Discourse Routines in Decision-Making Interactions

Conversational analysts claim that discourse is organized in routines that specify the most likely continuation to an utterance embodying a given discourse function. Such adjacency pairs operate as the elements of a frame or schema that enables efficient processing by guiding inferences and helping participants maintain common ground. We examine these routines within the framework of a problem-solving model that specifies the dependencies among orientations, suggestions, and evaluations, and that provides a coding system whereby utterances may be identified with different functions within the model. We demonstrate that both face-to-face and computer-mediated interactions adhere to the model, although we find greater adherence in the latter, arguably due to the increased processing and attentional demands calling for increased efficiency. The specific dependencies among functions are evaluated to test the claim that particular adjacency pair routines specify the preferred or unmarked continuations, and that dispreferred or marked continuations will require more linguistic form. Finally, we use Markov analyses to identify the most probable dependencies in our data, and to examine the extent to which the resulting Markov model is consistent with the problem-solving model we initially proposed.
 

A similar problem arises in approaches to discourse understanding that make use of scripts, frames, schemas, or memory organization packets (Anderson & Pearson 1984; Goffman 1981; Minsky 1975; Schank & Abelson 1977; Schank 1982; Turner & Cullingford 1989). Such structures are usually described as abstract knowledge structures or vague "understandings" that people have about their behavior, but the precise role these structures play in determining actual language processing has not been specified. In particular, we again find a failure to provide a good account of the properties associated with adjacency pairs.

Adjacency pairs have been studied most directly by conversation analysts concerned with describing the dynamics of natural conversation. Many analyses of conversation have focused on the routine sequencing of language in speech events like the beginnings of telephone interactions (Schegloff 1968, 1986). If patterns in language data are routine or script-like, they should occur frequently in appropriate samples. Yet after investigating 25 telephone openings, Hopper 1989 (p. 190) concludes, "When all exceptions are considered, one is left with a considerably greater number of non-routine calls than routine ones. What is the status of a model of a routine which is, in actual full occurrence, not all that frequent?" He observes that researchers should not expect language data to conform exclusively to routine patterns because as Schegloff 1986 remarks, participants in interaction may choose to violate the expectations of discourse routines to convey particular meanings. For example, the opening routines of telephone conversations may be abbreviated in expressions of intimacy or urgency. Yet routines can be violated only if they are established: Some reliable process must make routines recognizable to the participants who establish, learn, and employ them.

It seems clear that systematic quantitative analyses of discourse routines will contribute to researchers' understanding of sequential dependencies in conversation. Currently, the greatest efforts to quantify discourse phenomena are being made by computational linguists motivated by the possibility of exploiting discourse patterns in technologies for automated language processing. Moreover, the issues that Hopper raises should be of great concern to these researchers, since the patterns observed by language experts must also be "observed" by computer software in order to be useful. Therefore, this paper reports the results of some initial efforts to quantitatively examine sequential dependencies at the discourse level. We examine two modalities, face-to-face and synchronous computer-mediated, since previous work provides evidence that participants adopt the same routine sequences under both conditions (Condon & Cech 1996a,b). The cooperative decision-making tasks that participants engaged in allowed examination of a decision routine along with request/compliance pairs.

First, because routine continuations are expected, routines can contribute to the top down processing of language. We believe that routine continuations are not merely expected, but are actually assumed during processing unless an alternative function is signaled in the linguistic form (Condon & Cech 1996a). Therefore, a potentially infinite search for the function of an utterance can be avoided by a successful match to the expected function (thereby predicting faster processing of routine continuations).

Also, a theory of routine continuations makes testable predictions about the amount of linguistic form observed in routine and nonroutine continuations, in addition to predictions about adjacency. If routine continuations are anticipated, it is not necessary to expend linguistic form on encoding the expected function. In the extreme case, an expected continuation might not be encoded in linguistic form at all (see Condon & Cech 1996a, b, forthcoming, for examples). Therefore, the theory predicts that conversation will employ ellipsis, fragments, and other minimal linguistic forms like discourse markers and interjections and that these will occur more frequently in routine continuations than in nonroutine ones. In contrast, nonroutine continuations should require more linguistic encoding to achieve the unexpected function.

