Patrick Healey

Exploiting Under-specification as a Semantic Co-ordination Strategy

In the standard model of machine communication a set of states at a source (the message) is encoded in a signal that is transmitted to a receiver. The receiver then decodes the signal to reconstruct the message. This has two important weaknesses as a model of human communication (Cherry, 1966). First, the only kind of miscommunication it captures is 'noise' or uncertainty about what signal was transmitted. Mishearing is, however, only one of a rich variety of miscommunication phenomena evident in human interaction (e.g., Gumperz, 1996; Schegloff, 1987, 1992). Second, successful communication is possible only if the source and receiver possess identical copies of the code. This presupposes a level of linguistic homogeneity that is probably never found in any human population. The strength of the encoding-decoding model is that it provides a simple way to underwrite the possibility of successful communication. If we abandon the assumption of linguistic homogeneity (or the idealisation to it) then we need to develop an alternative account of communicative co-ordination that treats miscommunication as a ubiquitous, first class phenomenon. I will present two experiments on task-oriented dialogues, one involving exclusively graphical exchanges and one involving exclusively verbal exchanges. In both experiments patterns of interaction amongst participants are manipulated to create a number of equivalent `sub-communities'. Comparison of interactions within and across sub-communities shows that they develop community-specific 'sub-languages'. In these experiments participants have no explicit awareness that they are divided into sub-communities. In both the graphical and verbal exchanges, when participants communicate across sub-community boundaries they use a semantic strategy to cope with communication problems. Specifically, they reverse the convergence achieved in their sub-groups and shift to progressively less well-specified semantic models of the task domain. This strategy has the advantage of weakening the level of semantic co- ordination required to carry out the task at hand. I will also argue that it provides a basis on which co-ordination can be rebuilt that does not depend on a pre-existing shared basic code.

Pat Healey, Commentary on Pickering and Garrod "Toward a Mechanistic Psychology of Dialogue" Behavioral and Brain Sciences 2003: Dialogue in the Degenerate case? [PDF]

Paul Drew

Miscommunication? Turn design, accountability, repair and mis-alignment

My title includes four of the concepts in Conversation Analysis (which is the perspective within which I work) which seem to me to be most relevant to considering miscommunication. These concepts, or at least the first three (turn design, accountability and repair) are outlined in my overview of CA in a recent Handbook of Language and Social Interaction, edited by Fitch and Sanders. One aspect of turn design is that speakers can construct their turns so as to be understood by recipients precisely in the way they wish to be (accountability of conduct). Repair mechanisms can be mobilized when participants discover that there's been a 'fault' somewhere in the design/recipiency of a turn. So miscommunication (misunderstanding and the like) in CA investigations generally refers to problems, breakdowns in communication etc. which participants themselves recognise and orient to as such. This may be both a strength and limitation of our approach to miscommunication. The concept of mis-alignment stands slightly outside our usual requirement to show that, and how, phenomena are 'real' for (oriented to by) participants. It's a way we have, as analysts, to describe participants sliding past one another - engaging in what might seem to be a coherent spate of talk (sequence, whatever), but without having the same understanding of what's at issue, or not going in the same direction. I'll outline and illustrate how these concepts relate to miscommunication in ordinary conversation. This will include examples of:

I've added the question mark to Miscommunication? in the title partly as a way to highlight the issue of how we (as analysts) know/can tell that there has been a communication fault; and partly also to signal the matter of whether the mis- was 'designed' or 'deliberate'.

Drew, P. Conversational Analysis. Handbook of Language and Social Interaction, 4 (4), pages 71-102. [PDF]

Susan Brennan

Elements of Visual Co-Presence and their Impact on Miscommunication and Repair

Susan E. Brennam & Calion B. Lockridge
State University of New York at Stony Brook

Perceptual co-presence affects common ground and its use in conversation (Clark & Marshall, 1981). Two people in conversation may have varying degrees of perceptual co-presence. Auditory co-presence, especially the ability to hear one another speaking, is especially important for efficiently coordinating activity in real time (Chapanis, Ochsman, & Parrish, 1972), for conveying emphasis, and for avoiding and repairing misunderstandings; even brief temporal delays can be disruptive. The effects of visual co-presence, on the other hand, vary widely in their importance for collaboration, depending on the task and on which elements of the visual context are shared by collaborators. Previous research has found that some elements of visual co-presence (e.g., the ability to see what the other person is working on) are much more important to most collaborative tasks and to grounding in communication than others (e.g., the ability to see the other person's face; Whittaker, 1995; 2002). We present two studies that attempt to tease apart elements of visual co-presence that support achieving a joint focus of attention during referential communication, and consider how the presence or absence of visual evidence shapes miscommunication and repair.

