=Paper=
{{Paper
|id=Vol-1119/paper8
|storemode=property
|title=Capturing the Implicit ‐ An Iterative Approach to Enculturing Artificial Agents
|pdfUrl=https://ceur-ws.org/Vol-1119/paper8.pdf
|volume=Vol-1119
}}
==Capturing the Implicit ‐ An Iterative Approach to Enculturing Artificial Agents==
https://ceur-ws.org/Vol-1119/paper8.pdf
Capturing the Implicit – an iterative approach to
enculturing artificial agents
Peter Wallis and Bruce Edmonds
Centre for Policy Modelling
Manchester Metropolitan University
Manchester, United Kingdom
pwallis@acm.org, bruce@edmonds.name
Abstract. Artificial agents of many kinds increasingly intrude into the human
sphere. SatNavs, help systems, automatic telephone answering systems, and even
robotic vacuum cleaners are positioned to do more than exist on the side-lines
as potential tools. These devices, intentionally or not, often act in a way that in-
trudes into our social life. Virtual assistants pop up offering help when an error is
encountered, the robot vacuum cleaner starts to clean while one is having tea with
the vicar, and automated call handling systems refuse to let you do what you want
until you have answered a list of questions. This paper addresses the problem of
how to produce artificial agents that are less socially inept. A distinction is drawn
between things which are operationally available to us as human conversational-
ists and the things that are available to a third party (e.g. a scientists or engineer)
in terms of an explicit explanation or representation. The former implies a de-
tailed skill at recognising and negotiating the subtle and context-dependent rules
of human social interaction, but this skill is largely unconscious – we do not know
how we do it, in the sense of the later kind of understanding. The paper proposes
a process that bootstraps an incomplete formal functional understanding of hu-
man social interaction via an iterative approach using interaction with a native.
Each cycle of this iteration entering and correcting a narrative summary of what
is happening in recordings of interactions with the automatic agent. This interac-
tion is managed and guided through an “annotators’ work bench” that uses the
current functional understanding to highlight when user input is not consistent
with the current understanding, suggesting alternatives and accepting new sug-
gestions via a structured dialogue. This relies on the fact that people are much
better at noticing when dialogue is ”wrong” and in making alternate suggestions
than theorising about social language use. This, we argue, would allow the itera-
tive process to build up understanding and hence CA scripts that fit better within
the human social world. Some preliminary work in this direction is described.
1 Introduction
This paper is focused upon computers that inhabit roles with human origin. In particular,
computers that have to converse with people as social actors, in the course of their
interactions with them. This is not the only sort of interface of course and some will
argue that computers as we know them have perfectly satisfactory interfaces, e.g. those
based on the notion that the computers are a tool facilitated by a physical analogue (e.g.
83
a desktop). However a “social stance” – considering computers as social actors – may
allow for a new range of applications to emerge as well as giving new insights into
human behaviour, in particular the current limitations of our models of this.
However, when computers are compelled to work as social actors – for example
when they use language as the primary modality – they tend to fail grossly rather than
in detail. Indeed people get so frustrated by computers that they often swear at them [1].
When someone swears at a carefully crafted chat-bot, the human is unlikely to have
been upset by punctuation or a quirky use of pronouns. The challenge is that existing
qualitative techniques are good at the detail, but can fail to find a bigger picture.
In this paper the focus is on performative language and builds on the findings of
applied linguistics where the mechanism of language can be seen as part of the same
spectrum of communicative acts ranging from “body language” to semiotics. However
much of these communicative acts are learned in context with reference to their effect
rather than a putative explicit meaning.
This is contrary to approach that characterises human actors as rational actors. Ap-
plied to language this motivates the characterisation that natural languages are a “fallen”
version of something more pure – a messy version of First Order Predicate Calculus –
where elements of the language can be associated with their separate meaning. This
meaning-text model [2] has been largely rejected since the late 1980’s but has a latent
existence in the idea that it is possible to create sets of Dialogue Acts (DAs) that capture
in some way the primitive concepts from which any conversation can be constructed.
For a comprehensive description of the theory and lack thereof in this area, there are
several papers by Eduard Hovy [3].
