The SERA project put “robot rabbits” in older peoples homes and recorded what happened. The challenge is now to use that data to develop better rabbits, but how? We are currently working on a methodology for distilling this data down into explanatory narratives, but in the mean time we are working on the idea that the essential nature of the SERA interface (and other conversational agents) is that it is intentional - it is an interface that sets out to have people ascribe beliefs and desires to it. According to Tomasello, this is not enough however. An intentional interface also needs to intend to help - it needs to be cooperative. What this means in detail is fleshed out in the context of an IVR system - a computer that answers the telephone.
A year on from the SERA project - Social Engagement with Robots and Agents
this paper looks back on what we did, and attempts to put the lessons learned in
a historical context. Our vision was to use a talking robot rabbit (an augmented
Nabaztag) as long term “companion”. Obviously it is beyond us to create a
perfect simulation of a human conversational partner, but was current technology
able to capture the essence of what is needed? The answer was no, but the
experience certainly prompted some thinking about that essence. This paper
develops that thinking and in doing so, o↵ ers a “grand unified theory” of HCI
based on Dennett’s Intentional Stance [
SERA
The SERA project was funded under the FP7 Theme 2.2: Cognitive Systems,
Interaction and Robots, and the aim was to put real robots in real people’s
hallways and kitchens and record what happens. The work was done with the
School of Health and Related Research at She eld (ScHARR) which had
extensive experience recruiting subjects from the broader community, and which was
working with “smart homes” with a view to “life-style reassurance” in which
people living alone could be assured that, should something happen to them,
help would be at hand. Building a state-of-the-art companionable robot is a
project in itself [
That was the set-up, but another challenge was deciding what the rabbit
should say. We settled on the popular scenario of an “exercise companion”.
Using the Trans Theoretical Model of behaviour change (TTM) [
Before introducing the theory, the shared conclusions from the SERA project were, first, don’t try to use ASR in the wild - the Siri publicity (formal and viral) is a dream. That kitchen is not my kitchen: there are no kids practicing the tin whistle, no oil sizzling on the hob, no radio playing, no extractor fan, no tra c and no refrigerator humming away. Indeed she is perfect as well with a nice East Coast accent with no Yorkshire clipping or Australian vowels. And the recipe - no star annis, or dried paw paw; nothing out of the ordinary. Put a speech recognition system in a kitchen and, we discovered, word error rates are too low for even a handful of key phrases.
Second, managing attention is a big issue when the interface is sensing the
environment, is always on, and can be proactive. The classic HCI interface is
passive (see below) and needs to be “poked” before it behaves. Our existing
model for a proactive interface is the alarm that demands attention. In between
is the telephone that could be demanding when its location was fixed but
mobile phones would, ideally, be more “socially aware” of the context. The SERA
rabbits used a PIR security sensor to detect with people were near but is she in
a hurry? Is she making an omlette? [
Third, people have “idiosyncratic” behaviour. Where one person gets cross and yells, another laughs while another roles her eyes. Another subject may frown or not respond at all. Naturally the notion of a response to “the same” event is also problematic without a framing theory but this issue is well known; what stood out for us all was the huge range of responses across socio-economic backgrounds.
Finally, we can say that there is no consensus on what to do with the data we
collected. We could all publish papers, but how can the data be used to advance
the state-of-the-art? There has been some work on a better methodology for
looking at the data [
How the mind works1 The proposal is that the SERA rabbits were not simply a conventional humancomputer interface with a speech recognition front end, but were instead an attempt at an intentional interface. This is not to say that the SERA interface was unique – many have attempted similar things – the point is to introduce a class for interface for which SERA is an example. In order to compare and contrast, the observation is that we can classify human-computer interfaces based 1 Thanks to Steven Pinker for the title of this section and the next. on how the user goes about understanding the computer, and that interesting distinctions can be drawn by looking at Dennett’s position on intentional systems.
Dennett, argues that the study of minds is di↵ erent to the study of brains, and that the wide spread use of “folk psychology” in the Social Sciences is perfectly valid as science. For realists there is little doubt that minds reside in the hardware of brains, but studying brains is not necessarily going to provide explanations for why things are the way they are. As a scientist one might have a theory of id, ego and super-ego, or as a mathematician one might have an elegant Bayesian model of how brains work that is meant to explain things, but Dennett’s line is that the psychology we use in our everyday lives is equally valid as a scientific theory and more e cient .