Staffan Larsson

Semantic plasticity through feedback and accommmodation in dialogue

Dept. of linguistics
Gothenberg University, Sweden

I use the term "semantic plasticity" to refer to gradual change and adaptation of meanings of words, phrases and other linguistic constructs. Apart from the well-known phenomenon of large-scale semantic change over time (as studied in historical linguistics as part of the general phenomenon of linguistic change), semantic plasticity occurs in single dialogues, between specific dyads or communities of speakers, and in the adaptation of linguistic resources to specific activities.

I outline a basic abstract formal framework for describing how meanings change as a result of concrete language use, especially in spoken dialogue. The basic idea is that speakers have internalised (potentially complex) dispositions for when and how to use specific linguistic constructs. These dispositions depend, among other things, on observations of previous situations where the construct in question has been used, and on specific generalisations over these situations.

In dialogue, linguistic behaviour of speakers in a linguistic community becomes coordinated when speakers adapt their usage dispositions to each other. This adaptation is possible since usage dispositions are plastic. Two central processes in dialogue that enable this interactive adaptation are feedback and "meaning accommodation". Feedback concerns how reactions to previous usage affects future usage. "Meaning accommodation" concerns how hearers can (quietly) adapt to observed usage.

We may take miscommunication as standing in opposition to perfect communication where perception, understanding and agreement are unproblematic, and hence no feedback or accommodation is needed. On this view, miscommunication is the very basis for linguistic plasticity and adaptation; if there was no miscommunication, language would be static and unadaptable (unless changes where instantaneous and universal, which is not plausible).

Semantic plasticity and linguistic change

Taking inspiration from Kripke's (1982) interpretation of Wittgenstein (1953), I have sketched an account of semantic plasticity in terms of updates to individual usage-dispositions associated with linguistic constructs (words, phrases, syntactic categories etc.) triggered by feedback and accommodation in dialogue. The social counterpart to usage-dispositions we may call "usability conditions". Briefly put, usability conditions are (often implicit) social conventions that emerge from the coordination of individual usage dispositions in interaction. The usage disposition associated with a linguistic construct is defined as a function of (among other things) the situations where that construct has previously been used. However, this account is not limited to semantics since usability conditions also involve syntactic and other linguistic aspects of the situation.

Semantic plasticity/change is thus related to linguistic change in general insofar as they can all be regarded as special cases of changes in usability conditions. While semantic plasticity might involve e.g. available shared referents in the utterance situation, and phonetic plasticity involves structural properties of the phoneme system being used, the aspects of the situation relevant to syntactic change might be something like structural relations between linguistic constructs being uttered in the same sentence, or in previous sentences. For example, the usability conditions on a verb phrase might require that the sentence also includes a noun phrase, or that one can be found in the shared dialogue gameboard (e.g. in a stack of Questions Under Discussion).

Semantic plasticity occurs globally as well as locally, in the adaptation of linguistic resources to specific activities. Regarding the relation between "plasticity" and "change", perhaps "change" implies more of a global and historical perspective involving long-term dynamics whereas plasticity is perhaps primarily geared towards more local (e.g. dyadic) short-term changes. An account of semantic and linguistic plasticity should generalise to both local short-term and global long-term effects. One question that arises is whether syntactic plasticity is a useful concept - are syntactic rules ever adapted to local/specific situations, or are there only global syntactic changes? While the latter assumption might perhaps seem easier to accept, it is also reasonable to assume that all global changes start as local changes and then spread successively over the population of speakers.

Luc Steels

Using Re-entrance for Repairing Miscommunication.

Vrije Universiteit Brussel (AI Lab)
Sony Computer Science Laboratory - Paris

There is now general agreement that partners in dialog align their communication systems at all levels: phonetic, lexical, grammatical, conceptual, pragmatic. The question I will address is how we can make mechanistic models of this alignment process, not just short-term local optimisations (for example based on priming) but also deeper adaptations and innovations that may possibly give rise to long-term change. Mechanistic models require that we understand what triggers partners to change their communication system, and then how partners figure out what to change and how. Miscommunication clearly plays a very important role in the first aspect, namely triggering change and subsequent dialog gives crucial information to both partners how they may repair the miscommunication.