It is also going in a different direction to those focused on statistical and machine
learning techniques [4] that treat mental attitudes as a “hidden state” that can be derived
from corpora of human behaviour. The advantage of this approach to engineering dia-
logue systems is that we do not need to understand how language is used, the machine
will figure it out for itself (as far as it is able). The challenge is the amount of training
data required to cover all the necessary cases and, unlike a search engine where mea-
sured performance as low as 10% is useful, many errors social actors make are noticed
and need to be dealt with. The assumption here is that we want to know more about the
process of being a social actor, and know enough about it to be able to make a computer
to do it with sufficient competency.
Rather this paper is predicated on the notion that there is a wealth of vague, implicit,
context-dependent and often unconscious knowledge that is necessary for a social actor
to successfully inhabit a society [5, 6], and to show how such knowledge might be
incorporated into artificial agents. Such social knowledge is not immediately accessible
to an engineer as explicit knowledge and so the classic “waterfall” model of software
engineering, in which one starts by developing a detailed specification and follows up
with a development phase, is inappropriate. Instead, a process of entity enculturation –
learning how a CA should behave in context over a period of time – is required. Design
plays a part, but has to be leveraged by a substantial subsequent iterative process of trial
and repair [7]. This is not a “one off” method of making socially fluent agents, but a
method of repeatedly: (1) analysing records of their interaction in situ, then (2) affecting
84
a repair on the behaviour for this context. Thus, over time, embedding the agent into the
culture it perforce inhabits.
The core of this approach is the leveraging of a common narrative understanding
of interactions between people. In this non-scientists are asked to “tell the story” of
how a particular situation came about with a conversational agent as a starting point
in preparation for an iterative approach to repairing that agent. In subsequent iterations
they may be asked to comment upon an existing narrative, possibly entering alternative
descriptions at certain points. This interaction will be guided and constrained by a a
developer’s workbench that allows someone to both “script” future dialogue and analyse
recordings of past (real) dialogue with the machine by narrating the action, and would
capture the mechanism by which we social actors decide what to say and when.
2 Contributory Threads
Given the nature of the proposal, and its contrary direction it is useful to trace the
projects and results that have lead us in this direction.
2.1 The KT experiments
The KT experiments were a project to understand the issues and the potential for em-
bodied conversational agents (ECA) acting as virtual assistants [8]. As part of that
project, we conducted Wizard-of-Oz experiments, where a human covertly pretends
to be the conversational agent conducting the conversations, followed by interviewing
the wizard (KT) about her actions using a technique from applied psychology called
Applied Cognitive Task Analysis (ACTA) [9]. The aim was to populate a model of KT
doing the task, and then use that model to drive a virtual assistant performing the same
task. The model was “folk psychological” in that it her beliefs, desires and other mental
attitudes were used as theory to explain and identify the “causes” of her behaviour. For
these experiments the task was simply to have staff call our agent when they wanted to
use one of the cars from the Division’s car pool. Ultimately the task was a slot-filling
task: specifying which car, who was using it and the time.
The relevant results were twofold. Firstly, that politeness was more important than
getting the facts right. For various reasons KT’s “slot-fill rate” - how often she managed
to identify a piece of information in the caller’s utterances and enter it in the appropriate
slot - was just over 80%. A “fact error rate” of close to 20% might sound high but the
point is nobody minded and, although we didn’t measure it, we expect nobody noticed.
Why didn’t they mind or notice? Because of course KT would make appropriate apolo-
gies and gave explanations when she had forgotten what they said their phone number
was or where they were going. What is more, looking at the length of utterances, it was
easy to see how KT’s utterances could convey the same information in a more compact
form. Grice’s Maxims would suggest shorter is better (a principle popular with call cen-
tre industry) but KT did not want to use shorter utterances because, from the interview
process, it just wouldn’t be polite. As a scientist one might have theories about the con-
cept of face [10], but KT’s seems to have some system for doing social interaction that
85
uses politeness as an atomic concept. She had not read Brown and Levinson’s book [10]
and didn’t need to.
