This paper introduces a model of human communication in which `accounting-for' is the basis of meaning, and argues that interruptions should be handled in the same way as any other speech act. The model has at its core the idea that human languages are inherently intentional { we focus on our conversational partner's goals { and that what is needed is mixed initiative at the level of intent. It would seem interruptions can rea rm or contradict the speaker's current intent and the paper nishes with a description of our (very) shallow approach to intention recognition.
A speech interface for robots and computers has been part of the AI vision from
the very beginning but after 50 years of trying it turns out that talking, like
walking, is far more complex than the early visionaries anticipated. Over the
last few years, large corporations with commensurate research budgets have
decided the technology has come of age, but their approach to dialog management
however is hardly more sophisticated than chat-bots. These corporations
certainly employ the best and brightest but following perhaps the Microsoft foray
with the Desktop Assistant (the Paper-Clip) commercial ventures in the area
tend to be conservative. Google Assistant and Siri basically perform internet
searches and Amazon Echo is primarily a home automation controller. Indeed it
is interesting to note that, when the Amazon Echo needs to interrupt the user,
it does it via the well understood mechanism of setting an alarm. Technically
Alexa could say "Excuse me it has been 10 minutes" but instead the Echo plays
a ring tone. The Echo does this because, as we discovered in 2008 when we put
robot rabbits in older peoples homes, it is surprisingly hard to get a machine
performing as a social actor to initiate a conversation [
It turns out this kind of conditioning is endemic in the way we communicate.
In order to chart a course through this complex network of norms and social
relations we have turned to a suite of techniques from the human sciences broadly
under the banner of Conversation Analysis[
The conventional wisdom is that natural languages are the de nitive symbol
system and computers are, in a very literal sense, universal symbol manipulation
systems. So what could we be possibly missing when it comes to conversational
machines? The answer has been of course the notion of agency and situated
action. Computers can do something other than manipulate symbols; they can
implement arbitrary causal relations between sensors and actuators. Computers
might compute anything that is computable, but they can also implement the
decision processes of a thermostat. Behaviour-based robotics [
As an example of the phenomena of interest in CA, consider this classic example from the literature in a doctor/patient conversation:
Patient: this treatment, it won't have any e ect on us having kids will it?
Although it might seem reasonable to consider words to have meaning that can be looked up independent of context, the same is not true of silence and in the example that is certainly a meaningful silence. Whatever mechanism is at work here, it is also hard at work when we gure out the meaning of words.
Looking through the CA literature, human communication is is full of
normative behaviours { rules that can be broken, but to break them will be interpreted.
These rules are \behind the scenes" in that we do not consciously think about
them but we know they are there and shared by our \community of practice".
Making an apology is a complex process [
The idea that language use requires folk knowledge may be obvious but
the extent to which folk knowledge is core is perhaps borne out by the success
amateurs have in developing conversational agents for things like the Loebner
Prize [
In order to systematically analyse such phenomena CA has disavowed conjecture
about the mechanism or \rules in the head" that might have general applicability.
Instead the focus on what happens in particular instances of communication and
what observable behaviour contributes to choices made. The scienti c knowledge
is in how to study \folk" knowledge and, as with other ethnomethods, the point
is to capture the everyday knowledge that people use to do what they do.
Conversation Analysis provides a methodology but, having collected our butter ies,
engineers need generalizations in order to make something that can hold a
conversation. CA is strong on methodology and shy on theory, but Seedhouse [
B: First class or second class? Unless it is pointed out, people often do not notice that B's response is not an answer to the question. B's response can however be accounted for. Risks sanction If the utterance makes no sense, and the CP cannot gure out how it relates to what went before, then the CP will start working toward sanctioning the CP. It seems humans have a notion of fairness and feel justi ed in sanctioning the speaker if they think the speaker is not cooperating in the communication process. This is not a prescriptive rule taught to well brought up children; it is descriptive of what people do. The form of the sanction depends on many things and is highly culturally dependent. This is where notions of power and distance, roles and expectations comes into play.