I will argue that simulating the understanding process of the other dialog partner (speaker or hearer) is an important part of repair, particularly in the case of grammar. The speaker can simulate the hearer by re-entering the sentence back into his own language system and then re-parse or interpret to detect possible failures. The hearer can do this by entering partial, predicted or derived meanings into his own language production system to figure out the misalignment. Some examples of repair strategies that exploit these forms of re-entrace will be discussed within the context of experiments with embodied agents.

Steels, L. (2003) Evolving grounded communication for robots. Trends in Cognitive Science, 7 (7), pages 308-312, July 2003. [PDF]

Jonathan Ginzburg

(Mis)Communication: grammar and evolution

In linguistically based communicative interaction among adult humans there are two basic cases: (a) understanding---typically signalled by providing the expected reaction (acceptance of assertion/command, answer to question etc) (b) incomplete understanding---which leads to clarification interaction.

Consequently, any account of successful communication using utterance type S needs to be able to simultaneously characterize why and how a particular range of possible clarification requests can occur following uses of S :

(1)a. Did Bo kowtow?
b.Bo? / Your cousin?/ Did WHO kowtow? / Bo Jackson or Bo Didly? / kowtow? / What do you mean kowtow? Why? . . .
In this talk I will sketch a model of dialogue interaction that offers a characterization of these two aspects of utterance (mis)understanding, which brings out inter alia the fact that content conveyed by grammatically provided material has a distinct status from implicitly conveyed content (so this approach is, in a sense, a radically semantic approach.). This model emerged in recent years in collaborations with Matt Purver and Robin Cooper; Purver has implemented a prototype dialogue system CLARIE that demonstrates the main features of the approach

I will also consider how this system of metacommunicative interaction (MCI) has evolved. I will sketch a number of simpler MCI systems that allow us to speculate about cutoff points between adult humans, kids, and other interacting organisms. I will present some recent results of an Artificial Life simulation (done jointly with Zoran Macura) in which the lexicon dynamics of populations that possess and lack MCI capabilities are compared. We ran a series of experiments whose initial state involved agents possessing distinct lexicons and whose end state was one in which all agents associated meanings with each word in a lexicon. The main effect demonstrated, one we dub the Babel effect, is that the convergence rate of population that relies exclusively on introspection is intrinsically bounded and, moreover, this bound decreases with an increasing population. This bound seems to disappear once agents are endowed with clarification requests.

Ginzburg J. and Cooper R, Clarification, Ellipsis, and the Nature of Contextual Updates in Dialogue [PDF]
Ginzburg J. and Macura Z., Lexical Acquisition with and without Metacommunication [PDF]

Ruth Kempson

Production Pressures, Routinisation and Syntactic Change

This paper takes a well-known syntactic puzzle and argues that its explanation lies in the effect over time of production pressures in dialogue.

The case study is the shift from Latin through Medieval Spanish into Renaissance Spanish..

What I shall argue is that in the Latin starting point, argument expressions whose value is to be selected from context tend to be placed at the left periphery in constructing any new propositional domain (reflecting the constraint of minimising production costs). From there, routinisation effects - themselves a means of minimising production costs (one look up for a sequence of actions) - led to weak pronominal forms (clitics) being associated with a macro of actions immediately following the identification of a new propositional domain in the emergent logical structure.

The result is the second position clitics of Medieval Spanish, otherwise analysed as subject to a heterogeneous list of structural restrictions. Subsequent steps of routinisation led to conflation of actions of pronoun+verb, collapse of restrictions on pre-verbal pronouns and calcification into modern pre-verbal clitic pronouns.

The interest of the account lies in the spontaneous emergence of syntactic phenomena from general dialogue pressures without any necessary mismatch of speaker/hearer context or content choice.

Amanda Stent

Local and Global Adaptation in Hyperarticulation

In this talk I report on a Wizard-of-Oz experiment designed to examine the relationship between hyperarticulation in human speech directed at computers and the misrecognition of such speech. In the experiment, the location and type of speech recognition "errors" were controlled, resulted in a corpus of 1300 utterances which were coded by hand for phonetic features associated with clear and relaxed speech, and separately were run through a set of speech recognizers. Findings include:

I then describe a follow-on experiment looking at how system prompts and responses can be used to guide speakers away from maladaptive speech.