In this study we examine two types of routines: adjacency pairs associated with requests (including requests for information, action, and validation), and a slightly more complex decision routine involving a sequence of 3 functions. An overview of the decision routine appears in Figure 1. Simple as it is, it will prove to exert powerful constraints on the interpretation of linguistic forms in decision-making contexts.

The functions represented in block capitals in the first column of Figure 1 are intended to suggest that the structure observed in the interactions instantiates a more general schema in which an initial specific goal (GOAL) is selected or agreed upon, and this goal serves to focus the ensuing discourse. At the simplest level, a potentially satisfying solution is proposed (INPUT), which must be tested for adequacy by each participant (EVALUATION). As attainment of the goal requires a group decision, there must be some means for determining when there is group agreement (CONSENSUS). Subsequently, whatever wrap-up actions are required that relate to that goal may be taken (OUTPUT). Finally, the participants cycle on to the next goal. Given that most tasks will require a hierarchy of goals (in order to go on a picnic, one must determine what day to go, what foods to bring, whom to invite, what to do for entertainment, etc.), the participants will cycle through a number of tasks (NEXT GOAL) until all the planning has been accomplished to meet the constraints of the high-level goal (PLAN A PICNIC).

Schema Functions                                                                             Utterance Functions
 

The functions represented in lower case type in the second column of Figure 1 relate the general schema to utterance functions in actual discourse. Example (1) provides an excerpt from a computer-mediated interaction in which utterances are labeled to illustrate these functions. Below, P1 and P2 designate first and second speaker (an utterance that is a continuation by the previous speaker doesn't carry a code).

(1) a.   P2:     [orientation]     what do you want to do in the morning
     b.   P1:     [suggestion]      sleep
     c.   P2:     [agreement]      cool
     d.             [suggestion incorporates orientation]     I say we lay out in the afternoon
     e.   P1:     [agreement]      ok
     f.              [orientation/suggestion]     and at night we party
     g.   P2:     [agreement]      yea
     h.   P1:     [orientation]     ?
     i.              [orientation]     whats next?

Example (2) provides an annotated excerpt from a face-to-face interaction.
 
(2) a.   P1:      [orientation]    where would it be?
      b.             [suggestion]     uh rent a hall or have this at someone's house?
      c.   P2:     [agreement]      probably want to rent a hall 'cause people our age would destroy someone's house [laugh]
      d.   P1:     [orientation]     beverages
      e.             [suggestion]      beer kegs of beer [laugh]
      f.    P2:     [agreement]      exactly
      g.             [orientation]      well they have to eat um
      h.   P1:     [suggestion]       well cake definitely if it's going to be a birthday party
      i.    P2:     [agreement]       yeah

Orientations like (1a,2a) establish constraints for each decision while suggestions like (1b) formulate proposals within these constraints. Agreements (1c,e,g) and disagreements evaluate a proposal and test for consensus. Finally, the third column of Figure 1 is meant to suggest both that disagreements are departures from the expected course of the conversation and that disagreements will need to be explicitly marked. In this model, a lack of explicit consensus does not signal a disagreement, since consensus is the assumed or unmarked function. In contrast, disagreements are dispreferred seconds (Levinson 1983) with the characteristic features of these (including prefacing by well, the presence of mitigating devices, and especially, explanations of reasons for not accepting the suggestion). Consequently, more linguistic form is needed when utterances fail to conform to the decision routine, whereas little form is required when the talk satisfies routine expectations and a decision routine involving the simple sequence of orientation, suggestion, and agreement.
 

