<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-0218">
  <Title>Dialogue Macrogame Theory</Title>
  <Section position="1" start_page="0" end_page="1" type="metho">
    <SectionTitle>
2 Introduction: 1 Abstract:
</SectionTitle>
    <Paragraph position="0"> Dialogue Macrogame Theory is designed to enable analysis of particular natural dialogues. Some dialogues can be analyzed with DMT; some cannot. Where it fits, DMT gives a partial technical characterization of the classes of dialogues represented in the analyses. This paper is focused on presenting the elements of the theory. Other presentations are expected to describe its application and validation.</Paragraph>
    <Paragraph position="1"> This paper introduces Dialogue Macrogame Theory, a method for describing the organization of certain kinds of dialogues. Dialogue Macrogame Theory (DMT) is a successor to a theory sometimes called Dialogue Game Theory, developed in the 1970s and 1980s at USC-Information Sciences Institute (ISI). DMT is able to describe substantially more dialogues than its predecessor, and it identifies kinds of mechanisms not included in the predecessor. DMT is a step toward accounting for the coherence of entire dialogues.</Paragraph>
    <Paragraph position="2"> In the late 1970s a research team at USC-Information Sciences Institute (ISI) studied natural dialogues with particular interest in applying the results to human-computer interaction. The team produced a series of reports (Mann 1979) (Mann, Carlisle et al. 1977) (Mann, Carlisle et al. 1975), a published paper by Levin and Moore (Levin and Moore 1977) and later another paper, entitled Dialogue Games: Conventions of Human Interaction (Mann 1988). These publications arose out of the study of many natural dialogues. The final theoretical summary was presented in the form of a technical method for representing the short range and long range dialogue structures.</Paragraph>
    <Paragraph position="3"> Recent work has attempted to apply this set of structures to additional dialogues. This effort has identified a number of deficiencies of the framework, which have led to making a fresh start rather than adjustment.</Paragraph>
    <Paragraph position="4"> The major structures in DMT are based on intentions which are imputed to dialogue participants. The focus of this paper is on mechanisms. Dialogue Macrogames are defined.</Paragraph>
    <Paragraph position="5"> Another class of mechanisms, called Unilaterals, is also described.</Paragraph>
    <Paragraph position="6"> A DMT analysis is presented. The analyzed dialogue is an excerpt (41 turns) of actual dialogue from the Apollo 13 mission, from the emergency period after the explosion.</Paragraph>
    <Paragraph position="7"> DMT is then related to another dialogue analysis method (Carletta, Isard et al. 1997).</Paragraph>
    <Paragraph position="8"> DMT is an exercised framework, meaning that it has been applied to dialogues from a diversity of situations. These include various emergency communications, tutoring, administrative interactions, online human computer help, medical interviews, laboratory conversational tasks, courtroom questioning of witnesses and hostage negotiation. The paper reports work in progress, and also indicates likely courses of further development.</Paragraph>
    <Paragraph position="9"> The central constructs in the prior work were called Dialogue Games, a name inspired by terminology of Wittgenstein (Wittgenstein 1973) but with only a distant family resemblance to his usage. Since the end of that work this term has been used in many other ways that are not technically related to the ISI usage. In addition, the terms dialogue act, conversational act, and conversational game are in wide use (Carlson Philadelphia, July 2002, pp. 129-141. Association for Computational Linguistics. Proceedings of the Third SIGdial Workshop on Discourse and Dialogue, 1983) (Kreutel and Matheson 2001) (Traum 1994; Poesio and Traum 1998; Traum 1999) .</Paragraph>
    <Paragraph position="10"> To simply resume using the old terminology now would invite confusion and misunderstanding. So this paper introduces a new term, Dialogue Macrogame, which represents some structures that resemble the dialogue games of the predecessor model. In this paper, Dialogue Macrogame may be abbreviated to dialogue game or even game.</Paragraph>
    <Paragraph position="11"> The term dialogue represents two party immediate interaction.</Paragraph>
    <Paragraph position="12">  Studies of dialogue from various points of view are numerous. There are dozens of technical fields with the word communication in their names (Craig 1999). Many of them study dialogue. There are structural views, communication views and many others. Linguistics tends to produce structurally oriented studies, but not exclusively. Even restricting attention to studies of dialogue coherence, there are many radically different viewpoints. Conversational Coherence: Form, Structure and Strategy (Craig and Tracy 1983) is a particularly relevant collection, now somewhat dated but representing ideas that persist in the wider literature. In this book, and in the wider literature as well, the distinction between coherence and cohesion is often not made. Studies of coherence are often really about abstract cohesive devices, in the sense of (Halliday and Hasan 1976). Some studies (Ellis 1983; Goldberg 1983) assume that coherence is produced by design, by appropriate use of cohesive devices.</Paragraph>
