFAI works with formal representations,
GOFAI systems will never be able to understand
language as humans do

25
Simulation and NEFAI
26
What about NEFAI?

This argument only applies to GOFAI!
A lot of modern AI is not GOFAI
New-Fangled AI (NEFAI)
interactionist AI (Brooks, Chapman, Agre)
embodied AI (COG)
connectionism / neural networks
reinforcement learning
So maybe human language use and understanding
could be simulated if we give up GOFAI and take
up NEFAI?
Note that very few have tried this in the area of
dialogue
Simply augmenting a GOFAI system with statistics
is not enough

27
Progress?

Although NEFAI is more promising than GOFAI...
... most current learning techniques rely on the
previous availability of explicitly represented
knowledge the training data must be interpreted
and arranged by humans
in the case of learning the background, this
means that the background has to be represented
before it can be used for training
But as we have seen, Dreyfus argues that
commonsense background cannot be captured in
explicit representations

Russel Norvig, in Artificial Intelligence -A
Modern Approach (1999)
In a discussion of Dreyfus argument
In our view, this is a good reason for a serious
redesign of current models of neural processing
.... There has been some progress in this
direction.
But no such research is cited
So R N admit that this is a real problem. In
fact it is still the exact same problem that
Dreyfus pointed out originally
There is still nothing to indicate that Dreyfus
is wrong when arguing against the possibility of
getting computers to learn commonsense background
knowledge

But lets assume for the moment that the current
shortcomings of NEFAI could be overcome...
that learning mechanisms can be implemented who
learn in the same way humans do
and that appropriate initial structure of these
systems can be given
and that all this can be done without providing
predigested facts that rely on human
interpretation

30
Some factors influencing human language use

Embodiment
having a human body, being born and raised by
humans
Being trained into a culture
by interacting with other humans
Social responsibility
entering into social commitments with other people

31
What is needed to achieve simulation?

So, perhaps we can do real AI, provided we can
build robot infants that are raised by parents
and socialised into society by human beings who
treat them as equals
This probably requires people to actually think
that these AI systems are human
These systems will have the same ethical status
as humans
If we manage to do it, is there any reason to
assume that they would be more useful to us than
ordinary (biological) humans?
They are no more likely to take our orders...

It appears that the research methods required for
simulation are rather different from those
required for interface design
The convergence assumption appears very dubious

33
Formal dialogue research and dialogue systems
design
34
Consequences of the argument for the engineering
view

If we accept the argument that the background is
not formalisable and that computers (at least as
we know them) cannot simulate human language
understanding...
...what follows with respect to the relations
between
Formal semantics and pragmatics of dialogue
Non-formal theories of human language use
Dialogue systems design as interface engineering
Both (1) and (2) are still relevant to (3)

35
Winograd on language and computers

Even though computers cannot understand language
in the way humans can...
...computers are nevertheless useful tools in
areas of human activity where formal
representation and manipulation is crucial
e.g. word processing.
In addition, many practical AI-style applications
do not require human-level understanding of
language
e.g. programming a VCR, getting timetable
information
In such cases, it is possible to develop useful
systems that have a limited repertoire of
linguistic interaction.
This involves the creation of a systematic domain

36
Systematic domains

A systematic domain is a set of formal
representations that can be used in a computer
system
Embodies the researchers interpretation of the
situation in which the system will function.
Created on the basis of regularities in
conversational behaviour (domains of
recurrence)

37
so...

For certain regular and orderly activities and
language phenomena...
... it is possible to create formal
representations which capture them well enough to
build useful tools
Formal dialogue research can be regarded as the
creation of systematic domains in pragmatics and
semantics of dialogue

38
Formal semantics and pragmatics of dialogue as
systematic domains

Formal theories of language use should be
regarded as
the result of a creative process of constructing
formal representations (systematic domains)
based on observed regularities in language use
These theories can be used in dialogue systems to
enable new forms of human-machine interaction

39
Formal pragmatics

Pragmatic domains include e.g.
turntaking, feedback and grounding, referent
resolution, topic management
Winograd gives dialogue game structure as a prime
example of a systematic domain
Analysed along the lines of dialogue games
encoded in finite automata
ISU update approach is a variation of this,
intended to capture the same regularities in a
(possibly) more flexible way
It is likely that useful formal descriptions can
be created for many aspects of dialogue structure

