J.P. de Ruiter (Bielefeld University)
In my talk I present a general representational framework for modeling and understanding the phenomenon of miscommunication. For this project, it is necessary to first define what is meant by “communication”. It depends on one’s theoretical temperament whether this necessity is seen as a nuisance, a mission impossible, or an exciting theoretical challenge.
To prevent philosophical hair-splitting that would be outside of the scope of this workshop, I will concentrate on communication between a) intentional agents that b) is intentional, and c) involves an intentional agent as sender of a signal, and a possibly unknown number of intentional agents as receivers of that signal. Restriction a) exempts us from contemplating if the sound of a falling tree in an uninhabited wood is communicative or not, and b) if someone who sneezes is communicating. The focus here is on signals, not symptoms. In the words of Grice, we focus on the transmission of non-natural meaning. Restriction c) is a preamble for my attempt to reanimate the – in my view – underappreciated “conduit metaphor” (Reddy 1997), or “signal transmission approach” as in the famous communication model by Shannon (1948).
The conduit metaphor is out of fashion, because communication researchers have argued for decades that, among other arguments, a) the conduit metaphor is wrong because what is communicated is not a message, but rather the possibility to select between alternative possible messages (Reddy, 1997), b) communication is ‘holistic’, so senders and receivers cannot be conceptually separated (Clark, p.c.), and c) there is continuous feedback between senders and receivers, so communication is of a very complex and duplex nature (various authors, p.c.).
My counterarguments for nevertheless strongly defending the conduit metaphor as essential in communication research are the following. Re a): whatever is originating in senders and arriving at receivers (let’s call it X) may well have the function to enable receivers to select between possible messages, but that doesn’t change the fact that there is something called X somehow sent from A to B. Even though we could argue that X is, strictly speaking, information, and information is not a “thing” that can “transmitted”, the price we pay for avoiding the conduit metaphor is prohibitively high, as readers of Reddy’s alternative proposal will have discovered to their own peril. Re b): the point that we can’t separate senders and receivers, I’d like to quote a famous politician by observing: “yes we can”. And we should. Otherwise it would be impossible to model communication from the perspective of an individual agent, be it in a theoretical model of a human agent or in a processing model of an artificial agent. Additional arguments supporting a similar, shamelessly anti-holistic, position can be found in the excellent paper by Margaret Wilson (2002) on embodied cognition. As for c): there indeed is a lot of duplex communication between agents going on, but I contend that the atomic constituents of communication are still sequential, such that for every signal sent by a sender at time t, the receiver will process and respond to it at time t + ε, ε > 0. In face-to-face conversation, for instance, there are indeed many things happening overlapping in time, but for every individual signal, this sequentiality is always preserved. Therefore, although c) is indeed complicating things considerably, and presents a strong argument for incremental processing in communication, it is not a convincing argument against the conduit metaphor.
The problem with applying the conduit metaphor or Shannon’s model to intentionally communicating agents is not the metaphor, but rather what it is that is presumably communicated over the conduit. The signal that is sent should be such that it succeeds in transmitting meaning, and that’s where things get hairy. In Grice’s famous theory, the sender S communicates p to receiver R by sending signal x if S intends that
1. R comes to believe p,
2. R recognizes this intention (1), and
3. (1) occurs because of (2).
To ensure that this all works, senders have to carefully craft a signal, applying their knowledge about the receiver(s) to make it likely that 1-3 occur. This is not the same process as recipient/audience design, but it is a similar process. Instead of ensuring that references can be resolved, this process is aimed at making intention recognition possible for the recipient or audience (see e.g. Newman-Norlund et al., 2009; De Ruiter et al., 2010; Scott-Phillips, Kirby & Ritchie 2009).
But even if we buy into all that, and have it all modeled, the question that is still open is: what should be in p? Grice’s theory of meaning emphasizes the communication of the intention that receivers come to believe something, but not what they should come to believe. And it’s arguably not just some predicate p; social agents cannot sustain complex social relationships with each other by merely throwing propositions at each other, even if they are recognized as intentionally communicative.
Instead of just guessing what it is that receivers get out of senders’ signals, it is fruitful to study that question by looking at misunderstandings. If we know how, when, and why the communication system breaks down, we can find out more about how it is supposed to work properly. Schegloff’s classical work (e.g., Schegloff, 1987) on misunderstanding and repair reveals that the task-based dialogue experiments popular in psycholinguistics have a very serious restriction that can easily lead us astray. Schegloff identifies five types of misunderstandings, only one of which is related to reference. The other four (and there probably are more) are related to what in Schegloffese is called “problematic sequential implicativeness”, which corresponds quite well to the “underlying communicative intention”, or “illocution”. As Schegloff (p.c.) has often pointed out, task-based dialogue experiments will rarely, if ever, produce misunderstandings of that latter category, because if one thing is clear in a dialogue task, it is what the communicative intention of the participants’ utterances is; this has been stringently prescribed to them in the instruction of the dialogue task. So the only type of misunderstanding we get from tasks like the Map Task, are those related to “problematic reference”. In fact, the Map-Task and its variants are specifically designed to evoke reference problems, by deliberately giving unsuspecting participants a different, usually visually presented, believed-to-be-common ground.
So in order to properly model (mis)communication we need to assume that the signal contains information that enables receivers to decode not only the transmitted belief p, but also the communicative intention that underlies it. The principal claim here, one that I will defend and illustrate in my talk with examples is that every communicative signal produced “in the wild” (i.e., not in made-up examples) has an underlying communicative intention or illocution, and this communicative intention is essential in understanding signals. I therefore propose that the (implicit) representation of an intentional communicative signal has the following structure:
(I) An obligatory communicative intention L (e.g., threaten, warn, inform, ask, invite, insult, etc.)
(II) An optional predicate P to which L is linked. (E.g., I warn you that there is danger P)
(III) One or more arguments A1,…,AN occurring in the predicate P.
These three representational levels correspond to fundamentally different types of miscommunication. (I) corresponds to the perceived illocution, at a socio/psychological level, which if erroneous leads to the problems with sequential implicativeness as documented by Schegloff & colleagues. (II) corresponds to problems with grammatical (e.g. scope) ambiguities. This is the undisputed territory of classical (psycho)linguistics. Finally, (III) corresponds to problems in reference. It is at this particular level that the task-based dialogue research is relevant and informative, just as linguistics tests or sentence processing experiments are relevant and informative at (II), and qualitative (conversation analytic) work at (III).
Although there is lots of work to be done for fleshing out the details of this framework, it is implementable in principle and in practice (examples will be discussed in my talk) and it allows for a precise specification of the interface between communication (with or without language) and cognition (e.g., BDI based models). It also generates nontrivial predictions and provides a theoretically motivated perspective on the different varieties of the phenomenon of miscommunication.