Linguistic markers in coming to understanding

Several recent research projects in artificial intelligence are devoted to the development of natural, multi-modal user-machine interfaces. Most of the work in this field share, more or less explicitly, the assumption that artificial agents who are competent in multi-modal communication should possess both context awareness and meta-cognitive abilities such as being able to represent and make inferences about beliefs, intentions, and desires of their human interlocutors. In this paper the problem of identifying the complex relationships between context awareness and (meta-) cognitive abilities is approached from the point of view of the nature of the understanding process in conversation. On the basis of the assumption that this nature is inherently dynamic and dialogic, we will focus on the role played by some linguistic and dialogic markers. These devices are commonly resorted by interlocutors, in order to perform their interactional tasks and in order to signal each other the steps of the understanding process. The paper highlights the cognitive, conversational, and interactional functions of these markers by showing, on the basis of corpus-based analysis (both human and human-computer interactions), how they contribute to the process of achieving a mutual understanding by passing through several stages of a continuum, possibly including both failures and repairs in communication. The paper also claims that natural user interface to multi-modal systems should be able to deal with a wider range of linguistic markers. 1. The dynamic and dialogic nature of understanding in conversation Current research on multi-modal user interfaces has been establishing severe requirements on the naturalness of intelligent, task-oriented user interfaces. In this field, some projects are based on theories of collaborative discourse, while in other projects, the issues in devising complex, intelligent, and multi-modal user interfaces are approached by exploring solutions that may be scalable depending on the dialogue task complexity (see Allen et al. 2000). An interesting application field is that one of intelligent tutoring systems enhanced with discourse capabilities: actually, tutoring systems such AutoTutor and Atlas, compared with interactive but non-natural language systems have shown that conversational dialogue substantially improves learning. On the basis of comparative discussion of these different tutoring systems, we may foreseen that a generic dialogue style might be unsatisfactory for applications to different knowledge classes, and that there is an increasing need for personalized dialogue styles. In order to satisfy this need within full-scale application-independent frameworks, we believe that it is crucial to get a deeper understanding of how the dynamic nature of conversation is affected by the role of linguistic markers. In this paper, we will analyze linguistic markers such as discourse markers and dialogic repetition that are commonly resorted to by interlocutors, in order to signal each other the steps of the understanding process, among other fulfilled functions. The paper will show the cognitive, conversational, and interactional functions of these markers by discussing, on the basis of corpus-based analysis (both human, and human-computer interactions), how they 1 The authors wish to thank Alberto Baracco for fruitful discussions on the present paper, and for his analysis of part of the IPAR corpus with respect to linguistic discourse markers. Correspondence can be addressed to [email protected] 2 For a review of intelligent interfaces in a variety of multi-modal interactive systems the reader can refer to the Special Issue of the AI Magazine on Intelligent User Interfaces, vol. 22 n.4, Winter 2001. contribute to the process of achieving a mutual understanding by passing through several stages of a continuum, including possible failures in communication. Actually, as Yus Ramos (1999,217-218) says "Communication is not an easy task. Every day, people engage in conversations of many kinds with different purposes and, on every single occasion, addressees have to rely on their ability to overcome inherent obstacles that can arise in face-to-face interaction in order to arrive at the interpretation that their interlocutors intend them to process. [...] Instead of straightforward, machine-like decoding of utterances, it is now admitted that interpretation is a matter of raising hypotheses about which proposition the addresser intends to communicate, and success in reaching the correct interpretation is constrained (or helped) by a whole range of contextual features, some of which are difficult to pin down." Two main notions characterize the implicit search for mutual understanding during everyday conversation: its dynamic nature and its 'dialogicity'/intersubjectivity. In the remnants of this section we will explore the basic notions of understanding as a process, and the relationship between dialogicity and understanding. 