Secondly, it turned out that interviewing people about their everyday behaviour is
problematic. Interview techniques such as Applied Cognitive Task Analysis are in-
tended to make explicit expert knowledge that has become automatic. A fireman is
likely to be proud of his knowledge and pleased when the interviewer can identify some
piece of knowledge he had forgotten was special. Using ACTA to interview KT about
her “expertise” (which it is) in the use of language however, KT thinks of her knowledge
as just common-sense. The knowledge was implicit knowledge – a set of learnt skills as
to how to converse. What we were after was exactly that common-sense knowledge in
an explicit form so we could model and use it. Unfortunately it is common sense also in
that is common to all – it is knowledge that is shared and KT knows that. Interviewing
people about their common-sense knowledge, they quickly become suspicious about
the interviewer’s motivations. Why is he asking such “dumb” questions?
The lessons from this were that it was precisely the implicit social skills in conduct-
ing a conversation that were important but also difficult to get at in an explicit form. Just
as one can be able to ride a bicycle but not know how one does it, one can conduct a
sensible social interaction whilst not being able to specify how one does this. The very
ubiquity of this skill hides its subtlety and complexity.
2.2 Ethnomethods
The CA4NLP project applied an ethnomethodological variant of Conversation Analy-
sis [11] to analysing records of conversations such as those produced by the KT experi-
ments. This approach is predicated upon the notion that the researcher is a “member of
the same community of practice” as the discussants, and hence has access to the import
of their utterances. Thus, for example, a researcher’s introspections about whether or
not some communicative act of KT’s was polite is valid evidence because both get their
knowledge about the purpose of communicative acts from the same common pool. I do
not need to ask KT about her internal reasoning because it is the external effect that
matters and I have direct access to its significance. KT is right: I could give as good an
answer to my own dumb questions as she.
This method also implies a shift from a mechanistic view to a functional view. When
it comes to engineering spoken language interfaces, rather than trying to access the in-
ternal reasoning of the speaker as the KT experiments attempted, we want to look at
and model the way a social agent engages with the community of practice in which it
operates. Although engineering more as a process of adaption of function than design
will make some engineers uncomfortable, this is common practice for long-standing
artefacts that inhabit complex niches, such as sailing yatchs – nobody designs a yatch
from first-principles but rather adapts and tunes existing designs, tinkering with each
aspect in turn. What matters is how the yatch functions within the complex environ-
ment of winds and water. The same applies to computers that act in our social space.
A computer that says “no records match your request” might be being informative [12]
but is it playing by the rules of social engagement? Using the terminology from Con-
versation Analysis, what is the work done by “no records match your request” and is it
all and only what the expression was designed to do in the current context?
86
The methodology of Conversation Analysis is for the scientist to capture naturally
occurring text or speech of interest and ask “Why this, in this way, right here?” Whilst
using introspection as a means to assess scientific truth is a bad idea, introspection
about community knowledge is fine and provides detailed descriptions of the function
of utterances in context. Thus the CA4NLP project illustrated the use of introspection
to leverage understanding about utterances. It marks a shift away from attempting to
access an internal or foundational model, but rather capitalises upon the function of
utterances in context. It is the function of utterances that is constrained by common
usage, not the cognitive processes that give rise to them.
The trouble with Conversation Analysis however is exactly its strength in that it
provides a valid means of studying anything and everything. It does not provide any
guidance on what is critical to the structure of a conversation.
2.3 HCI and Grounded Theory
The SERA project put a talking “rabbit” in older people’s homes and collected video of
the resulting real human-robot interactions. 300 or so recordings of people interacting
with their rabbits were collected. The experiment had three iterations of: placing the
rabbit, recording the interactions, assessing the success of the system, and improving
the system software based on the assessment.
The motivation for the project was to see how different research groups would go
about this process. In general, all the groups could find interesting things to write about
the data, but the process of improving the system was primarily driven by those with an
HCI background who would, in the tradition of design-based engineering, simply have
an idea that could be tried. This creative process often worked, and would be followed
by a quantitative evaluation, but felt quite unsatisfactory when it came to understanding
what is going on.
The understanding that did feel like progress actually came from qualitative meth-
ods such as Grounded Theory [13] in the form of detailed analyses of how particular
conversations unfolded in those contexts. In particular people are very good at noticing
when a conversation is NOT right. As an expert I can tell you that I wouldn’t say “no
records match your request” in a given context and it is this data that needs to be the
raw material on which we base a science of machines in social spaces. However, this
micro-level of detail poses a problem when one needs to utilise the knowledge, for ex-
ample in terms of suggesting improvements to CA scripts. The detail needs to somehow
be accumulated in a more comprehensive social ability.