This process of working through the seen-but-unnoticed, noticed-and-accountedfor, to sanction is not something we think about; it is just what we do and as such it is hard to notice in action. For instance it is quite surprising just how hard we are willing to work at accounting for a speaker's utterance. What is more, and in contrast to the views of many CA practitioners, it requires thinking about intentions. 2.2
We humans read o the intent of others { it may be di cult to recognize the
intent of Hitler invading Poland, but seeing two children tugging at a teddy bear
the human observer will be quite sure they both want it [
B: There aren't any other solutions.
You will try very hard to nd 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 oating 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.
This example demonstrates just how hard human communicators are willing to
work at recognizing intent in the speech acts of others. By contrast consider the
example in Figure 1 from Mann [
We humans rarely get to the stage of sanctioning but conversational agents are another matter. Consider the conversation in Figure 2 between a human and a chat bot. On a rst pass what stands out is the way the conversation simply PROGRAM: G'day mate.
JUDGE02: G'day. How's it going? PROGRAM: My mother once said "How is it going is a mystery." She was a bit nuts.
JUDGE02: Uh-huh.
PROGRAM: Are you still learning to speak, my friend?
JUDGE02: NO, no. But your anecdote took me a bit by surprise.
JUDGE02: How was your morning? PROGRAM: Oh well, that's a pity! (1)
JUDGE02: I'm sorry; what's a pity? PROGRAM: Apologies accepted.
JUDGE02: I feel compelled to insult you, and we've only been talking a minute. PROGRAM: We are bloody well talking, aren't we?
JUDGE02: Oh baby. We sure are. So here are some keywords for you: shit mother baseball.
... falls apart. One can point to the place where things start to go wrong, but for a considerable number of turns, the human at least is working toward repairing the interaction with apologies and warnings. At (1) the machine utters something that the judge cannot account for. The judge tries to get the machine to explain, which fails, resulting in an explicit threat of sanction. In the end the threatened \keywords" include swearing even though the judge is well aware that the world is watching.
A standard response to this example is to think we just need to ensure the machine does not say anything that cannot be accounted for. Notice however that four lines prior to (1) the human says something the machine (acts as if it) cannot account for and the human's response is quite di erent. The events are mirror images of each other but the human's handling of the situation is so automatic for us that it is hard to notice the symmetry. Human language use is situated action in an environment. Getting machines to do conversation, the context is just as much part of the process as the code and that context is full of highly socialized people.
When someone picks up a phone or runs a screen-based conversational agent,
they are already attending to (engaged with) the agent. Setting this up with
a physical agent is discussed elsewhere [
The Belief, Desire and Intentions (BDI) agent architectures [
Most BDI systems do not do planning but rather manage plans obtained from a xed plan library. What is more, it is not expected that there is planning \all the way down". Indeed plans in the library may contain sets of chatbot-like pattern-action rules that simply \produce behaviour" that an agent might have when it has the relevant goal. There may be several plans that might achieve a particular goal, and thus plan failure does not necessarily mean goal failure { there is a level of commitment to the goal that is not seen in many of the more traditional approaches involving planning. Critically for dialog systems, the goal is explicit and can be used in explanations of behaviour.
We have been developing a dialog scripting language based on XML that uses
a combination of features of Voice XML [
Looking at the knock-knock joke example, consider what happens when the goal tell a joke is posted. The system nds this plan, or another which also tells a joke, and executes it. Upon completion, if the goal is not removed, the next plan to tell a joke is found and so on. At some point the user will get sick of knock-knock jokes and can quit the program at any time by saying `good bye' which matches the trigger grammar of the sayBye plan. The trigger mechanism provides an elegant solution to interruptions that does not entail explicitly checking for no match conditions at every step.
Accounting-for is a crucial part of how humans communicate. To make machines do this requires some form of intention recognition, and this paper describes out simple approach to accounting-for based on trigger grammars. Using this approach, the Conversational Partner can not only interrupt the system while it is talking, he or she can interrupt the system's current intent. Our implementation is lacking in many, many ways, but the framework captures the idea of language as intentional and cooperative and is a basis for our goal of having machines do really natural language processing.