Announcement for Workshop 4

"Grounding and CRification: theory and emergence" - Jonathan Ginzburg, KCL

In this talk I will present the basics of a theory of the communicative process which involve balancing successful grounding with the cases in which clarification requests get posed, a process I dub instead CRification. I will also make some speculative comments on how such a system emerges, based on data arising from joint work with Dimitra Kolliakou.

Matthew Stone

"Incremental reasoning about constraints on reference" - Matthew Stone, Rutgers

As speakers flesh out incomplete utterances, the descriptive material they add gradually picks out the discourse entities they are talking about. Conversely, as hearers process incoming utterances, the descriptive material they get provides moment-to-moment evidence of speakers' references. Of course, in natural conversation, the incremental work of identifying entities is often a collaboration between speaker and hearer; clarification is a case in point. To model such interactions, we need to characterize collaborative conversation simultaneously in terms of the the linguistic representations behind interlocutors' utterances and the social activity that makes those utterances coherent.

This talk describes an implemented system for collaborative reference that combines incremental understanding and incremental generation with the ability to ask and answer clarification questions through incremental contributions to an extended activity. In our model, collaboration on referring builds on the representations used for processing referring expressions, particularly utterance structure and associated ambiguities in interpretation. Even so, the implementation distinguishes individual reference from collaborative reference, for example through a difference in granularity between words, the units of linguistic knowledge for constructing referring expressions, and task moves, the units of social knowledge in collaborative reference. Our architecture therefore separates dialogue management, which chooses task moves, from language generation, which chooses words, even though both can reason incrementally about constraints on reference. How to connect these processes together effectively, as human speakers evidently do, looms as a central problem for future research.

(Joint work with David DeVault with input from Natalia Kariaeva, Anubha Kothari, Iris Oved, and Rich Thomason.)

Stanley Peters

"Clarification Responses" - Stanley Peters,

This talk will examine the effect that a general clarificational capability can have on human-computer dialogue. We will first describe the CHAT system, an application of an information-state-based dialogue manager to automotive systems. CHAT provides a driver with a conversational interface which allows natural two-way interaction with various in-car devices, currently including the audio system, the navigation system, the phone & address-book, and an information database. An important part of the system is its ability to employ clarificational dialogue to detect and correct understanding errors of various types, while allowing the conversation to proceed. We will describe the particular clarification strategies it can employ, together with the basic context-update mechanisms which underly them. We will then examine the responses of users to these strategies, and contrast them with the less natural, more disjointed dialogue that results from the more simple-minded error-handling common in human-computer dialogue systems.

Matthew Purver

"Clarification Requirements" - Matthew Purver, Stanford

This talk will examine dialogue data gathered from corpus analysis, both human-human and human-computer, and discuss the requirements which this imposes on theoretical and practical approaches to dialogue processing. We will first examine what can be clarified, which will allow us to express some requirements on the the understanding process and the representations it uses. We will then turn to \emph{when} it can be clarified, and draw some conclusions about the nature of incremental syntactic and semantic processing. Finally we will see some evidence about who can employ clarification -- contrary to the unspoken assumption made in most human-computer dialogue systems, we need to be prepared for users to ask systems for clarification -- and discuss what this means for system designers.

Matthew Saxton

"The Prompt Hypothesis: Clarification Questions as Corrective Input for Grammatical Errors" - Matthew Saxton, IoE, London

The potential of clarification questions (CQs) to act as a form of corrective input for young children's grammatical errors was examined. Corrective responses were operationalized as those occasions when child speech shifted from erroneous to correct (E → C) contingent on a clarification question. It was predicted that E → C sequences would prevail over shifts in the opposite direction (C → E), as can occur in the case of non-error-contingent CQs. This prediction was tested via a standard intervention paradigm, whereby every 60 seconds a sequence of two clarification requests (either specific or general) was introduced into conversation with a total of 45 2- and 4-year-old children. For 10 categories of grammatical structure, E → C sequences predominated over their C → E counterparts, with levels of E → C shifts increasing after two clarification questions. Children were also more reluctant to repeat erroneous forms than their correct counterparts, following the intervention of CQs. The findings provide support for Saxton's (2000) Prompt hypothesis, which predicts that error-contingent CQs bear the potential to cue recall of previously acquired grammatical forms.

Susan E. Brennan, Marie K. Huffman, Tanya Kraljic, & Stephanie Hannigan

"Adventures with Dialects: Convergence and Ambiguity Resolution by Conversational Partners" - Susan E. Brennan, State University of New York at Stony Brook

Speech from individuals who speak the same language is highly variable; some of this variation is idiosyncratic, and some, due to dialects. We see a dialect as a bundle of features with a high probability of co-occurrence. Some of these features, speakers and addressees are aware of, and others, not. I will describe two of our studies of adaptive language use: One focuses on speakers’ partner-specific convergence in referential communication, and the other, on listeners’ ability to use their knowledge of a speaker’s dialect to resolve words that are homophonic in one dialect but not in another. These preliminary findings enable us to lay out a set of experimental questions about the production and comprehension of dialectal variation.

Robin Cooper

"Clarification and incremental meaning/content refinement"- Robin Cooper, Göteborg University

This paper will examine the consequences of simple clarification exchanges such as

A: Bo left

B: Bo?

A: Bo Ralph

B: Oh

for notions of meaning and content in semantics.

B's first utterance would not make sense under the simplifying assumption adopted, for example, in classical Montague semantics that proper names correspond to logical constants. Instead we need something like the situation semantics treatment of proper names as restricted parameters.