Dyads who interacted electronically were seated at microcomputers in separate rooms. They received additional instruction about sending messages on the system and were asked to practice using the system before they began the tasks. At all times during their interactions, a portion of each participant's screen displayed the system command keys and their functions. These participants communicated by typing messages which appeared on the sender's monitor as they were typed, but did not appear on the receiver's monitor until the sender pressed a SEND key. The software incorporated this feature to provide well-defined turns and to make it possible to capture and change messages in future studies. The software recorded message source, message initiation and sent times, and message text (including corrections: see Condon & Cech 1996a for additional details).

Participants were randomly assigned one of two variants of two different planning tasks. In one task (the weekend task), they either planned an itinerary for a weekend getaway or to entertain a friend visiting southern Louisiana. In the other task (the social event task), they either planned a party or a barbecue. Participants completed answer sheets to ensure a minimum number of decisions. The answer sheets for the weekend task provided space to record locations and activities for the mornings, afternoons and evenings of two days. For the social event tasks, the answer sheets provided spaces designated time, location, food, beverages, activities and entertainment.

Task type and task order were counterbalanced across the 30 dyads in each interaction modality. The results below are based on 16 oral-condition and 16 electronic-condition conversations selected to represent 4 conversations for each problem type, two in which that task was the first task attempted, and two in which it was the second. At least one interaction from each dyad is included.
 

The data were then coded by students who received academic credit for their involvement in the project. The coding scheme we employ identifies three broad classes of discourse function: MOVE, RESPONSE, and OTHER. As utterances could often be coded in several different subcategories, coders assigned an utterance to the highest possible function within each of these three categories (the functions, their abbreviations, and their hierarchical arrangement are presented at the bottom of Figure 2). Each category is described briefly below. More complete descriptions can be found in Condon and Cech 1992, 1996a.

MOVE functions are those which invite a response, and so act as first pair-parts of adjacency pairs with obligatory second pair-parts. The highest MOVE function, Suggests Action, corresponds to the suggestion/input function in the decision routine, as in (1b,d,f) and (2b,e,h). Another important MOVE function is Requests Action, which often corresponds to the output function in the decision routine. We place utterances that propose behaviors in the speech event in this category. For example, most interactions included requests concerned with recording answers on forms provided to participants, as in Example (3).

(3)  a. write it in activities   (oral)
      b. Hey you write down the details   (electronic)
      c. well list your two down there   (oral)

We place utterances in the category Requests Information if they seek information not already provided in the discourse, as in (1a,i) and (2a). Utterances that seek confirmation or verification of provided information, however, are coded as Requests Validation. The final MOVE category, Elaborates-Repeats, serves as a catch-all for utterances with comprehensible content that do not serve any other MOVE or RESPONSE functions. Frequently these are repetitions, utterances that support or comment on suggestions, and utterances that elaborate responses.

RESPONSE functions generally represent second pair-parts of MOVE functions. The highest of these, Agrees with Suggestion, codes routine continuation of suggestions (in contrast to the self-explanatory Disagrees with Suggestion). Complies with Request identifies utterances that comply with any of the three types of requests; Acknowledges Only was restricted to forms like yeah that acknowledge or repeat a partner's previous utterance.

The OTHER function types combine categories designed to reflect discourse management strategies as well as two categories included to assess affective functions. The latter include Requests/Offers Personal Information, in which participants discuss personal information or make other personal comments not required to complete the task, and Jokes Exaggerates (utterances that inject humor). The highest OTHER function is Discourse Marker, which is used for a limited set of forms: Ok, well, anyway, so, now, let's see, and alright. Another category, Metalanguage, was used to code utterances about the talk, while the Orients Suggestion category identifies utterances that function to orient decisions, including requests for information like (1a), fronted orientations incorporated into suggestions like (1f), and orientations expressed as short phrases or statements (2d,g).

Each utterance was given a coding for each of these three categories. In cases involving no clear function, the utterance was assigned a No Clear code.

Coders were trained by coding and discussing excerpts from the data. Reliability tests were administered frequently during the coding process. Reliability scores were high (80-100% agreement with a standard) for frequently occurring move and response functions, discourse markers, and the two categories designed to identify affective functions. Scores for infrequent move and response functions, metalanguage, and orientations were somewhat less reliable. Consequently, each interaction was coded by two coders and disagreements were resolved by the first author.
 