Rosemarie McCabe

Communication breakdown in psychiatric consultations with schizophrenia patients: What goes wrong and why?

I will be talking about miscommunication in a specific, applied context, i.e., when talking about psychotic symptoms leads to a breakdown in communication between patients with a diagnosis of schizophrenia and their psychiatrists in outpatient consultations.

In mental health care, specific treatment effects, e.g., of medication, explain only a small part of the variance in treatment outcome (Roberts et al., 1993). A larger proportion of outcome is due to non-specific or placebo effects of treatment. It may be assumed that most of the relevant non-specific processes are constructed in interactions between patients and clinicians. Indeed, research shows that a positive therapeutic relationship is a powerful predictor of treatment adherence and more favourable outcome (Catty, 2004; McCabe & Priebe, 2004). However, little research has been conducted on what kinds of interactions promote positive relationships and a better outcome for the patient.

Identifying such interactions is particularly important in the treatment of people with schizophrenia. Of all patients with mental illness, they are the most difficult to engage in services and the most likely to lose contact with clinicians and 'slip through the net' with serious consequences (e.g., homelessness, imprisonment, death by suicide).

Successful psychiatrist-patient communication rests on mutual understanding about the nature of the patient's problem and its treatment. Communication between psychiatrists and patients with schizophrenia is problematic because there is not a shared understanding about the nature of the problem. Patients typically do not agree that they have an illness and do not accept the clinical explanation of their symptoms. They experience psychotic symptoms (hallucinations, delusions) as real. From a clinical perspective, these symptoms are considered to be 'un-understandable' and are explained clinically as part of a mental illness (in the patient's mind). Despite repeated attempts by clinicians to get patients to re-attribute their symptoms, however, patients persist in their explanation of their symptoms and do not modify their beliefs.

We have used conversation analytic (CA) techniques to analyse naturally occurring psychiatrist-patient consultations to explore whether this 'un-understandability' reveals itself and if so, how. The main premise of CA is that participants in talk display their understandings of the talk for one another on a turn-by-turn basis. Participants offer each other opportunities to contribute to the making of agreed meanings in a back-and-forth process of negotiation. Hence, each turn in a conversation provides a continuously updated display of intersubjective understanding.

In these consultations:

These findings pinpoint how the intelligibility of the patient's experience is a problem for others (Hinshelwood, 1999), something that patients themselves cite as problematic in their experience of services (Pinfold & Corry, 2003). That the meaning of patients' symptoms and their treatment is regularly disputed between clinicians and patients, leaving patients feeling ill-understood, may well be linked to the failure of services to successfully engage this group of patients (cf. Tyrer 2000). I will present data to illustrate the process that leads to a communication breakdown between the psychiatrist and patient in an attempt to identify the necessary conditions for establishing mutual intelligibility and what leads to its failure in this context.

Communicating about psychotic symptoms in the medical consultation. The Journal of Primary Care Mental Health Volume 7, Number1, 2003. [PDF]

Herbert Clark

Forestalling misunderstandings in spontaneous talk

Standard models of speaking assume that people first plan their utterances and then produce them. They begin with determinate but incomplete plans, and fill out these plans as they execute their speech. In reality, people's plans are usually not just incomplete, but indeterminate. People recognize that circumstances may change, not only mid-conversation, but mid-utterance, so they may need to revise or cancel what they have already said or alter what they were about to say. People monitor three sources of evidence about changing circumstances: their own mental states, their partner's actions, and outside events. They monitor both for unforeseen obstructions to their current plans and for new opportunities they might exploit. The result is a variety of strategies both solo and interactive for forestalling misunderstandings. One class is retrospective, revising speech already produced. A second class is prospective, warning of problems to come. A third class is opportunistic, taking advantage of new circumstances that allow shortcuts or improved plans. People in conversation, in brief, recognize that plans aren't determined ahead of time, but evolve on the fly.

Herbert H. Clark and Meredyth A. Krych, Speaking while monitoring addressees for understanding. [PDF]

Pragmatics of Language Performance. Handbook of Pragmatics, 16 (16), pages 365-382.[PDF]