B's first utterance also suggests that the notion "content of A's first utterance" is not well-defined since at this point in the dialogue A's content for the utterance and B's content for the utterance are distinct. Perhaps, as was suggested in the early situation semantics literature, the important notion to study is that of information which an agent can obtain from an utterance rather than meaning or content tout court.

In earlier work Ginzburg and I have argued that clarification exchanges like this motivate the analysis of meaning and content as semantic objects with greater structure than, for example, the classical analysis of meaning as functions from contexts to sets of possible worlds. You need at least to be able to recover the interpretation of constituents in an utterance.

Previously, I have thought that the Montague/Kaplan notion of meanings as functions from contexts to contents should be used in modelling meanings as functions from records (contexts) to record types (propositional contents). However, this approach involves the manipulation of these meaning functions to create new meaning functions depending on how much information is available in the context. In this paper I want to suggest that we can use record types to represent meanings more elegantly if we take a more unification-like approach to the combination of meaning and context which is related to the use of context parameters in HPSG. It allows the meaning/content of an utterance for a particular dialogue participant to be specified incrementally during the course of a dialogue. While I believe that this approach maintains the basic insight of the classical Montague/Kaplan functional approach to meaning, it calls into question the rigid division between meaning and content found there and perhaps therefore corresponds to more intuitive notions of meaning/content.

David Schlangen and Raquel Fernandez

"Speaking through a noisy channel-experiments on induced communication problems" - David Schlangen & Raquel Fernandez, Potsdam University

We will talk about work done within the DEAWU project (DEAling with Uncertainty; see http://www.ling.uni-potsdam.de/DEAWU ). We are interested in learning more about the strategies human speakers use to deal with understanding problems and ensure (good enough) intersubjectivity; our mid-term goal being to use what we learn to devise strategies that spoken dialogue systems can use. In previous work we have provided a classification for different types of clarification requests and have conducted a corpus study. Our current work is a continuation of this. Instead of analysing a corpus of "normal" conversation, we are creating a communication situation in which one channel is disturbed by random bits of noise. In the analysis, we are interested both in the immediate reactions to the acoustic problems and in the more global strategies that speakers choose to limit the damage of the defective channel.

Patrick G.T. Healey and Gregory J. Mills

"Experimenting with Dialogue: Miscommunication as a resource for co-ordination" - Patrick G.T. Healey & Gregory J. Mills, QMUL,London

The EPSRC funded DiET project ( http://www.dcs.qmul.ac.uk/research/imc/diet/ ) is developing a new tool for experimental studies of dialogue. It builds on two developments. First, the increasing use of text as a basic mode of interaction (viz. instant messaging, sms, chat rooms). Second, the development of natural language processing techniques that can process text turns in real time. This makes it possible to make fine-grained, context sensitive, interventions in natural text-based dialogue without overt disruption to the flow of the conversation. We will illustrate the potential of this approach with two 'maze task' experiments:

1. The introduction of 'spoof' probe questions concerning maze locations. The experimental contrast is between probe questions that appear to originate from a primary participant and those that appear to originate from a peripheral participant. The results show that the same question, in exactly the same dialogue context receives systematically different answers depending on it's apparent origin. We argue that this finding challenges accounts of dialogue co- ordination based on priming and suggests that participants must directly engage in detecting and resolving misunderstanding in order to co-ordinate.

2. The effects of spoof clarification questions, that appear to originate from the other participant, on coordination. We contrast relatively low specificity spoof questions ('what?' 'Eh?') with high specificity reprise fragments of single words from preceding turns. The results show that clarification questions have a significant impact on co-ordination and that this impact varies with the relative specificity of the clarification.

We argue that the combined results of these studies show that miscommunication or `misalignment' is a key resource for the development of semantic co-ordination. In particular, people must be able to detect and resolve problems with mutual-intelligibility before higher levels of co-ordination can develop and be sustained.

Hannes Rieser

"Fragments in Multi-modal Dialogue" - Hannes Rieser, Universität Bielefeld

In the Bielefeld spoken language corpora communication with fragments or non-sentential utterances (Jonathan Ginzburg’s NSUs) is the rule rather than the exception. This will be shown using a dialogue which comes from an experiment based on object identification games (OIGs). OIGs have two agents, the description giver (DG) and the object identifier (OI). The DG describes or names an object in the demonstration domain using a (multi-modal) dialogue move; the object identifier (OIG) tries to identify the object named. The dialogues resulting from the OIG experimental setting have the following prototypical structure: directive by the DG, verbal reaction by the OI and final verbal reaction of the DG. The dialogue discussed shows fragments at structurally central places: in the DG’s first-turn directive, in the OI’s clarification request of the second turn, and in the final acknowledgement of the DG. In this context, interesting regularities of German case distribution across turns are also discussed. Fragment resolution is modelled using so-called situated conventions based on situation types. Situated conventions are recipes for agent’s verbal actions. This procedure indicates that the whole approach is on the pragmatic side. Some effort is invested in explaining the notions “situated convention” and “situation type”. The working of situated conventions is modelled using SDRT: they are introduced into SDRT’s cognitive modelling section. The phenomena treated have at times been of interest to several of the workshop participants working on dialogue. A further topic taken up is the meaning of the pointing and the grasping gestures and their interaction with the meaning of the fragments. This interaction is also modelled using SDRT as an interface

Staffan Larsson

"Coordination in the

GoDiS dialogue system

" - Staffan Larsson, Göteborg University

In its current version, the dialogue system GoDiS deals with informational coordination ("grounding") using various interactive communication management devices such as feedback, question accommodation, and clarification questions. I will explain the current capabilities of the system, and discuss various possible future developments and associated problems. These include extended feedback capabilities, feedback in an asynchronous ("incremental") dialogue system architecture, as well as semantic coordination.