In some preliminary experiments in the SERA project, people were asked to say
what happened in a video recording we had of people interacting with one of the SERA
rabbits. This initially did not work very well because, although the plot in a film or
play is easily identified and summarised, natural recordings are just not that interesting
and rather messy. Instead recordings where things go wrong was chosen. This made the
‘crux’ of the story more salient.
87
2.4 Summary of threads
From the above experience we draw out several lessons. We see the importance of the
shared culture in terms of the common folk theory about what is happening, however
we also see that this common knowledge is implicit and not very accessible via direct
interrogation. We see the importance of examples learnt in context, in particular in terms
of their functional fit to the social circumstances. Finally this suggests that, in order
to transfer this implicit knowledge we might have to mimic the learning that usually
happens within the social sphere in terms of making mistakes and repairing them.
In order to make better conversational agents they will have to be inducted into
the society they are going to inhabit. Clearly, in general, this is extremely hard and
takes humans a couple of decades of time but here we might be aiming for an agent
that copes tolerably well (on the level of a polite 6 year old) in a single context (or a
very restricted range of contexts). Here we aim to imitate the cycle of trial, error and
repair on a small scale, hoping to make up for the small number of cycles with a more
intelligent repair stage composed of analysis with repair leveraging some of our own
innate understanding of social behaviour. Each iteration in a particular context will (on
the whole) result in an incremental improvement in social behaviour. The hard part of
this cycle (other than the number of times it may have to be done) is the analysis and
repair stage. We will thus concentrate on this in this section.
3 Capturing the implicit
The idea presented in Figure 1 is to iteratively improve an in situ CA, each iteration
through allowing a bit more of the explicit and implicit knowledge concerning the ap-
propriate social behaviour to be captured in the knowledge base and hence used to tune
the CA rules. Each iteration the CA, in its current state of development, will deployed
and new records of its conversation with humans made, since it is difficult to predict the
full social effect of any change. This iterative cycle imitates, in a rough manner, the way
humans learn appropriate social behaviour: observing others, noticing social mistakes
and iteratively adapting their behaviour.
Clearly there are several parts of this cycle that could be discussed in detail. How-
ever, here we will concentrate on motivating and outlining how the user-interface that
prompts and structures the review of the conversational records by the native expert.
The nub of this process is how to elicit the, largely implicit, knowledge about social
behaviour using the responses of the third party reading and reacting to the records of
the conversation.
3.1 Vygotski
Vygotski’s insight used here is that plays and novels exist because they provide plau-
sible accounts of human behaviour. Theatre is the flight-simulator of life [14] and pro-
vides a means of exercising our ability to understand the motivations and behaviour of
others. We do think about other minds when we communicate – indeed it turns out to
be a critical skill [15–17] – and we do it in terms of beliefs, desires and other mental
88
Fig. 1. Summary flow chart of the proposed iterative method
attitudes. What is more, we expect our conversational partners to do the same, with
the same model. When it comes to communication, the truth of our folk model of other
people’s thinking doesn’t matter; what matters is that it is shared. Rather than looking
inside KT’s head to see how she would deal with social relations, the idea is to look
at some kind of collective understanding of events – what is the shared knowledge that
creates the context against which a human social actor figures out the significance of
communicative acts? Rather than sitting in an arm-chair and classifying utterances ac-
cording to the effect they have on a idealised conversational partner, the idea is to look
at real interaction data and document the effect in context. Rather than classifying ut-
terances as REQU EST IN F ORM AT ION , or GREET IN G [18], the idea is to
record the “work done” by utterances in the place they are produced. This can be done
by any member of the community of communicators and does not require a scientific
theory. Consider this example of a conversation between a doctor and a patient taken
from the Conversation Analysis literature:
Patient: So, this treatment; it won’t have any effect on
us having kids will it?
Doctor: [silence]
Patient: It will?
Doctor: I’m afraid the...
89
The “work done” by the silence is of course to disagree and some might be tempted
to mark it up as an explicit answer, but there are many different things that the doctor
could say at this point, with a wide range of “semantics” but all with the same effect.
