Internet and Higher Education 11 (2008) 87–97

Contents lists available at ScienceDirect

Internet and Higher Education

Interactional coherence in asynchronous learning networks: A rhetorical approach

Andrew PotterSentar, Inc., 4900 University Square Suite 8, Huntsville, Alabama 35816 USA

E-mail address: apotter@sentar.com.

1096-7516/$ – see front matter © 2008 Elsevier Inc. Alldoi:10.1016/j.iheduc.2008.05.001

A B S T R A C T

A R T I C L E I N F O

Article history:

Numerous studies have affirm Accepted 6 May 2008

Keywords:Asynchronous learningDistance educationInteractional coherenceRhetorical structure theory

ed the value of asynchronous online communication as a learning resource. Severalinvestigations, however, have indicated that discussions in asynchronous environments are often neitherinteractive nor coherent. The research reported sought to develop an enhanced understanding of interactionalcoherence, argumentation, and topic drift in asynchronous learning environments. Rhetorical structure theory(RST) was used to analyze and assess the coherence of several asynchronous discussions. Findings include thatasynchronous discussions take the form of dynamic rhetorical structures which are continuously redefined asnewmessages are added to a thread, that argumentationmay bemore prevalent in somediscussions thanothers,that topic drift does not seem to occur as a matter of chance, but rather topics are manipulated to suit theindividual preferences of the participants, and that the use of threading differs considerably from one discussiongroup to another. By demonstrating the applicability of RST, argumentative analysis, and topic drift analysis toasynchronous discussion, this research provides a framework and a terminology for fine-grained analysis ofinteractional coherence. By showing the applicability of RST to asynchronous discussion, this study has offeredevidence that essay assessment technology could be developed for evaluating the quality of online discussions.The development of rhetorical networks as a graph theory for representing the semantics of asynchronousinteraction could lead to a richer knowledge representation technology for inter-agent collaboration.

© 2008 Elsevier Inc. All rights reserved.

1. Introduction

Numerous studies have affirmed the value of asynchronous onlineconferencing as a learning technology. Prominent among these, Harasim(1990) found that the attributes of asynchronous conferencing canbenefit the learning process through a combination of active learningand knowledge building, where active learning is fostered by interactiveand asynchronous aspects of the learning environment, and knowledgebuilding occurs through online idea generation and structuring. Hiltzand Wellman (1997) found that asynchronous environments cansupport communities of learning, enabling students to establishcognitive and emotional ties necessary for effective learning. Blanchette(2001) studied student interaction in asynchronous discussions andfound that the students engaged in higher levels of cognition than thosein face-to-face environments, and their use of language tended to bemore interactive than that of students in face-to-face environments.Rovai (2007) found that effective online learning could be achievedthrough course designs based on constructivist precepts and effectivefacilitation of online discussion, which could be accomplished throughgeneration of social presence, emphasizing student-to-student interac-tions, and giving attention to issues of social equity arising from use ofdifferent communication patterns.

These results were echoed by Fjermestad, Hiltz, and Zhang (2005),who surveyed over 20 studies in learning effectiveness and deter-

rights reserved.

mined that asynchronous learning networks tend to be at least aseffective as traditional teaching methods. More generally, theadvantages of the anytime–anywhere features of asynchronous onlineenvironments have been frequently mentioned in the literature (e.g.Arbaugh, 2004; Chute, 2003; Dalziel, 2003; Doherty, 1998; Dringus &Terrell, 1999; Engelbrecht, 2005; Garrison & Kanuka, 2004; Harasim,Hiltz, Teles, & Turoff, 1995; Kramer, 2001; Phillips & Santoro, 1989;Rovai, 2003; Weller, 2002; Westfall, 2003; Woo & Reeves, 2007). Inshort, the preponderance of evidence suggests that, not only areasynchronous environments conducive to learning, they may besuperior in some respects to traditional classroom pedagogy.