1.1 Understanding as a process Understanding, both as a cognitive and interactive social notion, is a process which involves intermediate stages, and which evolves in the interactional development. In contrast to the traditional dichotomic view, where understanding was opposed to nonunderstanding, the notion of coming to understanding has been proposed by Weigand 1999, specifically with reference to dialogue activities: "Dialogic action games are always interactive action games negotiating on an interactive level the general purpose of coming to understanding." Misunderstanding itself has been seen as a possible step in the process of understanding, and a "negotiation cycle" of misunderstanding has been recently proposed (see Bazzanella and Damiano 1999,823-4), where the dynamic features of the process are stressed. As Dascal writes in his introduction to the special issue of Journal of Pragmatics on misunderstanding: " [...] in so far as misunderstanding is subject to 'negotiation' as part of an ongoing process of 'coming to an understanding' [...], it would seem rather to be viewed as a continuum. [...] The gradation view is further supported by the fact that, from a logical point of view, misunderstanding and understanding are contraries, rather than contradictory, i.e., the propositions 'A misunderstands B's utterance’ and 'A understands B's utterance' may be both simultaneously true (regarding the same utterance)." Dascal (1999,756). Misunderstanding in human-human interaction is often neglected: “in fact [...] during the process of interpersonal communication, participants tolerate a rather high degree of non-acknowledged, unresolved potential misunderstanding." Blum-Kulka e Weizman, (1988,220). Unlike human-human interaction, human-computer interaction differs in flexibility: i.e. it has less skilled in recognizing intentions, and in understanding via inference. However, interlocutors’ imprecision and/or errors require that interlocutors’ contributions are to be taken into maximal account, on the basis of Davidson's Charity Principle. In the early Nineties, the development of 3 The same ranking of “to understand” as a verb ("Understand is to be ranked as an achievement verb, and know as a state verb." Vendler (1994,14)) underlines its dynamic nature. 4 A particular case is misunderstanding in CMC, see Bazzanella and Baracco forth. 5 It is noteworthy that "[...] in situations of deceit and harm, what two participants infer that the other person knows and what they want the other person to know varies markedly from situations in which accord and agreement are present." Stein and Bernas (1997,259). spoken dialogue systems cannot avoid to deal with misunderstanding: in human-computer spoken dialogue break-downs in communication are more likely to happen because the understanding process of artificial agents is often a «yes-no process», where break-downs are more common (see Reilly 1987). In addition, artificial agents are usually not able to detect the occurrence of a misunderstanding. Most miscommunication phenomena can only be recognized by humans, however the system has to be able to detect humans’ detection of miscommunication (see Danieli 1996). 1.2 ‘Dialogicity’ and understanding Convergence in understanding, in its difficult and essential intersubjectivity, was already neatly expressed by Humboldt in his Eileitungen: "[..] Nobody means by word precisely and exactly what his neighbour does, and the difference, be it ever so small, vibrates, like a ripple in water, throughout the entire language. Thus all understanding is always at the same time a not-understanding, all concurrence in thought and feeling at the same time a divergence." Humboldt (1999, 63). Understanding in conversation inherits, in a sense, ‘dialogicity’ from the conversation itself. Conversation involves two or more participants, and cannot escape its dialogic or intersubjective ‘nature’, which has been rightly stressed in ethnomethodology and in conversation analysis. Understanding in conversation concerns mainly two aspects: structure and information. On the structural level, it is an easily noticed fact about two-party conversations that their sequence is alternating; that is to say, conversational sequence can be described by the formula ababab, ('a' and 'b' being the parties to the conversation). As Heritage and Atkinson (1984,11) maintain: "[...] In examining talk the analyst is immediately confronted with an organization which is implemented on a turn-by-turn basis, and through which a context of publicly displayed and continuously updated intersubjective understanding is systematically sustained. It is through this turnby-turn character of talk that the participants display their understandings of the state of the talk for one another, and because these understandings are publicly produced, they are available for analytic treatment by social scientists. " On the information level, if we consider understanding in the interactional development, it unavoidably includes both the participants, so much so that the mutuality of understanding is continually construed and reinforced in interaction: it works as a common ground which allows the flow of conversation to go on. In Schegloff (1992, 1299)’s words: "The defense of intersubjectivity is locally managed, locally adapted, and recipient designed.[...it ] is interactional and sequential, coordinating the parties' activities in achieving a joint understanding." (ib., 1338, our bold). 