Contribution Graphs in Multiparty Discourse

David G. Novick, Lisa Walton and Karen Ward

Center for Spoken Language Understanding,
Oregon Graduate Institute of Science & Technology
P.O. Box 91000, Portland, Oregon 97291-1000 USA


Clark and Schaefer demonstrated that dyadic discourse could be characterized in terms of contributions, where a contribution consisted of a presentation and an acceptance. Their model, which grows out of the collaborative view of conversation, describes the structure of dyadic discourse through trees of contributions and represents a conversation as a collection of cooperative acts performed by the conversants. In this paper, we show how Clark and Schaefer's model can be adapted to account for the more complex contribution structures created by multiple conversants.

1. Introduction

In the collaborative view of conversation, the active participation of conversants is required to achieve sufficient grounding of the discourse. The kinds of grounding acts that pairs of conversants use have been discussed by Clark and Wilkes-Gibbs [3], Novick [5], and Traum and Hinkelman [9]. The necessity of interactive use of such acts has been demonstrated by Schober and Clark [8], who showed that overhearers did not achieve the comprehension levels of full participants in conversation; overhearers cannot prompt for elaboration when they have not achieved adequate grounding and cannot avoid possibly misleading over-elaboration when they achieve adequate grounding more quickly than "active" hearers. Likewise, speakers rely on feedback from hearers to control the form and extent of their linguistic production. For example, Oviatt and Cohen [7] showed that speakers who did not receive such feedback tended to over-elaborate.

Thus, models of true multiparty conversation will have to account for grounding to and by multiple conversants. The mere presence of an additional hearer can influence both the content of utterances and conversants' efforts to ensure that their contributions are adequately mutual. In the case of active multiparty interaction, the conversants' actions are shaped by feedback from multiple sources.

We suggest in this paper that an adequate model of multiparty contributions should be able to account for at least three cases: (1) where acceptance of a speaker's contribution is a collaborative action by multiple hearers; (2) where a presentation is not accepted; and (3)where some of the conversants engage in side sequences. We present a model of multiparty contribution based on extended definitions of contribution, presentation, and acceptance. We then illustrate the model by building contribution graphs of multiparty discourse. Our example conversation is drawn from transcripts of the White House tapes in the Watergate scandal [4]. We conclude by considering problems with multiparty models and by outlining some open questions.

2. Contribution Models For Multiparty Discourse

A contribution is anything that a person does during a conversation that both carries some content (content specification) and assists the participants in establishing a mutual belief that the hearer believes she understood the content of the presenter's act (grounding). Clark and Schaefer [1], [2] described the process of contributing as containing two phases, a presentation phase and an acceptance phase. Every contribution (except for the very first one) is both a presentation and part of the acceptance phase for some other contribution. More formally, Clark and Schaefer [(2], p. 265) defined their terms as follows:

Presentation Phase
A presents utterance u to consider. He does so on the assumption that, if B gives evidence e or stronger, he can believe that B understands what A means by u.
Acceptance Phase
B accepts utterance u by giving evidence e' that she believes she understands what A means by u. She does so on the assumption that once A registers evidence e', he will also believe that B understands.

In extending the definitions of contribution, presentation and acceptance for multiparty conversation, we make several assumptions about the nature of multiparty interaction. First, we assume that the speaker need not ensure that non-addressees understand the presentation. Second, we note that a hearer may believe that she is an addressee even if she is not addressed directly by the speaker. Finally, we assume that a hearer, even when she believes that she is an addressee, may present less-than-normally strong evidence of understanding if (a) other addressees present normally strong evidence and (b) the hearer believes the other addressees' understanding is sufficiently mutual.

Accordingly, we differentiate between evidence that is of the strength normally sufficient to indicate understanding by an addressee and evidence where strong indication of understanding is not normally required. Let the hearers of A's utterance u be represented as B1, ..., Bn. Then:

A Contribution
is an action by a speaker that has content intended to be conveyed to at least one hearer and that assists some subset of the conversants in establishing mutual belief.
Primary Evidence
is evidence e' presented by hearer Bi where she believes that she was an intended addressee of A's. That is, Bi believes that A requires evidence from her to believe that they mutually understand u.
Secondary Evidence
is evidence e' presented by hearer Bi when she believes that she was not an intended addressee of A's and/or she believes that A does not require primary evidence of understanding.