Dennett argues that humans use three di↵ erent approaches, or stances, when trying to predict the behaviour of something. When a system is fairly simple - balls on a level table perhaps - then we can use a causal model to predict future events. Tapping the white ball in a particular way will cause it to role over to the red ball and knock it into the centre pocket. Taking this physical stance, people use their knowledge of hundreds (if not thousands) of highly reliable “facts” about the way things behave to assemble chains of causal events to predict the future. Dennett was writing at the time of good old fashioned AI and so the nature of these facts, as we now know from the work in computer science, is problematic and (apparently) based on situated action. But possible enumerations and classification of the base facts is not the point; the point is we can and do reason causally. Taken to its extreme, this is the idea of a clockwork, deterministic universe and that ultimately “there is only physics”.
Another way we humans predict the future is by knowing what something
is designed to do. Pressing the brake pedal when driving, one does not reason
about hydraulic fluid, but simply knows what that pedal is meant to do. An
alarm clock is too complex to follow the internal workings in a causal sense but,
knowing what it is designed to do, one can set it in the evening and predict that
it will wake you in the morning. This is of course where classic HCI is based with
advice on how to create good interfaces being things like making sure that the
system works as designed, and that the user has a clear idea of the function of the
design (e.g. Interaction Design: beyond human-computer interaction (2ed) [
The intentional stance is what we use when a system is too complex to predict with the physical stance, and the purpose of system - what it is designed for - is inaccessible to us. We humans have a strong tendency to assume that something capable of autonomous action will do what it believes is in its interests. That is, that the system will have desires, and that it can plan its actions to achieve (some of) those desires given its beliefs about the current state of the world. This tendency is very strong in us. Seeing two children tugging at a teddy bear, the casual observer will assume they both want it. When playing chess against a computer, I do not reason about the causal behaviour of registers and electricity, but rather predict the future by reasoning along the lines of it wanting to take my bishop. The consequences of a rational agent wanting something do not need to be spelt out for us; we just know. We are also likely to explain things that are not rational action with this model and Dennett gives a lovely example of someone explaining that electricity normally wants to take the shortest path but sometimes it “gets confused”. 3.1
The Human-Computer Interface from the stances
Current HCI best practice can be critiqued as using a “tool” metaphor in which
the computer is wielded by the user to achieve his or her goals. This is fine as far
as it goes and has the advantage that as long as the tool does what it is designed
to do, the user is responsible for outcomes. Hit your thumb with a hammer
and there is only yourself to blame. Using such a metaphor, the guidance on
HCI design is about making the design clear, and the consequences of an action
explicit and immediate [
Extending the metaphor, the sexy human-computer interfaces are those based on the physical stance. On the surface there is a class of interface that exploits the “facts” we have about the physical world with desk tops as a place to “put” things temporarily, folders that “contain” stu↵ , and recycle bins for the things we don’t want any more. Today’s touch screens allow things to be “flicked” and multi touch screens allow things to be “stretched” in a way that tend to obey our facts about the (physical) world. At a deeper level, many modern interfaces - especially those designed for new markets such as children - not only allow, but actively encourage exploration. In e↵ ect they encourage the user to discover things about causality in the virtual world that mirror the “hundreds or thousands of facts” we know that support the physical stance in the physical world.
This exploration process is clearly what Suchman points to in her classic
work on situated action and the photocopier [
The proposal is that the essence of our rabbit interface is that (it looks as if) it behaves in accordance with our intentional stance. At first blush the distinguishing feature of the SERA rabbits was the speech recognition. On reflection the distinguishing feature ws that the PIR meant that our rabbits were proactive about initiating a conversation. In accordance with the call, our engineering aim was indeed to sense and react to the environment as a robot is expected to do and this meant that it is hard not to think of the rabbit as wanting to do things. From the user’s perspective, a rabbit has its own agenda and the user slips very easily into taking an intentional stance. Once a conversation was started - as was clear from the video evidence - the rabbits were never good at negotiating shared goals. The problem was that the system did not take an intentional stance on the functioning of its user and was thus not able to negotiate a shared intention.