Analyses of variance revealed a significant interaction of function type with discourse modality in all categories of function, F (4,120) = 24.1, p < .001 for MOVES; F (3,90) = 20.2, p < .001 for RESPONSES; F (4,120) = 22.9, p < .001 for OTHER. Subsequent tests of simple effects of modality type at each function indicated that all were significantly different in the two conditions except Requests Validation, Disagrees, and the two affective categories, Requests/Offers Personal Information and Jokes-Exaggerates.

In spite of these differences, the ratios of utterances linked in routines are similar: The ratio of utterances coded as suggestions to utterances coded as agreements is 1.64 in the oral interactions and 1.71 in the computer-mediated interactions. In fact, the ratio of utterances coded as requests (summing requests for action, information, and validation) to utterances coded as complying with requests is 1.86 in both conditions. We interpret these ratios as evidence that participants rely on routines even under novel conditions. However, if the interactions were to conform exactly to the decision and request routines, and if all routines were to require explicit linguistic coding for both the first pair and the second pair part, then the ratios reported above should be at unity, since each suggestion ought to be followed by an agreement and each request by an indication of compliance. Because the actual ratios are closer to 2, we can anticipate that analyses of actual sequences will show that suggestions and requests are adjacent to agreements and compliances on only about half of the possible occasions, in accord with a claim that expected or unmarked continuations need not always be coded in linguistic form.

Further evidence in support of our model may be obtained from examining the data of the individual interactions. In the face-to-face condition, all 16 dyads exhibited more numbers of Suggestions than Agreements, and the average utterance length of an Agreement was shorter in each. Moreover, of the 14 dyads who had a Disagreement, all exhibited longer utterance lengths for these than for Agreements (all also had more Agreements). In like fashion, Requests were more numerous than Compliances (all dyads), and longer (for 14 dyads).

Similar results were obtained for the computer-mediated condition, with the exceptions that there were too few Disagreements for meaningful comparisons (only 4 dyads), and Compliances, though less numerous than Requests, were not significantly different from Requests in length.

As expected, other functions such as Metalanguage or Orients Suggestion involved longer utterances than Compliances and Agreements. The data thus generally support our claim that unmarked continuations are predictable, and will thus require less (and in some cases, no) surface form.

                    Orients Suggestion     =>      Suggests Action
                    Suggests Action         =>      Agrees with Suggestion

        Request Routine:

                    Requests Information   =>     Complies with Request
                    Requests Action          =>     Complies with Request
                    Requests Validation     =>     Complies with Request

A stronger test of the model, however, involves examining adjacency predictions. Here, we will need to distinguish strict adjacency, when identifying the immediately following utterance regardless of its source, from strict turn adjacency, the first utterance produced by a different participant. Since adjacency pairs, including suggestions and agreements, usually require that the first pair-part be produced by a different participant than the second pair-part, the better measure of the proportion of agreements following suggestions is obtained by examining turn adjacency rather than simple adjacency. Figure 3 provides the adjacent pairs linked by the two routine types we examined.

We may assess conformity to the predictions in Figure 3 by examining the contingent probabilities that an utterance coded in some category will be continued by an utterance coded in another category. In order to distinguish routine from nonroutine continuations using contingent probabilities, we employed the criteria in (4).
 

(4) a. The probability that a routine function will be followed by a routine continuation should be higher than the probability that the function will be followed by all other nonroutine continuations combined.
     b. The probability that a routine function will be followed by a routine continuation should be significantly higher than the probability that the function will be followed by any single other continuation.
     c. Probabilities for continuations from nonroutine functions should not satisfy (4a) and (4b).