    <Paragraph position="13"> In some studies coherence is equated to topic continuity or to the appropriate use of topic shifting devices (Crow 1983; Sigman 1983). In others it is seen as conformity to expressive rules. Grice is often interpreted as believing that conversational coherence is based on rule (or maxim) following (Grice 1975). Still others see coherence as an identifiable outcome of rule governed (social or linguistic) behavior (Goldberg 1983). Hawes defends his views against this idea (Hawes 1983).</Paragraph>
    <Paragraph position="14"> Some studies of coherence in dialogue assume that people pursue a tacit goal of being coherent, in addition to any other goals, when they interact. It is a benefit which they seek. (Hopper attributes this orientation to interpreters of language, but not producers (Hopper 1983).) Others see coherence as an obligation that is attached to interaction. It is an added duty. (Sanders sees this view as widely accepted, and defends his views against it (Sanders 1983).) Some studies equate coherence with propositional consistency, see (Goldberg 1983) for citations.</Paragraph>
    <Paragraph position="15"> Others see coherence as a kind of summary impression that is a side effect of understanding an interaction, an understanding that is enabled by the processes that ordinarily govern interaction (Sanders 1983).</Paragraph>
    <Paragraph position="16"> The (Craig and Tracy 1983) book incorporates an admirable attempt to make the various approaches comparable. All of the chapter authors were given one particular 30 minute dialogue (included in the volume) and told to relate their approaches to it. In a volume summary, apropos of this paper, the editors note that &amp;quot;Conversationalists' goals must play a central role in any adequate explanation of discourse production and interpretation in conversations.&amp;quot; p. 22.</Paragraph>
    <Paragraph position="17"> Global approaches to coherence, ones that attempt to address entire texts or dialogues, are often associated with some notion of genre or tradition, such as Rummelhardt's story grammars (Rumelhart 1975) or Schank and Abelson's scripts (Schank and Abelson 1977). (DMT addresses whole dialogues, but without resemblance to those approaches.) There is more recent progress in many aspects of understanding dialogue. A rich array of formal approaches has been built on the Discourse Representation Theory of Kamp and colleagues (Kamp and Reyle 1993) (Traum 1994). Agency theory, along with various vigorous efforts to develop data annotation methods, are also producing insightful views of natural dialogue.</Paragraph>
    <Paragraph position="18">  Sometimes DMT structures can be applied to interactions that have more than two participants. Even so, dialogue macrogame theory would have to be significantly augmented in order to become a sound representation scheme for multiparty interaction. Knowing of these complications, we focus on two party interaction as a research tactic.</Paragraph>
    <Paragraph position="19"> Clearly there is no consensus on the nature of coherence in dialogue. Although comparing views is typically difficult, as researchers we always find some views more credible than others. We may be able to make certain alternatives more distinct by an analogy concerning oral dialogue. What is the status of breathing? Do people breathe in dialogue because they believe it will make the dialogue more beneficial? Or is there a duty to breathe in dialogue? Is breathing simply following a tradition, or an attempt to perform smoothly? Are there rules of dialogue that would, for example, make a dialogue ill-formed if it did not involve breathing? Or is breathing regulated by processes that interact with the speaking processes? DMT is designed following assumptions that most resemble this latter alternative.</Paragraph>
    <Paragraph position="20"> Instead, we seek to find a set of theories that can jointly account for the coherence of dialogues that arise in different kinds of situations. We hope that the set will be small, but also that the set of theories will be very informative. DMT is designed to be one such theory.</Paragraph>
    <Paragraph position="21"> Size limitations and incomplete development prevent this paper from describing all of the significant aspects of DMT. We choose to describe only the elements which are part of DMT analysis. In future presentations we expect to describe dialogue analysis methods, framework validation, some form/framework relationships, and relationships to prior methods. We also expect to provide a corpus of analyzed dialogues.</Paragraph>
    <Paragraph position="22"> This work puts a high value on expanding the coverage of the set of theories mentioned above.</Paragraph>
    <Paragraph position="23"> This is in effect a personal preference for placing a high value on breadth rather than depth or precision.</Paragraph>
  </Section>
  <Section position="2" start_page="1" end_page="2" type="metho">
    <SectionTitle>
3 Research Goals:
</SectionTitle>
    <Paragraph position="0"> A major goal of Dialogue Macrogame Theory (DMT) is to provide a descriptive account for the coherence of a wide diversity of natural dialogues.</Paragraph>