40
Formal semantics

Not a focus of Winograds formal analysis,
presumably because Winograd believes that
language understanding is not amenable to formal
analysis
However, even if one accepts the arguments such
as those above...
... it seems plausible that the idea of
systematic domains also applies to semantics
That is, for certain semantically regular task
domains it is indeed possible to create a formal
semantics
e.g. in the form of a formal ontology and formal
representations of utterance contents
This formal semantics will embody the
researchers interpretation of the domain

41
Relevant issues related to semantic domains

How to determine whether (and to what extent) a
task domain is amenable to formal semantic
description
How to decide, for a given task domain, what
level of sophistication is required by a formal
semantic framework in order for it to be useful
in that domain
In some domains, simple feature-value frames may
be sufficient while others may require something
along the lines of situation semantics, providing
treatments of intensional contexts etc.
Fine-grainedness and expressivity of the formal
semantic representation required for a domain or
group of domains
e.g. database search, device programming,
collaborative planning, ...
Creation of application-specific ontologies
How to extract applications ontologies from
available data of the domain, e.g. transcripts of
dialogues.

42
but...

Even though some aspects of language use may
indeed be susceptible to formal description
This does not mean that human language use
actually relies on such formal descriptions
represented in the brain or elsewhere
So implementations based on such formalisations
are not simulations of human language use and
cognition

43
Limits of formalisation

Formalisation will only be useful in areas of
language use which are sufficiently regular to
allow the creation of systematic domains
So, repeated failures to formally capture some
aspect of human language may be due to the limits
of formal theory when it comes to human language
use, rather than to some aspect of the theory
that just needs a little more tweaking.

44
Non-formalisable language phenomena

For other activities and phenomena, it may not
possible to come up with formal descriptions that
can be implemented
e.g. human language understanding in general,
since it requires a background which cannot be
formalised
also perhaps aspects of implicit communication,
conversational style, politeness in general,
creative analogy, creative metaphor, some
implicatures
This does not mean that they are inaccessible to
science.
They can be described non-formally and understood
by other humans
Their general abstract features may be
formalisable

45
Usefulness of non-formal theory

Non-formal theories of human language use are
still useful for dialogue systems design
Dialogue systems will need to be designed on the
basis of theories of human language
They will, after all, interact with a human
May also be useful to have human-like systems
(cf. Cassell)
This does not require that implementations of
these theories have to be (even partial)
simulations of human language use and cognition
Also, observations of human-human dialogue can of
course be a source of inspiration for dialogue
systems design

46
Conclusions
47

In important ways the simulation view and the
engineering view are different projects requiring
different research methods
For the simulation project, the usefulness of
systems based on formal representations is
questionable
Instead, formal dialogue research can be regarded
as the creation of systematic domains that can be
used in the engineering of flexible
human-computer interfaces
In addition, non-formal theory of human language
use can be useful in dialogue systems design

If interface engineering is liberated from
concerns related to simulation...
...it can instead be focused on the creation of
new forms of human-computer (and
computer-mediated) communication...
... adapting to and exploring the respective
limitations and strengths of humans and
computers.

49
fin
50
Other views of what FDR is
51
A variant of the simulation view

The goal of formal dialogue research is a
complete computational theory of language and
cognition for machines
cf. Luc Steels
Robots evolving communication
Not intended to describe human language use
although some aspects may be similar
Arguably interesting in its own right

One may even be able to implement computational
models that capture some abstract aspects of
human language use and understanding
That are not based on symbol manipulation, but
involve subsymbolic computation
For example, the evolution of shared language use
in robots (Steels et al)
However, such formal models and simulations will
never be complete simulations of human language
understanding (to the extend required by the
Turing test)
unless the machines they run on are human in
all aspects relevant to language, i.e. physical,
biological, psychological, and social

53
A variant of the simulation view

The goal of formal dialogue research is a
complete computational theory of language and
cognition in general
either such a theory subsumes a theory of human
language
and thus as difficult or more difficult
or not
and thus coherent with idea that only some
aspects of language are formalisable,
although it remains to show that the same
features are the ones that are essential for
language