We posit a third requirement in (4c) to insure that the contingent probabilities associated with routine continuations are distinct from those associated with all nonroutine continuations. Even if the contingent probabilities were nearly 1, the routine sequences identified there would not be distinguishable from nonroutine sequences if those, too, approached 1. Therefore, it is important to compare the values obtained in routine sequences to values for all other sequences in the data. This task is complicated by our use of the No Clear MOVE, RESPONSE and OTHER categories, which may occur in high frequencies due to the structure of the coding system rather than the structure of the interactions. It is this need for a method of computing probabilities of continuations relative to probabilities of other continuations that motivates our use of Markov analyses reported in the next section.

Table 1 presents the proportions needed to estimate contingent probabilities for the criteria in (4a) and (4b). (We exclude the No Clear MOVE, RESPONSE and OTHER categories from the right-most columns for reasons discussed above). To examine the proportions of agreements that follow suggestions and the proportions of compliance indicators that follow requests, turn adjacency is employed in Table 1. In addition, the proportions of orientations followed by suggestions is provided in Table 1, but using strict adjacency, since the routine does not specify that orientations and suggestions must be produced by separate participants.

Many values associated with routine continuations approach or exceed .5, satisfying (4a), whereas very few of the values in the non-continuation columns do so, despite the fact that the coding system we use allows the possibility of several continuations exceeding .5 (since we track frequencies separately for three broad categories for each utterance, and not just one category). Thus, despite some noise in the data, the results support the claim that certain continuations are expected, once the first pair-parts have been identified.

R*   =>   CR                                                        .46             .48                     .14             .34

To address criterion (4c), we can estimate contingent probabilities like those in Table 1 for each code category using both strict adjacency and strict turn adjacency. For the most part, the proportions in Table 1 are distinct from proportions obtained for other sequences in the data. While these results present some encouraging evidence for our ideas about routines and for the possibility of identifying routine sequences using frequency-based measures, it would be preferable to devise more systematic ways of using the information provided in our coded data, especially to address the need to compare relative frequencies in assessing (4c). Consequently, we used Markov analyses to obtain a more precise picture of sequential dependencies in our data.
 

A zero-order analysis provides the proportions of triples that occur in the data. Of 180 possible, only 79 occur in the oral corpus; moreover, only 10 of these account for 80% of the 4141 utterance events. In the computer-mediated corpus, in like fashion, 12 of the observed 53 triples account for 80% of the 918 utterances.

A first-order analysis yields the proportions of pairs of utterance events (pairs of triples) in the data. There are 180² or 32,400 possible paired sequences of utterances permitted by the coding system, almost 8 times greater than the number of utterance events in the oral sample and more than 30 times greater than the number in the computer-mediated sample. Accordingly, sequences that occur as often as once per discourse (16 in each condition) would be extraordinarily rare under chance conditions, even if their proportions are low. In fact, the 4156 sequences in the face-to-face condition represent only 1921 distinct pairs, 57 of which occur with a frequency of 16 or greater and 6 of which occur more than 100 times in the corpus. In contrast, the 934 sequences in the computer-mediated condition represent only 342 distinct pairs, and only 7 sequences occur with a frequency of 16 or greater (the highest of which occurs with a frequency of 42).

Since the ratio of distinct pairs to the number of possible pairs in the sample is .46 in the oral condition and .36 in the electronic condition, participants in the computer-mediated interactions conformed to the same sequences more than participants in the face-to-face interactions, consistent with previous observations (Condon & Cech 1996a,b) that computer-mediated interactions contained higher proportions of functions linked in routines. However, the frequencies of specific sequences show that the computer-mediated interactions do not have higher proportions of sequences with a frequency of 16 or greater. Since the sample size in the oral condition is 4 times greater than in the electronic condition, we might expect there to be 4 times as many utterances with a frequency of 16 or greater, but the obtained frequencies are 8 times greater. These analyses thus demonstrate further differences among oral and electronic interactions.

The results of the first- and second-order Markov analyses can be presented as lists of events ordered from most to least probable. Figure 4 provides an example by listing the 7 most frequent 2nd-order sequences in the computer-mediated data. Abbreviations for the categories involved in the decision routine are expanded to facilitate observation of the routine. Other abbreviations are listed at the bottom of Figure 2.