By no means, however, should this be taken to indicate thatasynchronous learning has reached its full potential. Open issuesremain, which, if better understood, could lead to more effectivecommunication in the virtual classroom, improved technologyutilization, and new directions for future technology development.Among these are known problems in sustaining coherence. It is thisarea, called interactional coherence, which was the focus of theresearch reported here.

The term interactional coherence has been used to denote mattersof coherence and incoherence as they pertain to online discussion(Farrell, 2002; Herring, 1999, 2001; Jones, 2000; Jones, Ravid, &Rafaeli, 2001; van der Meij, de Vries, Boersma, Pieters, & Wegerif,2005; Van der Pol, Admiraal, & Simons, 2006). The early investigationsof Henri (1992), Herring, and others suggest that asynchronousdiscussions are too often neither interactive nor coherent. Theaccustomed orderliness of turn-by-turn conversation disappears

88 A. Potter / Internet and Higher Education 11 (2008) 87–97

when participants in a discussion make overlapping exchanges, replyto multiple previous messages within a single message, or simply failto respond at all (Herring, 1999; van der Meij et al., 2005). Discussionsseem to drift aimlessly from one topic to another, without returning tokey points or questions raised earlier (Herring, 1999; Hewitt, 2001;Severinson Eklundh & Rodriguez, 2004). Threads may diverge intonumerous sub-threads, with no prospect for eventual convergence(Hewitt, 2001). Participants routinely ignore the contributions ofothers, so that the resulting transcript reads more like a collection ofmonologues than a discussion (Henri, 1995; Hew & Cheung, 2003;Pena-Shaff & Nicholls, 2004). Clearly, maintaining coherence isproblematic. And yet, despite a wealth of research relating to mattersof coherence in asynchronous discussion, no clear understanding ofwhat is meant by coherence in this context has been articulated.

The research reported here sought to develop an enhanced under-standing of interactional coherence in asynchronous learning environ-ments. Thiswas accomplished using rhetorical structure theory (RST) toanalyze the coherence of a several asynchronous discussions. As will beexplained in detail later in this paper, RST is descriptive theory of textcoherence. RSTassumes that the coherence of a text canbeaccounted forin terms of theway the elements of the text relate to one another (Mann&Thompson,1988). In a coherent text, the relations among the elementsform an integrated abstract structure. This structure may be discoveredthrough a process of RST analysis.

The research used RST to analyze the structural patterns,argumentation, topic drift, and comparative characteristics of threesets of asynchronous discussions created using three differentcomputer conferencing technologies. The analysis revealed that thediscussions were structurally dynamic. While RST structures resultingfrom static documents are acyclic tree-shaped structures, therhetorical networks representing asynchronous threads are frequentlycyclic. Thus, the analysis required a modified form of RST based onrelaxed constraints. By this means, structural models of the discus-sions were created and analyzed.

The following research questions were used to motivate theinvestigation of the applicability of RST to asynchronous discussion,the role and extent of argumentative structures, the structure of topicdrift, and the interplay between interactional coherence and thefeatures of the computer conferencing environment:

1. What rhetorical structure theory modifications, if any, are requiredfor the analysis of asynchronous discussion? The objective of thisresearch questionwas to establish the applicability of RST for analysisof asynchronous discussions. The question left open the possibilitythat the theorymight requiremodification in order to continue to thesubsequent research questions. This was necessary because, whileRST has been applied to a wide range of problems, it had notpreviously beenused for in-depth studyof asynchronous discussions,and there had been no studies, insofar as this researcher was beenable to discover, using RST to study discussions in asynchronouslearning environments.

2. What are the role and extent of argumentative structures inasynchronous discussion? Azar (1999) showed that RST could beused to examine argumentative texts and to distinguish argumen-tative from non-argumentative texts. Azar (1999) found that only afew RST relations should be regarded as argumentative, includingEVIDENCE, MOTIVATION, JUSTIFY, ANTITHESIS, and CONCESSION.What distinguishes these relations is that their loci of effect are inthe nucleus, and further, that the intended effect is to persuade,move, or otherwise influence the reader to accept the content ofthe nucleus. In other words, the satellite provides some impetus foraccepting the nucleus.