To accommodate multiple hearers, the notions of the presentation and acceptance phases are extended from the dyadic definitions:

Presentation Phase
A presents utterance u for some subset of B1, ..., Bn to consider based on the assumption that if that same subset of hearers collectively gives enough primary evidence e, he can believe that they understand what he meant by u.
Acceptance Phase
For all hearers 1 <= i <= n, Bi accepts utterance u by giving either primary or secondary evidence that she understands what A means by u. She does so on the assumption that if A registers the evidence, he will believe that A understands.

Thus from the speaker's perspective, a successful contribution requires only that some sufficient proportion of the addressees understand. From the hearer's perspective, it is sufficient to present evidence that she believes she understands if she were an addressee and if not enough of the other hearers presented evidence that they understand. The practical rationale for this is that if every addressee attempted to provide primary evidence of understanding, the speaker would be overwhelmed by acceptances.

These definitions shed additional light on what it means to contribute to a conversation. In dyadic discourse, contributions are always the joint product of two conversants; a contribution requires both presentation and acceptance, and both parties must participate in its creation. In Clark and Schaefer's dyadic model, this is reflected in the binary tree structure of the contribution graph: the dyadic contribution node always has two children, representing a single presentation and a single acceptance phase (although the acceptance phase may itself consist of multiple contributions in the case of, for example, a clarification side-sequence). In multiparty conversation, however, all of the hearers have the possibility of accepting. Conversants may independently offer evidence of understanding, or several may jointly construct a single collaborative acceptance. Similarly, two or more conversants may collaborate on a presentation which may then be accepted by other conversants. Accordingly, multiparty contribution graphs may show multiple acceptances for a single presentation, and a contribution may be the product of two or more conversants.

Another interesting consequence of the multiparty model is the possibility that a single acceptance may provide evidence of understanding for two or more presentations made by different conversants. Clark and Schaefer demonstrated that when closing a dyadic side-sequence, a single acceptance might both accept the immediate previous presentation and accept the side-sequence as a whole. In that case, though, the two accepted presentations are in a hierarchical relationship to one another. In the multiparty case, we have the possibility of a single acceptance being directed toward two different presentations at the same hierarchical level of the conversation. For example, one conversant may make a suggestion. A second conversant may agree, thus offering primary evidence of understanding. A third conversant may follow up with "Me, too," thus offering evidence of understanding both the original suggestion and the subsequent agreement and so resulting in two presentations accepted by a single utterance.

3. Multiparty Contribution Graphs

Representations of contributions following our multiparty model can be constructed from observed conversation. In this section we first summarize our notation for multiparty contribution graphs. We then illustrate application of the model to an excerpt of a conversation from the "Watergate" tapes corpus [4]. As in Clark and Schaefer's original studies [2], the contribution structures reflect patterns of grounding rather than focus. Also, it should be noted that various participants in a conversation may have different beliefs about the structure and content of a particular conversation. The contribution graphs are constructed from the perspective of an idealized overhearer and so may not reflect the perspective of any individual conversant. Furthermore, these transcripts are from audio tape, so they do not capture nonverbal cues such as gesture or gaze that can play an important role in signalling, for example, continued attention. Although our analysis is therefore incomplete in this regard, the transcripts are adequate for illustrating the salient points of our model.

3.1. Notation and Corpus

In the contribution graph diagram (Figure 1) that presents our analysis, the conversants are then-President Richard M. Nixon (N), chief of staff H. R. Haldeman (H), chief domestic affairs adviser John Ehrlichman (E), and presidential counsel John Dean (D). As in Clark and Schaefer's original notation, squares represent contributions, triangles pointing right are presentations, and triangles pointing left are acceptances. Primary evidence of understanding is shown with unbroken lines; secondary evidence of understanding is shown with dotted lines. Initials representing the contributor(s), presenter(s), and accepter(s) are placed within the relevant contribution graph symbols; an `@' stands for all of the conversants. The notation is explained in more detail in [6].