The Vygotski argument is that human story telling gives sufficient detail of events
that any socialised human (who is part of the same community) can fill in the gaps to
produce a set of linked causal relationships for the story to make sense. This requires
contextual knowledge (e.g. teddy bears are toys and children like to play with toys)
and ”hard-wired” knowledge (e.g. children often want things and act in ways that bring
them about).
One might think that human-machine interactions would be less fraught and thus
simpler. Indeed those working on commercial spoken language interface design try
very hard to make this true using techniques such as menu choice. However, even in
the DARPA Communicator data where the systems were only slightly more natural
than those one might find in a bank, there are examples where the work done by an
utterance such as “no” goes well beyond what might be seen as the semantics of the
utterance [19]. One can not escape the importance of social etiquette.
It is this process, and the interface to support it, that we will now describe.
3.2 An example
Consider some video data captured spontaneously (Figure 2) during the development
of the SERA set-up.
Fig. 2. Mike and the rabbit talking with Peter
To set the context, “Peter and Mike have been talking in Peter’s office where he
has a robot rabbit that talks to you and that you can talk to using picture cards.” Two
narrators were given this sentence and asked to watch the video. They were then asked
to, independently, finish the story in around 200 words. The resulting stories appear in
Figure 3.
There are many differences, and many things were left out entirely. There does
however appear to be general agreement on core events. Neither narrator mentioned
90
Narrator 1 Narrator 2
It is time to go home so Peter takes his keys from the Peter is about to do something to wake the rabbit up again
rabbit. Mike notices this and says “Isn’t it supposed to and as he is about to speak, it says hello. Peter gestures to
say hello?” Peter is about to say something when the Mike that it is now talking as expected. Peter presses the
rabbit says: “Hello, are you going out?” Peter replies video button to record the interaction. Mike laughs as it
that he is (using the card and verbally) and the rab- talks. It asks Peter if he is going out, to which he responds
bit tells him to have a good time, bye. Mike picks verbally that he is, showing the rabbit the card meaning
up a card and shows it to the rabbit, but nothing hap- yes. Seeing Peter’s interaction, Mike tries using the cards
pens. He thinks this make sense as the rabbit has said to interact with the rabbit himself. It does not respond and
goodbye but Peter thinks it should work and shows Mike suggests that this is because it has said goodbye and
the rabbit another card. Mike sees that he has been finished the conversation. Peter tries to reawaken the rabbit
showing the cards to the wrong part of the rabbit and with another card. Mike sees that he had put the card in
gives it another go. Still nothing happens and Mike the wrong place. He tries again with a card, after joking
tries to wake it up with an exaggerated “HELLO!”. that the face card means “I am drunk”. Peter laughs. When
Peter stops packing his bag and pays attention. Mike the rabbit does not respond, Mike says “hello” loudly up
tries getting the rabbits attention by waving his hand to the camera. Peter says he is not sure why there is no
at it. Still nothing happens. Mike looks enquiringly at response while Mike tries to get a reaction moving his hand
Peter as if to ask “what’s happening” He says “that’s in front of the system. They wait to see if anything happens,
a new one” and goes back to his packing. Mike takes Mike looking between the rabbit and Peter. When nothing
his leave at this point. Peter finishes his packing, and, happens, Peter changes topic and they both start to walk
as he leaves says to the rabbit “You’re looking quite away. Mike leaves. As Peter collects some things together,
broken.” walking past the rabbit, he looks at it. Before leaving the
room he says to the rabbit “you’re looking quite broken”.
Fig. 3. Two narrative descriptions of the same event.
1. Peter is about to say something and is interrupted by the rabbit
2. the rabbit asks if he is going out, Peter’s verbal and card response
3. the rabbit says bye
4. Mike’s attempt to use a card and the non-response of the rabbit
5. Mike’s explanation (that the rabbit has already said bye)
6. and Peter showing the rabbit another card
7. Mike sees that he has been showing the card to the wrong part of the rabbit and has another go
8. the rabbit does not respond
9. Mike says “Hello” loudly
10. Peter acknowledges it doesn’t look right
11. Mike tries again by waving his hand in front of the rabbit
12. no response from the rabbit
13. Mike looks at Peter
14. They give up
15. Mike leaves
16. Peter leaves saying “You’re looking quite broken” to the rabbit
Fig. 4. The third-party common ground.