    <Paragraph position="1"> A dialogue is said to be coherent if a person who has good access to the dialogue is left with the impression that every part of the dialogue contributed to the remainder, or equivalently that there are no parts whose presence is not easily explained.</Paragraph>
    <Paragraph position="2">  This definition of coherence is parallel to a definition for monologue texts. That definition says that every part of the text has an evident role, and that there are no apparent deletions.</Paragraph>
    <Paragraph position="3"> We can say something preliminary about the degree of confirmation of DMT. As mentioned above, although the predecessor theory expressed some insights about dialogues, it was generally not precise enough to be applied reliably. Details had been missed. DMT has been applied to a much wider range of dialogues. The forthcoming release of a corpus of analyzed dialogues will show this, and the structures described below all reflect details of actual natural dialogues.</Paragraph>
    <Paragraph position="4"> Near the end of this paper there is an example of an analysis of a natural dialogue, illustrating use of each kind of element of DMT.</Paragraph>
    <Paragraph position="5"> Notice that the definition of coherence does not refer to turns, does not require that there be distinct turns, and in particular does not require that turns be coherent monologues. 4 Intentions in Dialogue: Dialogue coherence arises from the intentions (also called goals) of the dialogue participants. It arises especially from the way that the conventions of dialogue cause the participants to adopt and dismiss groups of intentions. Grouping of intentions is the foundation for coordination of the activities of dialogue participants. It provides a way of introducing joint goals into a dialogue, and, equally important, a way of dismissing goals from a dialogue, all under the shared control of both participants.</Paragraph>
    <Paragraph position="6"> The assumptions of DMT include saying that not all dialogue situations are alike, and that the differences between dialogue situations affect the dynamics of the dialogues which occur in them.</Paragraph>
    <Paragraph position="7"> We therefore do not expect to discover a single theory that accurately describes the coherence of all natural dialogues.</Paragraph>
    <Paragraph position="8">  For research purposes, the impressions of dialogue participants themselves are generally not available. Researchers' impressions of dialogue transcripts and records are virtually the only source, so those impressions are the object of study.</Paragraph>
    <Paragraph position="9"> The view of intentions here follows (Mann 2001), which identified a set of attributes of intentions that can be found in the literature of intentions in language. That view makes use of (Bratman 1987; Clark 1996; Gibbs 1999). The view of joint intentions, and of joint actions where they occur, generally follows Clark, who follows (Tuomela 2000).</Paragraph>
  </Section>
  <Section position="3" start_page="2" end_page="2" type="metho">
    <SectionTitle>
5 Dialogue Macrogames:
</SectionTitle>
    <Paragraph position="0"> The major construct of Dialogue Macrogame Theory is, of` course, the dialogue macrogame. A (dialogue macro)game is defined as a set of three goals:  1. A goal of the initiator 2. A goal of the responder 3. A joint goal  A dialogue macrogame is a convention, loosely comparable to a lexical item or a grammatical pattern. DMT expects that people will hold hierarchic goal structures, in general but especially for embedded uses of games. However, these three goals are not in hierarchic configuration. When a game is used, the goal of the initiator and the joint goal will be in the memory of the initiator as commitments. (DMT does not constrain the relationship of these two.) Similarly, the goal of the responder and the joint goal will be in the memory of the responder as commitments. In each memory, the two goals are committed and uncommitted simultaneously, and at the same time each person's knowledge of the other's commitments is adjusted. How this happens, and how grounding can be continuous, are described below.</Paragraph>
    <Paragraph position="1"> One of the games is named the Information Offering game. Like all of the other games, a single occurrence of this game can be used to account for an indefinitely long interval of interaction. Currently in DMT there are about 19 defined games.</Paragraph>
    <Paragraph position="2"> The definition of the Information Offering game is:  1. goal of the initiator: to provide particular information to recipient 2. goal of the responder: to identify and receive the particular information offered 3. joint goal: the responder comes to possess  the particular information These, of course, are not fully specified goals. Rather, they (and all of the goals of the framework) have places for unspecified arguments, such as the particular information above. DMT does not assume that the dialogue is task oriented; if it is, the goals of a particular use of a game may be task goals, and a particular goal can appear in more than one position in a game use.</Paragraph>