It turns out that the intentional nature of human-human communication is well recognised in linguistics proper. What our rabbits need is a better approach to dialogue management. 4
The language instinct
Computer science as applied to natural language moved out of the arm chair in
1989/90 and that research community generally accept that data driven research
is the way forward. The critical mass however use statistical models and, like
the behaviourists of old, abhor any notion involving “mental attitudes”. In the
last 10 years this has been applied to dialogue and so, the argument goes, we
do not need to study how language works because (given enough data) machine
learning techniques will enable computers to simulate conversational behaviour
without theory. Partially Observable Markov Decision Processes (POMDP) have
been applied to the dialogue problem [
This is a noble aim but in practice human intervention is required to make
these systems work. In practice ML techniques are not trained on raw speech or
text, but rather on tag sequences where the tags are from a set of dialogue acts
or DAs. There is no consensus on what should go into these sets of tags and in
general each annotation scheme adapts an accepted set of dialogue acts to the
particular application domain. From a linguistics perspective the methodology
is sequence analysis [
Much of Linguistics and the social scientists however do data driven research
but take the line that mental attitudes are causal in human a↵ airs and, as
Dennett argues, that a valid science can be based on such concepts. The argument
is made beautifully by ten Have [
As is often the case there are notable exceptions - classically Grosz and Sidner
[
The need for cooperation is made clear when we take a closer look at what
linguists have said about the mechanism of language. Conversation Analysis [
A: What’s that floating in the wine? B: There aren’t any other solutions.
You will try very hard to find a way of interpreting B’s turn as somehow
an answer to A’s question, even though there is no obvious link between
them, apart from their appearance in sequence. Perhaps you will have
decided that B took a common solution to a resistant wine cork and
poked it through into the bottle, and it was floating in the wine.
Whatever explanation you came up with, it is unlikely that you looked at
the example and simply said ‘it doesn’t make sense’, so strong is the
implication that adjacent turns relate to each other [
The appearance of an utterance immediately after another in an interaction
to which the partners are committed (that is, a conversation) causes the hearer to
work hard at recognising the intent of the speaker. The social pressure on doing
this and cooperating in general is captured by Tomasello. Quoting at length:
Thus, from the production side, we humans must communicate with
others or we will be thought pathological; we must request only things
that are reasonable or we will be thought rude; and we must attempt to
inform and share things with others in ways that are relevant and
appropriate or we will be thought socially weird and will have no friends.
From the comprehension side, we again must participate, or we will be
thought pathological; and we must help, accept o↵ ered help and
information, and share feelings with others, or we risk social estrangement.
The simple fact is that, as in many domains of human social life, mutual
expectations, when put into the public arena, turn into policable social
norms and obligations. The evolutionary bases of this normative
dimension of human communication in terms of public reputation, will be ...
[
Looking again at Eggins and Slade’s example, apes it seems are not hardwired, preprogrammed and/or socialized into putting the e↵ ort in and would simply say “it doesn’t make sense” and move on. What our computers need as social actors is the ability to account for the communicative acts of its human companions and to do that requires intention recognition and a willingness to put in the e↵ ort. 5
Practical intention recognition
Intention recognition and pro-active cooperation are core to human
communication of all kinds but it is not enough to say this or even prove it. If researchers in
an engineering faculty are going to embrace it, there needs to be a means of
implementing it, and the rest of this paper shows how this can be done for limited,
but useful, cases. As is often the case with non-incremental development [
For the next project the aim is to demonstrate an intentional interface with
an IVR system and the scenario under consideration is the classic directory
assistance application. With these systems a caller can ring the institution and
talk to a computer which puts them through to the required individual. The
classic approach would hold the relevant information in a relational database
and, in the spirit of Meaning-Text Theory [
M/C Welcome to University of She eld directory assistance. Who do you wish to contact?
USR Mark Hepple please.
An ASR module would produce the text from the voice signal, a parser normalize the grammar, and a language understanding module might map that into a canonical representation of the meaning. In the case of database access, the canonical form might take the shape of the SQL query:
SELECT ALL FROM phonebook WHERE “familyName=’Hepple’ & givenName=’mark”’ For the University of She eld phone book, such a query returns: givenName familyName dept extn
Mark Hepple DCS 21829 From this result a text generation system that uses some form of pronoun rewriting could say “His number is 21829.”
This is great when things go well. Trouble occurs however when the user’s query does not return a single row. 5.1
If the caller asks for Mark Hawley in Computer Science, the resulting query returns no rows as Hawley is not in the Department of Computer Science. What should the system do?