3. What are the rhetorical relations or structures of topic drift, andwhat relations are used tomanage it? Hobbs (1990) identified threestrategies that account for topic drift: parallel association, metatalk,and chained explanation. Parallel association occurs when adjacent

text text-spans are associated with one another by virtue of theirsimilarity to one another. Metatalk occurs when one text text-spancomments on another with regard to the objectives of theconversation. Chained explanation is a complex mechanisminvolving a series of interlinked explanations, with each newexplanation displacing the topic of its predecessor, without evergetting back to the original topic. Hobbs (1990) claimed that thesestrategies could account for most topic drift in spoken conversa-tion. The objective of RQ3 was to determine whether the samemaybe said of asynchronous discussion.

4. Do the characteristics of the software used to support asynchro-nous discussions affect the characteristics of interactional coher-ence in asynchronous discussions? Whittaker (2003) and othershave observed that the features offered by a computer conferencingenvironment influence the nature of the interactions occurring inthe environment. The RQ4 investigation examined the rhetoricalstructures used by participants in three different environments inan effort to discover how the features led to differences ininteractional coherence.

2. Related research

The claim that coherence in asynchronous discussion is proble-matic is easily established. Reaching consensus as to precisely whatconstitutes coherence remains difficult. In studies of online commu-nication, two lines of research have predominated. The first of thesederives from Grice's philosophical work in logic and conversation,which led to his famous cooperative principle (Grice, 1975). Thesecond is based on conversation analysis, as defined by Sacks,Schegloff, and Jefferson (1974).

Grice's (1975) cooperative principle and its associated maximshave been used as a standard for assessing spoken conversationalcoherence. The cooperative principle, as defined by Grice, proposesthat participants in a conversation will abide by such norms as arenecessary to ensure that the conversation will achieve its objectives.The maxims that support this notion include admonitions to be asinformative as necessary, but no more so (maxim of quantity); to besincere (maxim of quality); to be relevant (maxim of relevance); andto avoid obscurity, ambiguity, unnecessary wordiness, or disorderli-ness (maxim of clarity). Grice's influence can be found in a variety ofworks in linguistics (Lindblom, 2001; Simner & Pickering, 2005),philosophy (Baccarini, 1991; Neale, 2004), artificial intelligence(Hoadley & Enyedy, 1999; Hulstijn, Dignum, & Dastani, 2004; Kelleher,Costello, & van Genabith, 2005;Walker, 1996), and psychology (Brisch,2002; Kempler, 2004).

As such there should be little surprise that the cooperative principleshould also play a prominent role in defining coherence in onlinecommunication. Herring (1999), in her study of interactional coher-ence in computer-mediated communication, uses Grice's maxims,especially the maxim of relevance, to establish her claims. Pincas(1999) used Grice's maxims as a coherence model in her study ofsequential integrity in asynchronous discussions. Brennan and Ohaeri(1999) invoked the cooperative principle as the basis of theirconceptual framework in their study of rudeness in online discussions.Greenfield and Subrahmanyam (2003), in their study of discourse inchatrooms, base their definition of coherence onGrice's principle. Cechand Condon (2004) use Grice's maxim of relevance in their study ofturn-taking in synchronous computer-mediated communication.Schallert et al. (1996) investigated what they see as a duality incoherence, one being a social activity defined consistently with Grice'sprinciple and the other being a sense-making activity, through whichindividuals respond to discourse interpretively, as theorized by VanDijk (1977) and others. Ho and Swan (2007) analyzed messages in anonline graduate-level English grammar class to determine theirconformance to Grice'smaxims and found that the quality ofmessageswere useful in predicting direct responses to a posting, and that

89A. Potter / Internet and Higher Education 11 (2008) 87–97

students whose messages were rated high in quality and manner alsoreceived higher final grades than their counterparts. Thus Grice'sphilosophical speculations have been adopted as the foundation for avariety of investigations pertaining to interactional coherence.