    <Paragraph position="3"> All of the individual participants' goals that are used in DMT games have the following attributes in the (Mann 2001) framework: partialness, priorness, tacitness, immediacy , interactionconfiguring, intended to be recognized, structuredness, complementarity, conventionality. The joint goals have the additional attribute of jointness, and they lack complementarity. The course of a dialogue between peers is generally under the control of both participants. They coordinate and jointly control by means of particular kinds of actions. DMT uses a negotiation metaphor to describe this. A game is bid by the initiator, and the responder accepts or refuses the bid. These actions are almost always implicit, but certain situations (especially the diagnosis of misunderstanding) can cause them to become explicit. Similarly, games terminate by negotiation. If a game has been bid and the bid has been accepted the game is open.</Paragraph>
    <Paragraph position="4"> Either party can bid termination of a game, and the other can either accept termination or refuse it by continuing to pursue the game. Games often end by apparent accomplishment of the joint goals of the game, which is the most common form of bidding termination. Abandoning the goals of the game accepts termination. So for example giving an answer to a question will generally be seen by both parties as the moment when pursuing the question should end. Thus apparent satisfaction of the game goals bids termination. There are other ways to terminate a game. The scope of a game is the entire interval during which it is in use, including the initial bid of the game and the final acceptance of termination.</Paragraph>
    <Paragraph position="5"> If a bid to initiate a game is refused, as can happen for example when one person offers information that the other person does not need, then the game never becomes open, and the joint goal of providing the offered information is never taken up.</Paragraph>
  </Section>
  <Section position="4" start_page="2" end_page="2" type="metho">
    <SectionTitle>
7 An Example of DMT Analysis: Occurrences of games can contain occurrences of
</SectionTitle>
    <Paragraph position="0"> other games or the same game within their scopes; they are mutually recursive. The representations in DMT are really only about the beginnings and ends of game scopes, not about how the goals are pursued. It is thus a very partial view of dialogue structure. It generally does not represent all of the goals of a participant. If the dialogue is task oriented, there are task goals that may not be represented. The control of the interaction will tend to be well represented, but private and long term goals less so. Since DMT applies to entire dialogues, it supplements views of sentences, turns or inherently small collections of turns in dialogue.</Paragraph>
    <Paragraph position="1"> As an example, consider the dialogue excerpt below. It is part of the actual transcript of the Apollo 13 mission to the moon, not the movie script.</Paragraph>
    <Paragraph position="2"> The dialogue occurs near the end of the mission, after the major explosion and after many emergency measures have been completed. The spacecraft is traveling toward earth, and the dialogue is concerned with completing the mission. The central concerns at this moment are managing fuel and adjusting the path of the spacecraft. (The indicated turns are from the original transcript.</Paragraph>
    <Paragraph position="3"> Successive turns by the same speaker sometimes indicate a silent interval.) CMP and CC are NASA designators for the communicating parties; &amp;quot;Roger&amp;quot; is jargon for &amp;quot;I heard you.&amp;quot; The communication channel is continuously open, but radio quality varies.</Paragraph>
  </Section>
  <Section position="5" start_page="2" end_page="41" type="metho">
    <SectionTitle>
6 Unilaterals:
</SectionTitle>
    <Paragraph position="0"> It turns out that most dialogues do not consist entirely of goal pursuit that could be represented by dialogue macrogames. Other things happen, even when the major activity is joint goal pursuit.</Paragraph>
    <Paragraph position="1"> DMT has a class of actions that do not involve joint goals and that are generally confined to a single turn. They are called Unilateral Actions or Unilaterals. The specific unilaterals currently recognized are called Action and Tell.</Paragraph>
    <Paragraph position="2"> The entire text is part of an analysis diagram below.</Paragraph>
    <Paragraph position="3"> The text is an excerpt; the beginning and end of the dialogue do not have the usual sorts of greetings, in part because the radio communication medium remains open even when all parties are silent. By training and because of the life threatening situation, everyone communicates extremely carefully. Extensive use of &amp;quot;Roger&amp;quot; and multiple use of Clarification Seeking both indicate this.</Paragraph>
    <Paragraph position="4"> There are other Unilaterals that are presently represented by a category, Media Management, representing communication that is primarily involved with the medium of communication.</Paragraph>
    <Paragraph position="5"> There are also Unilaterals in categories called Politeness and Acknowledgement. These should be differentiated further someday.</Paragraph>
    <Paragraph position="6"> All of the technical elements in the analysis have definitions in the DMT framework. This example dialogue analysis is based on the definitions, but they are too numerous to present. The names are representative, and the definitions are being made available. The Analysis Summary below shows the scopes (durations of use) of each instance of use of a game according to the analysis.</Paragraph>