1.3 The "evolving structure" of conversation as trigger and marker of understanding The "evolving structure" of conversation (see Hopper 1987's "emergent grammar") can be used as a cue to the process of understanding, because its development both triggers and marks the process of understanding. The dialogic and dynamic nature of understanding is 'mirrored' in the evolving structure of conversation, which both triggers and marks this process, via different devices. As Schegloff (1992,130) says: "The understandings are displayed en passant for the most part (although there is also a distinct type of utterance overtly designed to check its speaker's understanding of preceding talk), as by product of bits of 6 "Intersubjectivity would not, then, be merely convergence between multiple interpreters of the world [...] but potentially convergence between the 'doers' of an action or bit of conduct and its recipient, as coproducers of an increment of interactional and social reality." Schegloff (1992, 1299). talk designed in the first instance to do some action such as agreeing, answering, assessing, responding, requesting, and so on." On the one hand, the "conversational exchange" favors comprehension since it gives both information and relevant context: The contextual effects, which have been recently referred to as relevant to understanding (see, among others, Sperber and Wilson 1986, Light and Butterworth 1992), can be considered as including both linguistic and extra-linguistic (see Bazzanella 1998 and Akman and Bazzanella forth.). On the other hand, the on-going process of comprehension may be displayed at least via three kinds of devices activated during the conversation: conversational moves, such as question-answer sequences, interruption, and repair; linguistic markers, such as "Dialogic repetition" or "allo-repetition" (i.e. " the verbatim repetition of an item uttered by another speaker, usually the previous one" Bazzanella 1996, p. ix), paraphrases/ reformulation, and discourse markers; cognitive devices, such as inferences about the hearer's current belief-and-intention states (see Sperber and Wilson 1986, Givon 1997), non-literal meaning, and coherence as an 'interpretive' notion (see Bublitz et al.1999). From now on only the linguistic resources which are used to mark the process of understanding, and, for the sake of space, only two of them (i.e. Dialogic repetition and discourse markers) will be taken into account. 2. Linguistic markers: Dialogic repetition and discourse markers 2.1 The case of "Dialogic repetition" In the increasing number of studies on repetition in recent years, from 1987 (see Johnstone, Norrick, Tannen a and b) to the present (see Johnstone 1994, Fischer 1995, Bazzanella 1996), we almost always find an interest both in the use of language in social interaction, and in the cognitive perspective. Repeating own/someone else’s words has several different macroand micro-functions; in fact: "The constitutive ambiguity of R arises from its unstable balance between variance and invariance, sameness and difference: from the very moment when something is repeated, it ceases to be the same, 7 "A more thorny issue concerns the representation, in the mind of the speaker, of the hearer's current belief-andintention states. Consider first the conventional use of declarative, interrogative and manipulative speech-acts (Grice 1968/1975 inter alia). This use is inconceivable without assuming that speakers construct and carry a running mental model of hearer's current belief-and-intentions. Such a mental model must be extremely dynamic, and must shift constantly between one clause uttered by the speaker and the next." Givon 1977 (100-101). 8 "Indirect utterances are also useful for signaling and stressing mutually shared information." Yus Ramos, 223 "Besides, various studies have demonstrated that figurative language need not to take longer to process than literal utterances if the contextual features are appropriate (Gibbs 1979, 1994 The poetics of mind: 113)" id. 224 9 As Egner (1996,70) rightly states: "The echo-exchange is exclusively a feature of face-to-face interaction and occurs most frequently in phatic conversation, i.e. to confirm and maintain the social relationship through verbal interaction. It has not been found to occur in reported conversations, such as in the dialogues between the protagonists of a folk tale. Moreover, it seems to be absent from all formal types of face-to-face interaction." 10 "[...] to quite an extent the dichotomy between discourse as interaction and discourse as information-flow is an artificial one, reflecting accidents of history and methodology, as well as perhaps pre-empirical philosophical or ideological perspectives. [...] There is something both unnecessary and undesirable about the intellectual segregation between the two sub-fields of discourse studies. To begin with, the episodic memory system that makes coherent multiclausal (or multi-turn) communication possible is the very same cognitive-neurological system, supporting the production and comprehension of monologue and conversation alike. [...] Fundamentally, then, the dichotomy between 'situational' and 'cognitive' is false." Dickinson and Givon 1997,92-93. not only on a semantic level, but also on a pragmatic one: the speech act performed by the original utterance usually differs from the speech act performed by the repeated utterance." Bazzanella (1996, p.ix). Table 1 recalls/presents three out of the seven macro-functions which have been ascribed to repetition: the Cognitive, Conversational, and Interactional ones; the Textual, Stylistic, Argumentative and Ethnic macro-functions have been deleted for the sake of space.