The second seminal source used in examining asynchronouscoherence is conversation analysis, as defined by Sacks et al. (1974).A primary focus in this area of research is the concept of turn-taking.Turn-taking is viewed as fundamental to human social interaction. Notonly is it used in playing games, but in allocating political office,controlling traffic flows, waiting on customers, and regulating speechexchange systems, such as debates, interviews, meetings, and–importantly–conversations. Because conversation occupies a promi-nent position among speech exchange systems, understanding turn-taking in conversations is viewed essential to understanding thedynamics of speech exchange systems. Turns are treated as a resource,such that conversational participants make seek turns, try to avoidthem, or allocate their turn to some other speaker. Therefore, the turn-taking system may have an economic dimension. If that is so, theorganization and distribution of turn-taking will have effects on theoutcome of the conversation. Since conversation is a central instru-ment in political, scientific, business, and educational discourse, it isimportant that conversation qua instrument be understood.

Moran (1991) cited Sacks et al. (1974), not to identify similaritiesbetween conversation and asynchronous discussion, but to highlight thedifferences. According to Sacks et al., in conversation, only seldom doesmore than one person speak at a time. When two people do findthemselves speaking at the same time, one of them stops abruptly torepair the situation. There is no analog for this in asynchronousdiscussion. Further, in face-to-face conversation, in order to get anopportunity to speak, onemust also listen, at least attentively enough tosegue from one topic to another. However, in asynchronous discussion,there is nothing about the technology or conventions that govern its usethat obliges the participants to read the contributions of others. Aconsequence of this is online discussions tend to be divergent ratherthan convergent.

In a study of turn-taking in synchronous online conversations,Phillips (2000) hypothesized that the classic notion of alternating andorderly turns between dialogue participants is inapplicable for theanalysis of collaborative conversations and an ineffective means forachieving collaborative objectives. Phillips performed a number ofexperiments that indicated that even minor changes in the interac-tional parameters can lead to significant changes in the effectivenessof communication. An implication of this is that little may be inferredfrom face-to-face or synchronous interaction as to what may bereasonably be expected in asynchronous learning environments.

Although the cooperative principle and conversation analysis havebeen predominate in defining the coherence in studies of asynchro-nous discussion, conversational metaphors are not the only resourceavailable to researchers in this area. Crystal (2001) observed thatwhile online communication borrows some characteristics fromconversation, it also has properties from written texts. Therefore,theories of text coherence might provide useful tools for analysis ofasynchronous discussion. Theories of text coherence are used todescribe how the parts of a text are interrelated to produce a wholegreater than the parts. Within this realm, a wide variety of theories,models, and relation sets have been postulated (Hovy & Maier, 1993).Several of these have proven durable, including theories of coherencerelations, cognitive coherence relations, and rhetorical structuretheory. These theories share a common bond: textual coherencederives from the ways in which the elements of a text relate to oneanother, these relationships may be explicitly identified, and a test ofcoherence for any text is the extent to which these interrelationshipsand structures may, upon analysis, be identified. Any text that defiesanalysis is by definition incoherent.

Among the various theories of coherence relations, rhetoricalstructure theory (Mann & Thompson, 1988) is distinctive in several

respects. It provides a well-defined methodology for constructingcomprehensive structures representing an entire text, and it specifiesa rigorous protocol for defining relations, but without making anytheoretic commitment as to what the specific relations should be.Taking a neutral stance on relation set has kept RST free from some ofthe problems that have beset other research (Mann & Thompson,1988). Theories of coherence relations have been seldom used forconversation, much less asynchronous discussions. Rhetorical struc-ture theory has been used to analyze dialogue in a few studies (Stent,2000; Stent & Allen, 2000; Taboada, 2004). To the extent thatasynchronous communication may be viewed as a hybrid mode ofexpression, exhibiting properties of both written text and conversa-tion, theories of coherence relations should prove useful in the studyof interactional coherence.