    <Paragraph position="7"> We noted above that DMT does not make direct use of the notion of turns in dialogue, and that it does not require that turns be coherent monologues. Given the bidding and termination of games as outlined above, DMT would predict that some turns would terminate one game and bid another, thus producing an incoherent fragment.</Paragraph>
    <Paragraph position="8"> Although this pattern is not seen in the example dialogue below, it occurs. Turns are a presentational convenience, but DMT analysis sometimes requires other boundaries.</Paragraph>
    <Paragraph position="9">  It is somewhat problematic to be definite about the nature of turns. In addition to game opening and closing, there are overlapping, echoing, failed interruptions, completions and competition for &amp;quot;the floor.&amp;quot; All of these create variations on the simple notion of turns. Presentationally we try to use it only where there are no subtleties.</Paragraph>
    <Paragraph position="10">  The detailed analysis appears in a two page diagram below.</Paragraph>
    <Paragraph position="11"> The following are diagram abbreviations: The graphic convention is described in the table below. The text is given, divided into turns, which may represent speaker alternations or some finer divisions. For each turn, there is a graphic anchor point. There are columns which indicate the depth of embedding of the game. The highest level, not embedded, is labeled depth 1.</Paragraph>
    <Paragraph position="12">  Refer to the two-page analysis diagram.</Paragraph>
    <Paragraph position="13"> The first four turns are identified as Media Management, one class of Unilaterals. They establish which ground control station is calling, and that the radio quality is adequate. (Turn numbers are from the original NASA transcript.) In turn 5, CC is offering information to CMP.</Paragraph>
    <Paragraph position="14"> DMT analysis represents this as a bid of the Information Offering game. It suggests that the three goals of the Information Offering game, as specified by the identification of what information is offered, be adopted for immediate pursuit. The initiator, who is the speaker of turn 5, would take up the initiator's goal and the joint goal. The responder would take up the responder's goal and the joint goal. The goals are not negotiated individually.</Paragraph>
    <Paragraph position="15"> In turn 6, CMP explicitly accepts the bid. Explicit acceptance is rare. Most often there is explicitly only an acknowledgement or, for most games, initial pursuit of the goals of the game. This giving of information under this bid and acceptance is, in principle, indefinitely long. Termination can occur in various ways, the most common being accomplishment of the goals of the game.</Paragraph>
    <Paragraph position="16"> The analysis diagram shows this use of the Information Offering game as extending from turn 5 through turn 16. Within this scope, in turn 13, CMP bids the Clarification Seeking game. This is not quite a typical clarification seeking because the bidder knows the answer to his question. The turn is saying, indirectly, that CC has failed to be explicit about the sign of the number, plus or minus, and so it requests that CC be specific. In context, the question is thus humorous. Turn 14 is Media Management, and so at an analytic level the bid has not been accepted or rejected. Turn 15 begins to provide the answer, and so accepts the bid in turn 13. Turn 16 acknowledges accomplishment of the goals of the Clarification Seeking game, the inner, most recently started game, and also the goals of the Information Offering game. The termination of these games can be identified by the fact that neither party continues to pursue goals of the games.</Paragraph>
    <Paragraph position="17"> In turn 17 a new episode starts with CMP requesting an action by CC and the team on the ground. This is a bid of the Action Seeking game. Since turn 18 is an Acknowledgement, another Unilateral, it does not introduce or remove any goals from being active. Turn 19 accepts the bid, and as a promise to do the action it also bids termination of the game. Termination is accepted in turn 20.</Paragraph>
    <Paragraph position="18"> Turn 21 is obscure. It may involve fuel-burning parts of the spacecraft. Assuming that, it is background information for some bid of a game in which there is talk about those parts or about fuel. Thus turns 21 and 23 together constitute a bid, in this case of the Information Seeking game. Turns 24 through 28 manage a pause, one that is long enough that the ground control station identifies itself again in turn 26. Turns 29 through 31 begin to provide the information, and turns 33 through 36 complete it and terminate the Information Seeking game in the normal way.</Paragraph>
    <Paragraph position="19"> However, turn 32 attempts to get a clarification. As such it is a bid of the Clarification Seeking game. It is ignored, which is equivalent to being refused.</Paragraph>
    <Paragraph position="20">  In turn 38 CMP again seeks a clarification, which is a new bid of the Clarification Seeking game. It is pursued to closure in turn 41.</Paragraph>
  </Section>
class="xml-element"></Paper>