Using a classic HCI approach the aim would be to make it clear to the user that he or she is using a relational database and to point out that it is the user’s query that is resulting in unhelpful output. For many of the DARPA Communicator systems a common solution to no result (no rows in the resulting table) was to remind the caller that he or she could change their query by adjusting the parameters. The user is using a tool, and it is the user’s responsibility to use it as designed.
At the other extreme, if the caller asks for Mark in Computer Science, the resulting query returns: givenName familyName dept extn Mark Hepple DCS 21829 Mark Stevenson DCS 21921
Mark Ellerby DCS 21856 But which Mark does he or she mean? When the user’s query matches multiple rows a graphical user interface can present all the rows and this is sometimes attempted with IVR systems. Once again in the spirit of the Communicator systems the system might say:
M/C: “There are 16 people with that name, the first is Mark Heppe in Computer
Science. Is that who you are after?”
Usr: No M/C: The second is Mark Stevenson in Computer Science. Is that the person you are after? Usr: No
...
The computer as tool metaphor may work, but can an intentional approach provide an alternative? 5.2
It seems a computer behaving as a social actor needs not only to be right, but also seen to be helpful, and the challenge in the first instance is to come up with helping strategies that the system can introduce. Introducing a strategy requires mixed initiative, not just at the information level, but at the level of intent. The following discussion shows what this means and does it in terms of conversational strategies implemented as plans in a BDI architecture.
The Belief, Desire and Intention architecture was introduced by the software agents community as a means of balancing reactive and deliberative behaviour in a constantly changing environment. The approach does not do planning in the traditional AI sense, but rather manages commitment to plans. The usual
This strategy is good as far as it goes, but the machine’s question assumes the user knows. Thus, this helping strategy might fail if the user is unsure. For a BDI architecture this is not a problem — the architecture was introduced to handle plan failure — and the system simply looks for another plan. The success or failure of this plan will of course depend on what the user says next. Failure however is not bad; what is important is that the system is seen to be be trying to help. Consider:
M/C: There is a Mark Hawley in Health? (1
second) M/C: I can give you the number for the Departmental Secretary in Computer Science?
Usr: Mark Hawley please M/C: Professor Mark Hawley in the School of Health and Related Research is on 219...” Once again the point is the “unfolding” of conversation and a socially ept intentional interface has a responsibility to help.
The fourth case is where there are many rows in the table - 0,1,2,many - and when this happens there is often a misunderstanding. Consider someone who thinks he is calling the number for the Department of Computer Science and says:
Usr: Mark please.
M/C: Err you have called directory assistance
for the University of She eld.
M/C: I’m sorry, who are you after? Putting on one’s CA hat, the “work done” but the machine’s response is to appeal to the caller’s sense of fairness. As Tomasello says, people have a sense of fairness and the strategy here is for the system to explain what its job is, suggesting that it is unfair to expect it to be able to help in this case.
Intention recognition is hard for a machine but we can get some way there by working from a fixed set of plans. At this stage the above conversational strategies have been implemented but the system has not been evaluated in an operational setting at this stage. The point of this paper however has been to introduce an alternate model for HCI, and to demonstrate that it is not just hand waving - Tomasello’s claims are concrete and implementable. 6
Conclusion ICT is amazingly versatile, enabling us to create the information systems we want, with the interfaces we want. Without limitations, the designer is ultimately responsible for any problems. It is very tempting in these circumstances for us to favour interfaces that exploit the user’s design stance which shifts some responsibility to the user - the user ought to RTFM (read the manual) and then wield the tool as we designed it to be wielded.
The sexy new interfaces - be it 2010 or 1985 - exploit the user’s physical stance in which our understanding of cause and e↵ ect in the physical world is mapped onto virtual events.
The claim being made is that the essence of “human-like” interfaces — from embodied conversational agents to robot companions through chat-bots to speech interfaces and IVR systems — is that the user takes an intentional stance. Although making these systems more like humans is interesting in its own right — adding micro movements to ECA, emotion or persona to chat-bots — the feature of human communication that provides an opportunity for HCI is the intentional nature of the human interface. This is not enough however because, according to Tomasello, a social actor in human society also needs to be cooperative.
Such claims might be seen as too abstract, but the paper gives an interpretation of these principles in the context of an IVR system providing directory assistance.