3. Methodology

For this investigation, rhetorical structure theory was used toanalyze three sets of transcripts, most of which were from asynchro-nous discussions that occurred in masters courses offered at a majoruniversity between 2003 and 2005. An additional transcript, one froma discussion that occurred outside of any formal educational program,was used to further illuminate the subject matter. The RST analysisserved as the basis for investigating argumentation, topic drift, andconferencing system factors in the discussions. The following sectionsprovide an overview of RST, a description of the transcripts, and anexplanation of the analysis.

3.1. Rhetorical structure theory

As a theory of coherence (Mann & Thompson, 1988), rhetoricalstructure theory defines the coherence of a text in terms of the way itsparts, or text-spans, relate to one another. A text-span is either anindividual segment or sentence, or it may be a structure consisting ofseveral segments interrelated by one or more relations. RST postulatesa small number of schemas for defining the possible structuralrelationships among spans and defines a set of rhetorical relations thatmay be used when applying a schema to a set of text-spans.

An RST analysis of a coherent document defines a hierarchicalstructure representing the rhetorical interrelationships of the spanscomprising the text. A text-span may be either an individual segmentor it may be a structure consisting of several segments interrelated byone or more relations. Most relations are binary, consisting of twotext-spans, with one designated as the nucleus and the other as thesatellite. The nucleus is the more salient of the two. The exampleshown in Fig. 1 uses the EVIDENCE relation, where the satelliteprovides information that makes the nucleus more believable.

Schemas are abstract patterns defining one of several possible RSTstructures. In this analysis, four schemas were used: satellite–nucleus,nucleus–satellite, satellite–nucleus–satellite, and multi-nuclear. Thesatellite–nucleus schema describes a binary relation where thesatellite precedes the nucleus. The nucleus–satellite schema describesa binary relation where the nucleus precedes the satellite. Satellite–nucleus–satellite describes a nucleus flanked both sides by satellites.In addition, the multi-nuclear schema describes any relation withmultiple nuclei. The nucleus of a binary relation may participate inmultiple binary schemas. That is to say, the nucleus may havemultiplesatellites. An application of a schema to a text is sometimes called aschema application.

Coherence is defined in terms of four constraints: completeness,connectedness, uniqueness, and adjacency (Mann & Thompson, 1988).The completeness constraint requires that all units in the text beincluded in the structure. Connectedness requires that all units berelated, either directly or by means of nested spans. Uniquenessstipulates that each text-span will be engaged in no more than onerelation. Adjacency requires that for any relation, the nucleus and

Fig. 1. An example RST diagram.

90 A. Potter / Internet and Higher Education 11 (2008) 87–97

satellite text-spans must be adjacent to one another, or that if notadjacent, any intervening text-spans must be satellites of the samenucleus. Thus, a judgment as to the coherency of a text is based onwhether it meets the constraints of completeness, connectedness,uniqueness, and adjacency. A result of this is that a coherent text willform a tree tree-shaped structure, such the one shown in Fig. 1. Inapplying RST to asynchronous discussions, it was necessary to relaxboth the adjacency and the uniqueness constraints. This will bediscussed further in the results section of this paper.

3.2. Transcripts

The principal transcripts used in this analysis were from a course inHuman–Computer Interaction (HCI). The course was part of the corecurriculum in the Master of Science in Management InformationSystems program at Nova Southeastern University. Salient parametersregarding each of the transcripts are summarized in Table 1. Theprogram is offered entirely online; participation in discussions are arequired part of the coursework. The transcripts were archived andderive from two separate offerings of the course. The first sectionincluded discussions held using the Allaire Forums conferencingsystem, and the second section included discussions held using theWebCT conferencing system. To facilitate comparison, the topicsdiscussed in the WebCT transcript were the same as those of theAllaire transcripts, and the same instructor moderated all.