54
Applied science?

Formal dialogue research as applied science
c.f. medicine
theories of interface design
theories of (linguistic) human-computer
interaction - LHCI

55
The role of human-human communication in LHCI

What aspects of natural dialogue are
formalisable
implementable
useful in HCI

56
Scientific status of formal descriptions

Formal descriptions may have some scientific
value as theories of human language use and
cognition
However, they are
not useful as a basis for simulation of human
language use and congition, since this is not
based on explicit rules and representations
(except for novices and breakdowns)
often radical simplifications (as many other
scientific theories)
limited in scope and describe special cases only
Even if the creation of a systematic domain is
possible for some linguistic phenomena, this does
not mean that human language use is based on
formal representations

57
Formal dialogue research vs. Dialogue systems
research

Both share the assumption that human language use
and meaning can be captured in formal symbol
manipulation systems
Human language use and meaning relies on
background
Background cannot be formalised

58
Language use vs. cognition

Turing test tests only behaviour cognition is a
black box
So whats the justification for talking about
cognition?
Turings test intended as an operational
definition of thinking, i.e. cognition
Possible underlying intuition
There is no way of passing the Turing test for a
system with a style of cognition which is very
different from human cognition
Turing assumed that human cognition was based on
symbol manipulation

59
Domain-specific simulation?

In a regular domain, can a program based on a
formalisation of these regularities be regarded
as a simulation of human performance in that
domain?
Even if there are regularities that can be
captured to a useful extent in rules, this does
not mean that humans use such rules
unless they are complete novices who have been
taught the rules explicitly but have not yet had
time to descend down the learning hierarchy

60
General vs. domain-specific intelligence

Weizenbaum there is no such thing as general
intelligence
intelligence is always relative to a domain
(math, music, playing cards, cooking, ...)
Therefore, the question whether computers can be
intelligent is meaningless
one must ask this question in individual domains

61
The Feigenbaum test

Replace the general Turing test with a similar
test in limited domains? (proposed by Feigenbaum)
Certainly seems more manageable, especially in
systematic domains
On the other hand, it could be argued that it is
exactly in the non-systematic domains that the
most interesting and unique aspects of human
being are to be found
So this test is very different from the original
Turing test

62
More on skills vs. rules
63
Everyday skills vs. rules

Dreyfus suggests testing the assumption that the
background can be formalised
by looking at the phenomenology of everyday
know-how
Heidegger, Merleau-Ponty, Pierre Bourdieu
What counts as facts depends on our skills e.g.
gift-giving (Bourdieu)
If it is not to constitute an insult, the
counter-gift must be deferred and different,
because the immediate return of an exact
identical object clearly amounts to a refusal....
It is all a question of style, which means in
this case timing and choice of occasion...
...the same act giving, giving in return,
offering ones services, etc. can have
completely different meanings at different times.

64
Everyday skills vs. rules

Having acquired the necessary social skill,
one does not need to recognize the situation as
appropriate for gift-giving, and decide
rationally what gift to give
one simply responds in the appropriate
circumstances by giving an appropriate gift
Humans can
skilfully cope with changing events and
motivations
project understanding onto new situations
understand social innovations
one can do something that has not so far counted
as appropriate...
...and have it recognized in retrospect as having
been just the right thing to do

65
The B.A.B. objection - background
66
The argument from infant development (Weizenbaum)

(Based on writings by child psychologist Erik
Erikson)
The essence of human being depends crucially on
the fact that humans are born of a mother, are
raised by a mother and father, and have a human
body
Every organism is socialized by dealing with
problems that confront it (Weizenbaum)
For humans, the problems include breaking the
symbiosis with the mother after the infant period
This is fundamental to the human constitution it
lays the ground for all future dealings with
other people
Men and machines have radically different
constitutions and origins
Humans are born by a mother and father
Machines are built by humans
OK, so we need to give AI systems a human or
human-like body, and let human parents raise them

67
The argument from language as social commitment
(Winograd)

The essence of human communication is commitment,
an essentially social and moral attitude
Speech acts work by imposing commitments on
speaker and hearer
If one cannot be held (morally) responsible for
ones actions, one cannot enter into commitments
Computers are not human
so they cannot be held morally responsible
therefore, they cannot enter into commitments
Therefore, machines can never be made to truly
and fully understand language
OK, so we need to treat these AI computers
exactly as humans, and hold them morally
responsible