In this conversation, Nixon and members of his staff are discussing how to deal with the unfolding crisis that eventually drove Nixon from office. The discourse excerpted for Figure 1 occurred at the beginning of a meeting in the Oval Office on March 21, 1973 [(4], p. 181-182). The conversants are confronting the problem that the investigation of the Watergate break-in has gotten out of their control because both the District of Columbia and the United States Senate have started investigations. Nixon is concerned that the investigations will lead to the White House; he would like to insulate as many people as possible from prosecution.

[Multiparty Contribution Graph]
Figure 1: Multiparty Contribution Graph

3.2. Application of the Model

The extended definitions of contribution, presentation and acceptance suggest that multiparty discourse diverges significantly from the dyadic model in several ways. For example, some conversants may engage in side sequences, or several conversants may collaborate in accepting a presentation. Some multiparty conversations have structures that may be represented by traditional contribution trees in a straightforward manner.

Side sequences. This excerpt illustrates one aspect of multiparty discourse: the engagement of a subset of the conversants in side sequences (subdialogues). Beginning with turn 5 ("In a Grand Jury?") Nixon and Ehrlichman engage in a side sequence to clarify Ehrlichman's previous utterance. The contributions that constitute such side sequences are the product of the relevant conversational subgroup. The collaborative nature of the side-sequence is expressed by noting multiple conversants in the acceptance symbol, as in the acceptance to Ehrlichman's utterance in turn 4 ("Well, I just don't think the immunity thing will wash.") In some cases, the side-sequence involves repair for the benefit of members of the contributing subgroup. In other cases, the side-sequence is produced for the benefit of another conversant.

Collaborative acceptance. In a collaborative acceptance, multiple conversants present primary and secondary evidence of understanding. This situation could occur because the speaker has directly addressed multiple hearers, some of whom rely on each other to produce sufficient evidence of acceptance; secondary evidence might be provided by one hearer and primary evidence by another hearer. Collaborative acceptance can also occur where a non-addressee retakes the initiative to provide normally strong evidence of acceptance, thus leading to multiple acceptances of a single presentation. Finally, multiple addressees could each produce normally strong evidence of acceptance.

Most of the acceptances in Figure 1 are primary, in that they provide normally strong evidence by an addressee. However, after Ehrlichman completes his turn (10) by saying

I think you have to figure that that is out of the picture. I just don't believe that we can do that. It can't be carried off.

he is waiting for Nixon to accept his presentation with primary evidence of understanding. Instead, Haldeman (turn 11) provides secondary evidence accepting Nixon's utterance in turn 5 via initiation of a next relevant contribution, which is a weak form of acceptance. It is not until after Dean, Haldeman and Nixon's subdialogue on immunity in a special panel (turns 11-14) that Nixon accepts Ehrlichman's contribution (turn 15):

Well, let's take the Grand Jury now, without immunity, and what are your ideas for getting out of it?

At this point, Ehrlichman gets his primary evidence that the President understood his completed utterance.

Thus in this excerpt, Nixon's question in turn 5 ("In a Grand Jury?") is accepted two ways: first through a colloquy (turns 6-9) between Nixon and Ehrlichman that clarifies the utterance that Nixon was questioning, and then through a secondary acceptance from Haldeman in turn 11. Ehrlichman's presentation in turn 10 ("I think you have to figure...") is also accepted in two ways: first through the clarification colloquy (turns 11-14) and second by Nixon's final, responsive utterance in turn 15. Note that one conversant, Nixon, contributed to the acceptance of Ehrlichman's presentations in two ways, both by contributing to the jointly-produced acceptance in the clarification colloquy and by independently producing primary evidence of understanding in turn 15.

Dyadic-like structures in multiparty discourse. Straightforward exchanges can occur in multiparty discourse. In the first four presentations in the excerpt (turns 1-4), the conversation is jointly advanced by three conversants. Each presentation is accepted by a single utterance, although different conversants produce different acceptances. In Figure 1, Nixon's first utterance (turn 1) is to all of the other conversants, asking the group for conclusions. Haldeman offers primary evidence of understanding by responding critically to the President's request (turn 2).