91
the filing cabinet nor the clothes participants were wearing. No comment on accents
or word usage; no comment on grammatical structure nor grounding, nor forward and
backward looking function. Whatever it is that the narrators attend to, it is different
to the type of thing that appear in classic annotation schemes. It does however seem
to be shared and, the claim is, shared by the community of practice at large. Both the
narrators and the participants are working from a shared theoretical framework – not
from raw undifferentiated data – that guides and selects which sense-data is attended
to. However this shared framework is implicit.
Accounts of the action in the video data as written down by the narrators are of
course descriptive in that they are written to ‘fit’ past events. The claim is that they are
also predictive. If Mike wants to use the system, then it would be surprising if others
did not want to. If failure to work causes disappointment in Mike, it is likely to also
cause it in others. Having a predictive model of events we are well on the way to having
prescriptive rules that can be used to drive conversational behaviour.
But first however let’s look at how we might move more formally from the stories
in Figure 3 to the summary in Figure 4.
3.3 An interface to support capture of social knowledge
The problem is that even if they observe the same thing, they may not describe it in the
same way and, unless the descriptions are the same, a machine cannot recognise them as
the same. In the example above the two observers produced two narrative descriptions
and it is claimed they are the same, but how would one measure the sameness? Without
a machine that can understand what is written, human judgement is involved and claims
of researcher bias are possible. How might comparative narratives [20] be produced that
are the same to the exacting standards required for machine understanding?
The proposal, should one want to re-do this preliminary experiment properly, is to
use the techniques seen in industrial machine translation for the production of opera-
tor and repair manuals. Companies like Mitsubishi and Caterpillar [21] have systems
that allow them to produce manuals in one language and then, at the push of a button,
produce the same manual in all of the languages for countries to which they export.
The way this is done is to have the author of the manual write in a restricted version of
the source language and provide the tools to guide the writing process. The process of
authoring with such tools will be familiar to us all because modern text editors provide
spelling and grammar checking assistance in much the same way. The primary differ-
ences being of course that the list of recognised words is much smaller and the grammar
rules much stricter, and the process of breaking those rules is not simply for the system
to ignore it, but to ask the user to add the new word or expression to the system. For
instance the author might really want to use the term “airator” and the system would
allow that but ask the author if it is a noun or an adjective, a count noun, what its se-
mantic preferences are, and if it is masculine or feminine in French. The word would
be added to the lexicon and, the next time an author wanted to use it, the system would
have enough detail to translate it correctly or ask this new author how it should be used
in the current context.
If one wanted to re-do the experiment above more formally, the approach would be
to reproduce a “translators work bench” and, rather than having it translate to another
92
language, have it “translate” to a different style in the same language. This authoring
process works for machine understanding for translation; there is no reason to think it
wouldn’t work for this new application if one really wanted to do it. But why bother?
The ultimate aim is to script dialogue for synthetic characters and the proposal is that,
rather than stopping at narrative descriptions, the system would go on to explore coun-
terfactuals.
3.4 Narrative descriptions capturing context
The aim is to classify utterances as the same in context and hence be able to program an
agent to give a particular response to any input from the same class. Using a classic an-
notation scheme one might decide that if its conversational partner produces something
in the class of QU EST ION , then the agent should produce an AN SW ER. This func-
tionalist model of sameness applies to everything from chatbots in which something like
regular expressions are used to recognise inputs are the same, through to full planning
systems such as TRAINS [22] in which input recognition is set against the current goals
of the system. The variation proposed here is that the functionalist definition of same-
ness is embedded in narrative. Two expressions are the same if and only if, for every
narrative in which expression #1 occurs the outcome of the story would not change if
expression #2 was used.
Given such a definition of sameness, it is only in trivial cases that expressions will be
universally the same. It is far more likely that expression #1 and #2 will be equivalent for
some narratives and not others – the equivalence is context dependent, and this provides
an opportunity to question an observer about the features of the context that determine
when an existing response to input might or might not be appropriate for another input.