Consistent with our model, all of the continuations in Figure 4 include one or more of the predicted continuations in Figure 3. A similar listing for the oral interactions would reveal that only 2 of the 7 most frequent sequences include one of the predicted continuations due to the fact that the largest proportion of utterances in the oral condition is coded as Elaborates, Repeats, a catch-all category that resembles the No Clear MOVE, RESPONSE, and OTHER functions. The five other most frequent sequences involve this category, and are thus uninformative about adjacency sequences, although strictly speaking, an elaboration or repetition may be regarded as a continuation of the first part of an adjacency sequence, rather than the response to it. In any case, as Figure 2 illustrates, the proportions of Suggests Action and Elaborates, Repeats are nearly reversed in the two conditions. The goal of minimizing expenditures of linguistic form due to the increased cost of computer-mediated communication seems to motivate less elaboration and repetition.

                Sequence                                                                                             Probability

nc - Agrees - nc    =>   Suggests - nc - nc   =>
            nc - Agrees - nc                                                                                             .014

Suggests - nc - nc    =>   nc - Agrees - nc    =>
            ri - nc - Orients                                                                                              .014

nc - Agrees - nc   =>    ri - nc - Orients    =>
            Suggests - cr - nc                                                                                           .012

ri - nc - Orients   =>   Suggests - cr - nc    =>
            nc - Agrees - nc                                                                                     `        .011

nc - Agrees - nc    =>   ri - nc - Orients    =>
            Suggests - nc - nc                                                                                            .008

Suggests - nc - nc    =>   Suggests - nc - nc    =>
            nc - Agrees - nc                                                                                               .008

Suggests - nc - nc    =>   nc - Agrees - nc    =>
            Suggests - nc - nc                                                                                             .006

The sequences in Figure 4 not only include single routine continuations, but also sequences of two routine continuations. In fact, if we revise the decision routine as in Figure 5, then each sequence in Figure 4 traces a path in the model in Figure 5.

One drawback of the Markov analyses is that the presentation of utterances in triples makes it difficult to obtain information for a single function. There are several methods we could use to try to reduce the information available in the coded triples, but few that would preserve the relative frequencies that the Markov analyses provide. Nevertheless, one measure seems promising. Given a model like Figure 5, and a listing like Figure 4, we can calculate the proportion of events that trace a path in the model. Using the 0-order analyses gives the proportion of utterances associated with 1 or more of the functions linked in the model. The 1st-order analyses then give the proportion of sequences from one function to another, the proportion of adjacent pairs of utterances, that match one of the 5 links in the model. The 2nd-order analyses provide proportions for sequences of three utterances.

Table 2 presents the results of performing the measure just described on the oral and electronic interactions. The proportions therefore reflect the average (and standard deviation) per discourse of events that conform to the routine in Figure 5. Since conforming to the model is less and less likely as more functions are linked in sequence, it is not surprising that the proportions decrease as the order of the Markov analysis increases. Still, it is encouraging that the proportions of routine continuations in the 1st-order analyses are approximately equal to the proportions of suggestions in the two types of interactions, which suggests that nearly every utterance that could participate in a routine like the one in Figure 5 does participate in at least one routine continuation. It is also impressive that the proportion does not decrease nearly as dramatically in the computer-mediated condition under the more stringent 2nd-order analysis.

1 (Sequence of Two)                                                                     .16    (.06)                              .32    (.13)

2 (Sequence of Three)                                                                   .07    (.04)                               .21    (.11)

Condon and Cech 1996b adopt the term compression to describe the use of significantly less linguistic form to accomplish the task in computer-mediated interaction. This efficiency seems to reflect greater reliance on discourse routines and on strategies that combine discourse functions. Since decision-making is the goal of the task, the category Suggests Action provides an estimate of the proportion of language that is engaged in decision routines, and Figure 2 shows that the proportion of suggestions is almost twice as high in the computer-mediated interactions. The proportions of requests are also greater. The contingent proportions in Table 1 provide additional evidence of greater conformity to routine sequences in computer-mediated interactions, since most proportions in Table 1 increase in this condition. Finally, the higher proportions associated with computer-mediated interactions in Table 2 provide even stronger evidence of greater conformance to discourse routines than in face-to-face interactions. Consequently, we are discovering that computer-mediated interactions are unexpectedly useful for investigating discourse routines.