68
The argument from human being/Dasein
(Heidegger, Dreyfus)

Heideggers project in Being and Time
Develop an ontology for describing human being
What its like to be human
This can, according to Heidegger, only be
understood from the inside
Hs text is not intended to be understandable by
anyone who is not a human
Such an explanation is not possible, according to
H. human being cannot be understood from
scratch
Yet it is exactly such an explanation that is the
goal of AI
According to Heidegger/Dreyfus, AI is impossible
because (among other things)
Infants are, strictly speaking, not yet fully
human they must first be socialised into a
society and a social world
Only humans so socialized can fully understand
other humans
Since cultures are different, humans socialized
into one culture may have problems understanding
humans from another culture
Machines are not socialised, they are programmed
by humans
OK, so we need to socialise AI systems into
society!

69
Arguments related to evolution
70
The humans are animals argument

What reason do we have to think that
non-conscious reasoning operates by formal
reasoning?
Humans have evolved from animals, so presumably
some non-formal thinking is still part of the
human mind
Hard to tell a priori how much

71
The argument from the role of emotions

Classical AI deals first with rationality
Possibly, we might want to add emotions as an
additional layer of complexity
However, it seems plausible to assume that
emotions are more basic than rationality
(Damasio The Feeling of what happens)
Animals have emotions but not abstract rational
reasoning
The human infant is emotional but not rational
So machines should be emotional before they are
made rational
unfortunately, no-one has a clue how to make
machines emotional

72
The argument from brain matter and evolution

Weak AI assumes that physical-level simulation is
unnecessary for intelligence
However, evolution has a reputation for finding
and exploiting available shortcuts
works by patching on previous mechanisms
If there are any unique properties of biological
brain-matter that offers some possible
improvement to cognition, it is likely they have
been exploited
If so, it is not clear if these properties can be
emulated by silicon-based computers

73
The argument from giving a damn

Humans care machines dont give a damn
(Haugeland)
Caring (about surviving, for example) comes from
instincts (drives) which animals, but not
machines, have
Caring about things is intimately related to the
evolution of living organisms
Having a biological body
So, can evolution be simulated?
Winograd argues that the only simulation that
would do the job would need to be as complex as
real evolution
So in 3,5 billion years, we can have AI!

74
More on CYC
75
Problems with formalising commonsense background

How is everyday knowledge organized so that one
can make inferences from it?
Ontological engineering finding the primitive
elements in which the ontology bottoms out
How can skills or know-how be represented as
knowing-that?
How can relevant knowledge be brought to bear in
particular situations?

76
CYC (Lenat) and natural language

Formalise common sense
The kind of knowledge we need to understand NL
using general categories that make no reference
to specific uses of the knowledge (context free)
Lenats ambitions
its premature to try to give a computer skills
and feelings required for actually coping with
things and people
L. is satisfied if CYC can understand books and
articles and answer questions about them

77
CYC vs. NL

Example (Lenat)
Mary saw a dog in the window. She wanted it.
Dreyfus
this sentence seems to appeal to
our ability to imagine how we would feel in the
situation
know-how for getting around in the world (e.g.
getting closer to something on the other side of
a barrier)
rather than requiring us to consult facts about
dogs and windows and normal human reactions
So feelings and coping skills that were excluded
to simplify the problem return
We shouldnt be surprised this is the
presupposition behind the Turing Test that
understanding human language cannot be isolated
from other human capabilities

78
CYC vs. NL

How can relevant knowledge be brought to bear in
particular situations?
categorize the situation
search through all facts, following rules to find
the facts possibly relevant in this situation
deduce which facts are actually relevant
How deal with complexity?
Lenat add meta-knowledge
Dreyfus
meta-knowledge just makes things worse more
meaningless facts
CYC is based on an untested traditional
assumption that people store context-free facts
and use meta-rules to cut down the search space