As an example of the type of thing we have in mind Figure refCTAprobes gives a
table showing the type of question that was asked of KT. It would appear that some
of these questions would be a useful way to explore context with our observers and,
importantly, the questioning could be automated. An observer might provide a narra-
tive description of a particular recording of an interaction and, at some point in that
description the computer might say S where the rules being used by the machine might
have equally produced S 0 . An annotator’s work bench could ask the human if S and S 0
would be functionally equivalent in the narrative given. If not, the workbench could ask
what (in the context) makes S 0 inappropriate, and perhaps ask the annotator to develop
a rule that distinguishes the context for S and S 0 . Similarly the system could ask the
observer if he or she can formulate an alternative to S and S 0 that would be better, and
develop a rule to distinguish the alternative utterance from S and S 0 .
The above gives a flavour for the proposed work bench designed to enable non
scientists to use their expert knowledge of language use to create context dependent
rules so the system can decide what to say when. The aim is to combine the direct
contact with the data normally seen in an annotation tool such as Anvil [23] with the
creative process of scripting conversation for the agent. In effect the aim is to formalise
the process (and add some theory) that people use when they script chat-bots using
AMIL by pouring over log files.
93
Fig. 5. O’Hare et al 1998 - the revised CDM probes.
Goal specification What were your specific goals at the various decision points?
Cue identification What features were you looking at when you formulated your decision?
Expectancy Where you expecting to make this type of decision during the course of
the event?
Describe how this affected your decision-making process
Conceptual model Are there any situations in which your decision would have turned out
differently?
Describe the nature of these situations and the characteristics that would
have changed the outcome of your decision.
Influence of uncertainty At any stage, wee you uncertain about either the reliability or the rele-
vance of the information that you had available?
At any stage, were you uncertain about the appropriateness of the deci-
sion?
Information integration What was the most important piece of information that you used to for-
mulate the decision?
Situation awareness What information did you have available to you at the time of the deci-
sion?
What information did you have available to you when formulating the
decision?
Situation assessment Did you use all the information available to you when formulating the
decision?
Was there any additional information that you might have used to assist
in the formulation of the decision?
Options Were there any other alternatives available to you other than the decision
that you made?
Why were these alternatives considered inappropriate?
Decision blocking - stress Was there any stage during the decision-making process in which you
found it difficult to process and integrate the information available?
Describe precisely the nature of this situation.
Basis of choice Do you think that you could develop a rule, based on your experience,
which could assist another person to make the same decision success-
fully?
Why/Why not?
Analogy/generalization Were you at any time, reminded of previous experiences in which a
similar decision was made? Were you at any time, reminded of previous
experiences in which a different decision was made?
94
4 Conclusion – Towards the Iterative Embedding of Implicit
Social Knowledge
This proposed approach seeks to take seriously the subtlety of social behaviour, result-
ing from the “double hermeneutic” which relies on the fact that encultured actors will
have a ready framework of how to interpret the social behaviour of others, including the
expectations that others will have of them. In particular it is important how it is that so-
cial knowledge is embedded within a complex of social relations and knowledge, which
makes it hard to formalise in general. We do not expect that this will be easily captured
in a “one-off” analysis but require a iterative approach based on repair. The difficulty of
the task means that a number of approaches will need to be tried to leverage little bits
of social knowledge each iteration. The key parts of this are the interactive capture of
social information from a third party and the use of that knowledge to inform an update
of the CA rules. We have not talked about the latter here – currently it will require sig-
nificant programming skill. The ultimate aim would be to eliminate this programmer,
so that this iterative process could be used by non-experts, utilising their own implicit
expertise, to socially ”educate” their own CA. This is illustrated in Figure 6.
Fig. 6. Flow chart of the process without the programming expert
95
References
1. de Angeli, A.: Stupid computer! abuse and social identity. In de Angeli, A., Brahnam, S.,
Wallis, P., eds.: Abuse: the darker side of Human-Computer Interaction (INTERACT ’05),
Rome (September 2005) http://www.agentabuse.org/.
2. Mel’cuk, I.: Meaning-text models: a recent trend in soviet linguistics. Annual Review of
Anthropology 10 (1981) 27–62
3. Hovy, E.: Injecting linguistics into nlp by annotation (July 2010) Invited talk, ACL Workshop
6, NLP and Linguistics: Finding the Common Ground.