It seems likely that Markov analyses can provide sensitive measures that will be useful for identifying differences between interactions and for measuring the effects of experimental factors on interactions. The analyses also clarified the effects of coding conventions such as those involving the category Elaborates-Repeats, which raises the possibility that measures based on Markov analyses can be used to investigate such difficult issues as the systematic effects of code design and coding conventions on experimental results. It may be that further anaysis will suggest other Markov-like models of the process, for example hidden Markov models (Charniak 1993, Chapter 3). Clearly, Markov analyses provide many options both for developing and testing models of discourse routines.
 
 
 

Charniak, E. 1993. Statistical Language Learning. Cambridge, MA: MIT Press.

Condon, S., and Cech, C. 1992. Manual for Coding Decision-Making Interactions. Rev. 1995. Unpublished manuscript available at Discourse Resource Initiative website at http://www.georgetown.edu/luperfoy/DiscourseTreebank/dri-home.html

Condon, S., and Cech, C. 1996a. Functional Comparison of Face-to-Face and Computer-Mediated Decision-Making Interactions. In Herring, S. (ed.), Computer-Mediated Communication: Linguistic, Social, and Cross-Cultural Perspectives. Philadelphia: John Benjamin.

Condon, S., and Cech, C. 1996b. Discourse Management in Face-to-Face and Computer-Mediated Decision-Making Interactions. Electronic Journal of Communication/La Revue Electronique de Communication, 6, 3.

Condon, S., and Cech, C. (forthcoming). Discourse Management Functions in Three Modalities. In Herring, S. (ed.), Computer-Mediated Conversation Analysis.

Condon, S., Cooper, C., and Grotevant, H. 1984. Manual for the Analysis of Family Discourse. Psychological Documents 14.1, Document no. 2616.

Geis, M. 1995. Speech Acts and Conversational Interaction: Toward a Theory of Conversational Competence. Cambridge: Cambridge University Press.

Goffman, I. 1981. Forms of Talk. Philadelphia: University of Pennsylvania Press.

Gordon, D., and Lakoff, G. 1975. Conversational Postulates. In Cole, P. and Morgan, J. (eds.), Syntax and Semantics. Vol. 3: Speech Acts. New York: Academic Press.

Hopper, R. 1989. Speech in Telephone Openings: Emergent Interaction v. Routines. Western Journal of Speech Communication, 53: 178-194.

Labov, W., and Fanshel, D. 1977. Therapeutic Discourse: Psychotherapy as Conversation. New York: Academic Press.

Levinson, B. 1983. Pragmatics. Cambridge: Cambridge University Press.

Minsky, M. 1975. A Framework for Representing Knowledge. In Winston, P. (ed.) The Psychology of Human Vision. New York: Mcgraw-Hill.

Schank, R. 1982. Dynamic Memory. New York: Cambridge University Press.

Schank, R., and Abelson, R. 1977. Scripts, Plans, Goals and Understanding. Hillsdale, New Jersey: Lawrence Erlbaum.

Schegloff, E. 1968. Sequencing in Conversational Openings. American Anthropologist, 70: 1075-1095.

Schegloff, E. 1986. The Routine as Achievement. Human Studies, 9: 111-151.

Searle, J. 1975. Indirect speech acts. In Cole, P. and Morgan, J. (eds.), Syntax and Semantics. Vol. 3: Speech Acts. New York: Academic Press.

Turner, E., and Cullingford, R. 1989. Using Conversation MOPs in Natural Language Interfaces. Discourse Processes, 12: 63-90.