79
Analogy and metaphor

... pervade language (example from Lenat)
Texaco lost a major ruling in its legal battle
with Pennzoil. The supreme court dismantled
Texacos protection against having to post a
crippling 12 billion appeals bond, pushing
Texaco to the brink of a Chapter 11 filing (Wall
Street Journal)
The example drives home the point that,
far from overinflating the need for real-world
knowledge in language understanding,
the usual arguments about disambiguation barely
scratch the surface

80
Analogy and metaphor

... pervade language (example from Lenat)
Texaco lost a major ruling in its legal battle
with Pennzoil. The supreme court dismantled
Texacos protection against having to post a
crippling 12 billion appeals bond, pushing
Texaco to the brink of a Chapter 11 filing (Wall
Street Journal)
The example drives home the point that,
far from overinflating the need for real-world
knowledge in language understanding,
the usual arguments about disambiguation barely
scratch the surface

81
Analogy and metaphor

Dealing with metaphors is a non-representational
mental capacity (Searle)
Sally is a block of ice could not be analyzed
by listing the features that Sally and ice have
in common
Metaphors function by association
We have to learn from vast experience how to
respond to thousands of typical cases
Mention approaches to metaphor, e.g. Abduction -
the boston office called - isnt this a
solution? Dead vs. Creative metaphors

82
Neural nets
83

Helge!
But people also interpret things differently (but
not wildly differently)
Not many researchers believe in tabula rasa
Evolutionary alogrithms but so far not combined
with learning
Nns can learn without prior strong
symbolisation of learning data, but pehaps not
very complex stuff like dialogue?
Data can be either discrete or continuous
How does this relate to predigestion? Is
selection of data (e.g. Dividing into frequency
ranges predig.?
Main obstacle now puny amounts of neurons,
little knowledge of interaction of evolved
initial structure learning

neural nets can learn some things without prior
conceptualisation (but some discretisation is
necessary, e.g. representation in the weak sense)
strong and weak sense of represenation
Other problems with connectionism
Current neural networks are much less complex
that brains
but maybe this will change
Even if we had a working neural network, we would
not understand how it works
the scientific goal of AI would thus still not
have been reached

85
Learning generalisation

Take in dialogue systems based solely on
statistics (superhal?)
Mention hybrid vs totally nonsymbolic systems
Learning depends on the ability to generalise
Good generalisation cannot be achieved without a
good deal of background knowledge
Example trees/hidden tanks
A network must share our commonsense
understanding ot the world if it is to share our
sense of appropriate generalisation

Some counter-counter-arguments
The more the computer knows, the longer it will
take to find the right information
The more a human knows, the easier it is to
retrieve relevant information

87
Non-symbolic approaches to AI and dialogue
88
Interactionist AI

No need for a representation of the world
instead, look to the world as we experience it
Behaviour can be purposive without the agent
having in mind a goal or purpose
In many situations, it is obvious what needs to
be done
Once youve done that, the next thing is likely
to be obvious too
Complex series of actions result, without the
need for complex decisions or planning
However, Interactionist AI does not address
problem of informal background familiarity
programmers have to predigest the domain and
decide what is relevant
systems lack ability to discriminate relevant
distinctions in the skill domain...
... and learn new distinctions from experience

89
Connectionism

Apparently does not require being given a theory
of a domain in order to behave intelligently
Finding a theory finding invariant features in
terms of which situations can be mapped onto
responses
Starting with random weights, will neural nets
trained on same date pick out the same
invariants?
No it appears the tabula rasa assumption
(random initial weights) is wrong
Little research on how (possibly evolved) initial
structure interact with learnin

90
Learning generalisation

Learning depends on the ability to generalise
Good generalisation cannot be achieved without a
good deal of background knowledge
Example trees/hidden tanks
A network must share our commonsense
understanding ot the world if it is to share our
sense of appropriate generalisation

91
Reinforcement learning

Idea learn from interacting with the world
Feed back reinforcement signal measuring the
immediate cost or benefit of an action
Enables unsupervised learning
(The target representation in humans is neural
networks)
Dreyfus To build human intelligence, need to
improve this method
assigning fairly accurate actions to novel
situations
reinforcement-learning device must exhibit
global sensitivity by encountering situations
under a perspective and actively seeking relevant
input

