Everybody is involved many times a day in conversations. Sometimes when we stop talking and begin to listen, we can hear the incredible noise made by other speaking humans. We spend much time in this linguistic activity which seems to be a distinctive characteristic of our species. So it is quite surprising to observe that spontaneous conversations are the object of a limited number of studies compared to the huge number of papers devoted to other aspects of natural language (e.g. syntax or semantics). Indeed several authors consider that most models of "higher" aspects of language are still unsatisfactory (e.g. [Trognon & Brassac 1988:212]). H. Bunt [1991] puts it this way :
"dialogue theories need other devices than, say, constituent structure diagrams and truth-conditional semantic rules. Instruments are needed to model such things as what each of the partners knows and believes and what communicative intentions they pursue. The representation of such things and how they can be used by an intelligent linguistic agent to perform successfully in a dialogue is far beyond the means of traditional linguistic theories".
Are all aspects of conversations so complex that they appear totally unconstrained ? The aim of this paper is to suggest that we may find some reliable structure if we look at the content of arguments. Some scientists do study higher aspects of natural language without making use of any corpus, e.g. Discourse Analysis (see [Coulthard 1977]) and many studies using the paradigm of Speech Acts [Searle 1969]. However for our purpose, such a limitation would be an obstacle. Our approach is closer to the Conversation Analysis method for at least one reason : we consider unconstrained conversation as a natural behavior that has to be studied as it occurs, and thus we decide to work only on data gathered on real situations. All the excerpts given in this paper come from conversations that really occurred (for discussions about differences between these approaches, see [Reichman 1989], [Goodwin & Heritage 1990], [Moeschler 1990], [Norman & Thomas 1991]).
Conversations can be studied from very different points of view. Many authors are concerned with more sociological aspects : rules governing turn taking (how we know if we can or even must utter a reply at a given moment), conversational style (describing recurrent differences between speakers), how social relations are expressed in conversation (e.g. dominance, aggressiveness, politeness, cooperation), what kind of social acts are performed during conversation, etc. Other authors are more concerned with micro-features of conversations: interruptions, intonation, hesitations, etc. [Clark 1995]. Some others study relationships between structure and function in conversations : how explanations, complaints, etc. are expressed through utterances, how decisions or truth are negotiated in interaction, and so on. For instance Drew & Holt [1987] show how idiomatic expressions are used under certain circumstances to express and summarize complaint during conversation ; Reichman [1985] tries to identify " a conversation 'deep structure' in terms of structural relations between discourse elements " ; Heritage [1990] shows how explanations are expressed depending on the social situation.
Our approach is closer to this last trend. In the present study, we are looking at the content of utterances, trying to describe how they are functionally linked together and with the topic. On one side, we are not describing social aspects of conversations. We do not even consider the intentions of interlocutors. Our concern here is, as A. Trognon puts it, conversations, not people involved in conversation. On the other side, we are considering the content of arguments, not wording details. The model we will present will allow us to make some predictions about what an utterance can or cannot express at a given moment of a conversation.
But what do we consider here as a conversation ? An extreme position may be to consider that every naturally occurring use of language has to be considered as a conversation. Our aim is of course much more limited : our results come from the observation of typical casual conversations, with the following characteristics :
- speakers with equal status
- relaxed speakers
- speakers knowing each other well
- speakers meeting frequently
- speakers not involved in a task-oriented conversation
- speakers being (at least momentarily) serious
The extension of our model to other types of interactions is not straightforward. Suprisingly, it appears that casual conversations seem to be more constrained than less relaxed dialogues (cf. [Dessalles 1992a]), like a quarrel between wife and husband (as in [Schank & Lehnert 1979]), or work dialogues (such as operative language or dialogues occurring during design activities, studied for instance by Falzon [1991]). In contrast, our own data consist mainly of more than 30 hours of family conversations involving most often more than three adult speakers. This approach to conversation claims to be "ecological", since it starts from the unprejudiced observation of human beings behaving normally. Some of the recordings were made "secretly", but most of the time the recorder was placed on the table and quickly forgotten, even by the observer. It was simply presented in fact as a way of taking a "sound photo".
I was always present and often participated in these recorded situations. One may be surprised by the fact that I did not keep my distance from the data. But this was necessary for at least two reasons : first it is understandable from an ethical point of view, and secondly, as we shall see, it is the best way to have a perfect knowledge of the context, and it was thus the only way we found to gain some insight into some of the rules governing conversations. But we will also use here data from other corpora whenever it is possible.
This last point which concerns the context must be stressed : our study of logical aspects of argumentation in spontaneous conversations begins as soon as the context is known. Having currently no way to extract relevant contextual knowledge, we study only conversations for which this context is unambiguously known.