In this part of the conversation, Clark and Schaefer's binary contribution-tree model would produce a plausible structure. We note, though, that the authorship of the different contributions goes unstated in the dyadic model; in the multiparty model, the complexity of the collaborative creation is made explicit.

4. Conclusions

In this paper, we have adapted Clark and Schaefer's model of dyadic discourse to account for the more complex contribution structures created by multiple conversants. In particular, we distinguished primary and secondary evidence of understanding, and we extended the definitions of presentation and acceptance to account for collaborative acceptance. We showed that in actual multiparty discourse there are instances of collaborative acceptance and side sequences; the multiparty contribution graph model we proposed can account for these phenomena.

Our model, while accounting for the principal elements that we have distinguished, is far from a comprehensive explanation of multiparty discourse. There are several open issues. How is the content of a presentation affected by the presence of multiple hearers, each of whom the speaker may wish to leave with a different interpretation of the act? How does the level of evidence required by the speaker change when there are several hearers present? Does the speaker require stronger evidence of understanding because he cannot watch everyone at once? Or does he require less evidence from any one hearer as long as he receives enough total evidence to convince himself that he was understood? Do speakers aggregate acceptance of their presentations or do they require independent levels of acceptance from each addressee? How do the non-verbal presentations (e.g., raised eyebrows) and acceptances (e.g., continued attention) of face-to-face discourse function in the multiparty setting? This analysis, like Clark and Schaefer's, is from transcript and so non-verbal evidence is not represented or considered. How do participants in face-to-face discourse adapt their conversational skills in the presence of multiple targets for mutual gaze?

Despite the limitations of their dyadic origins, Clark and Schaefer's fundamental ideas of presentation, acceptance and contribution continue to account for conversational structure. Their own "contribution" forms the basis of the extended model for multiparty discourse presented here.

5. Acknowledgments

This research was supported by NSF Grant No. IRI-911079 and the member companies of CSLU. The authors thank Brian Hansen for his helpful comments on an earlier draft of this paper.

6. References

  1. Clark, H. & Schaefer, E. "Collaborating on Contributions to Conversations," Language and Cognitive Processes, 2: 19- 41, 1987.
  2. Clark, H. & Schaefer, E. "Contributing to Discourse," Cognitive Science, 13: 259-294, 1989.
  3. Clark, H. H., & Wilkes-Gibbs, D. "Referring as a Collaborative Process," Cognition, 22: 1-39, 1986.
  4. New York Times, The White House transcripts: Submission of Recorded Presidential Conversations to the Committee on the Judiciary of the House of Representatives by President Richard Nixon. Viking, New York, 1974.
  5. Novick, D., Control of Mixed-Initiative Discourse through Meta-locutionary Acts: A Computational Model. Doctoral Dissertation, published in abridged form as Technical Report CIS-TR-88-18, University of Oregon, 1988.
  6. Novick, D. G., Walton, L., & Ward, K. "Contribution Graphs in Multiparty Discourse," Technical Report CS/E 93-015, Computer Science and Engineering Department, Oregon Graduate Institute of Science & Technology, 1993.
  7. Oviatt S., & Cohen, P. "Discourse Structure and Performance Efficiency in Interactive and Non-interactive Spoken Modalities," Computer Speech and Language, 5: 297-326, 1991.
  8. Schober, M., & Clark, H. "Understanding by Addressees and Overhearers," Cognitive Psychology, 21: 211-232, 1989.
  9. Traum, D., & Hinkelman, E. "Conversation Acts in Task- Oriented Spoken Dialogue, Computational Intelligence, 8:3, 575-599, 1992.

This paper was published as: Novick, D., Walton, L., & Ward, K. (1996). Contribution graphs in multiparty conversations, Proceedings of the International Symposium on Spoken Dialogue (ISSD-96), Philadelphia, PA, October, 1996, 53-56.

DGN, March 5, 1997
Valid HTML 4.0!