4. Young, S.J.: Spoken dialogue management using partially observable markov decision pro-
cesses (2007) EPSRC Reference: EP/F013930/1.
5. Edmonds, B.: Complexity and scientific modelling. Foundations of Science 5 (2000) 379–
390
6. Edmonds, B., Cershenson, C.: Learning, social intelligence and the turing test: Why an ‘out-
of-the-box’ turing machine will not pass the turing test. In Cooper, S.B., Dawar, A., Lwe,
B., eds.: Computers in Education. Springer (2012) 183–193 LNCS 7318.
7. Edmonds, B., Bryson, J.: The insufficiency of formal design methods – the necessity of an
experimental approach for the understanding and control of complex mas. In: Proceedings
of the 3rd International Joint Conference on Autonomous Agents and Multi Agent Systems
(AAMAS’04). ACM Press, New York (July 2004) 938–945
8. Wallis, P., Mitchard, H., O’Dea, D., Das, J.: Dialogue modelling for a conversational agent.
In Stumptner, M., Corbett, D., Brooks, M., eds.: AI2001: Advances in Artificial Intelligence,
14th Australian Joint Conference on Artificial Intelligence, Adelaide, Australia, Springer
(LNAI 2256) (2001)
9. Militello, L.G., Hutton, R.J.: Applied cognitive task analysis (ACTA): a practitioner’s toolkit
for understanding cognitive task demands. Ergonomics 41(11) (November 1998) 1618–1641
10. Brown, P., Levinson, S.C.: Politeness: Some Universals in Language Usage. Cambridge
University Press (1987)
11. ten Have, P.: Doing Conversation Analysis: A Practical Guide (Introducing Qualitative Meth-
ods). SAGE Publications (1999)
12. Traum, D., Bos, J., Cooper, R., Larson, S., Lewin, I., Matheson, C., Poesio, M.: A model of
dialogue moves and information state revision. Technical Report D2.1, Human Communi-
cation Research Centre, Edinbrough University (1999)
13. Urquhart, C., Lehmann, H., Myers, M.: Putting the theory back into grounded theory: Guide-
lines for grounded theory studies in information systems. Information Systems Journal 20(4)
(2010) 357–381
14. Bennett, C.: But mr darcy, shouldn’t we be taking precautions? Keith Oatley (quoted in)
(July 2011)
15. Grosz, B., Sidner, C.: Attention, intention, and the structure of discourse. Computational
Linguistics 12(3) (1986) 175–204
16. Eggins, S., Slade, D.: Analysing Casual Conversation. Cassell, Wellington House, 125
Strand, London (1997)
17. Tomasello, M.: Origins of Human Communication. The MIT Press, Cambridge, Mas-
sachusetts (2008)
18. Jurafsky, D., Shriberg, E., Biasca, D.: Switchboard swbd-damsl shallow- discourse-function
annotation coders manual. Technical Report 97-01, University of Colorado Institute of Cog-
nitive Science, Colorado (1997)
19. Wallis, P.: Revisiting the DARPA communicator data using Conversation Analysis. Interac-
tion Studies 9(3) (October 2008)
96
20. Abell, P.: Comparing case studies: an introduction to comparative narratives. Technical
Report CEPDP 103, Centre for Economic Performance, London School of Economics and
Political Science, London, UK (1992)
21. Kamprath, C., Adolphson, E., Mitamura, T., Nyberg, E.: Controlled language for multilin-
gual document production: Experience with caterpillar technical english. Proceedings of the
Second International Workshop on Controlled Language Applications 146 (1998)
22. Allen, J.F., Schubert, L.K., Ferguson, G., Heeman, P., Hwang, C.H., Kato, T., Light, M.,
Martin, N.G., Miller, B.W., Poesio, M., Traum, D.R.: The TRAINS project: A case study in
defining a conversational planning agent. Journal of Experimental and Theoretical AI 7(7)
(1995) 7–48
23. Kipp, M.: Gesture Generation by Imitation - From Human Behavior to Computer Character
Animation. PhD thesis, Saarland University, Saarbruecken, Germany (2004)
97