92
AI and Turing Test
93
Prehistory of AI

Plato / Socrates
All knowledge must be statable in explicit
definitions which anyone could apply (cf.
definition of algorithm)
Descartes again
Understanding and thinking is forming and using
symbolic representations the rational
manipulation of symbols by means of rules
Thinking as computation
Leibniz
universal calculus for representing reasoning
goal put an end to all conflicts, which are
caused by misunderstandings of inexact, informal
language
Kant
clear distinction between external world
(noumena) and inner life (phenomena,
representations)

94
Artificial Intelligence

Goal
simulate human/intelligent behaviour/thinking
Weak AI
Machines can be made to act as if they were
intelligent
Strong AI
Agents that act intelligently have real,
conscious minds
(It is possible to believe in strong AI but not
in weak AI)

95
Some arguments against weak AI (Turing 1950)

Ada Lovelaces objection
computers can only do what we tell them to
Argument from disability
claims (usually unsupported) of the form a
machine can never do X
The mathematical objection
based on Gödels incompleteness theorem
The argument from informality of behaviour
(Searles Chinese Room
argument concerns strong AI
purports to show that producing intelligent
behavoiur is not a sufficient condition for being
a mind)

96
The Turing test and dialogue

The Turing Test can be seen as the ultimate test
of a simulation of human language use
The ability to think is operationalised as the
ability to carry out a natural language dialogue
in a way that is indiscernible from that of a
human
The machine in question is assumed to be a Turing
machine, i.e. a general symbol manipulation
device, i.e. a computer

97
The Turing test and the Simulation view

According to the simulation view, the goal of
formal dialogue research is to reproduce, in a
machine, the human ability to use and understand
language
Thus, the Turing test can be regarded as a
potential method of evaluating theories of human
language use and understanding
The goal of formal dialogue research coincides
with the goal of AI (as originally perceived)

98
Misc slides
99

Non-formal theories of those aspects of language
use which resist formalisation can be used as a
basis for design of aspects of dialogue systems
that do not need to be modelled by the system
itself.
For example, it is likely that any speech
synthesizer voice has certain emotional or other
cognitive connotations
it might sound silly, angry, etc..
It is extremely difficult, if not impossible, to
design a completely neutral voice.
However, if we have some idea of how different
voices are perceived by humans, we can use this
(informal) knowledge to provide a dialogue system
application with an appropriate voice for that
application.

100
Dreyfus account of skill acquisition

5 stages
1. Beginner student Rule-based processing
learning and applying rules for manipulating
context-free elements
There is thus a grain of truth in GOFAI
2. Understanding the domain seeing meaningful
aspects, rather than context-free features
3. Setting goals and looking at the current
situation in terms of what is relevant
4. Seeing a situation as having a certain
significance toward a certain outcome
5. Expert The ability of instantaneously
selecting correct responses (dispositions)
(Note this is how adults typically learn
Infants, on the other hand...
learn by imitation
pick up on a style that pervades his/her
society)

101
(No Transcript)
102

The question is still open exactly how far it is
possible to go in the formal description of
phenomena related to language use
The only way to find out is to by trial-and-error
(i.e., research).
In this, I suggest one might be well-advised to
keep in mind the following points..

103

From observations of some domain or aspect of
language use...
... the researcher creates a formal
representation...
... and implements it in a dialogue system.
When this dialogue system is used and interacts
with humans,
new domain ?of interaction comes into being
depending on the design of the system, different
conversational patterns will arise
Domains of language use that may be susceptible
to formalisation (i.e. creation of systematic
domains) can be roughly divided into pragmatic
and semantic domains

104
Usefulness of formal semantics and pragmatics

Systems based on formal representations provide a
great potential for improving on human-computer
interaction

105
(An aside human reinforcement)
106
Overview

Introduction
Formal dialogue research and GOFAI
Phenomenological arguments against GOFAI
Simulation and NEFAI
Formal dialogue research and dialogue systems
design
Conclusions

107

Currently, programmer must supply machine with
rule formulating what to feed back as
reinforcement
What is the reinforcement signal for humans?
Survival?
Pleasure vs. pain?
Requires having needs, desires, emotions
Which in turn may depend on the abilities and
vulnerabilities of a biological body

108