Table 1Transcript parameters

Group Discussion topic Participants Messages Begin End

Allaire HCI Intuitiveness 26 35 02/02/2004 03/18/2004Usability concepts 25 53 01/26/2004 02/18/2004HCI and the Web 26 39 01/22/2004 03/20/2004

WebCT HCI Intuitiveness 24 61 01/31/2005 03/24/2005Usability concepts 20 73 01/24/2005 03/04/2005HCI and the Web 21 62 02/07/2005 03/24/2005

STS STS under attack 60 152 10/03/1994 11/9/1994

In addition to the course transcripts, the investigation wasamplified by a study of an additional transcript, one that occurredoutside a formal educational program. This transcript was from awell-documented asynchronous scholarly debate (Dusek,1998; Hert,1997). The debate took place in 1994 on an email list devoted to thetopic of science, technology, and society (STS). It attracted theattention and participation of numerous noted scholars in the field.Including a discussion of this nature in the research offers theopportunity to discover evidence that the findings of the research areeither unique or not unique to institutionally offered programs—itbecomes possible to gain some broader perspective as to thesignificance of the findings.

3.3. Transcript analysis

The investigation included analysis of both individual message andinter-message rhetorical analysis. The inter-message rhetorical ana-lysis is the focus of the results reported here. The method foridentification of inter-message RST structures proceeded as follows:each message was anonymized and assigned a unique identifier; non-relational information, such as message headers was removed andthreads were identified, and then each thread was analyzed usingRSTTool and Protégé. RSTTool was used for simple interactions, wherethe inter-message relations could be captured in a few diagrams.Protégé was used to develop rhetorical networks representing each ofthe threads in the discussions.

Argumentation was analyzed both in terms of explicit agreementand disagreement and in terms of argumentative structures. For thepurposes of this analysis, an inter-message structure was consideredargumentative if it used the EVIDENCE, MOTIVATION, JUSTIFY,ANTITHESIS, or CONCESSION relations. The basis for this was derivedfrom Azar (1999), who argued that only a few RST relations should beregarded as argumentative. What distinguishes these relations asargumentative is that their loci of effect are in the nucleus, andfurther, that the intended effect is to persuade, move, or otherwiseinfluence the reader to accept the content of the nucleus (see Table 2).In other words, the satellite provides some impetus for accepting thenucleus.

Table 2Argumentative RST relations

EVIDENCE The reader's belief of the nucleus is increased by information providedin the satellite.

ANTITHESIS The situations presented in the nucleus and the satellite are in contrast.Because of an incompatibility that arises from the contrast, one cannothave positive regard for both the situations presented in the nucleus andsatellite; comprehending the satellite and the incompatibility betweenthe situations increases reader's positive regard for the situationpresented in nucleus

JUSTIFY Comprehension of the satellite increases the reader's readiness to acceptwriter's right to present nucleus

MOTIVATION Comprehending the satellite increases the reader's inclination toperform action presented in the nucleus.

CONCESSION The writer acknowledges a potential or apparent incompatibilitybetween the situations presented in the nucleus and the satellite; thewriter regards the situations presented in the nucleus and satellite ascompatible; recognizing that the compatibility between the situationsin the nucleus and the satellite increases reader's positive regard for thesituation presented in the nucleus.

91A. Potter / Internet and Higher Education 11 (2008) 87–97

Topic drift, according to Hobbs (1990), occurs incrementally,through a series of minor modifications to the topic. Takenindividually these modifications are not necessarily problematic forcoherence, and in fact they rely on the same structures used in fullycoherent texts to maintain coherence and enrich communication(Hobbs, 1990; Lenk, 1998; O'Donnell, 2000). What distinguishes topicdrift, however, is that these relations are engaged without return tothe previous topic of discourse (Hobbs, 1990). As the name suggests,the topic drifts with no prospect for recovery. Hobbs identified threedevices that account for topic drift. These are parallel association,metatalk, and chained explanation.

Parallel association occurs between text-spans when the spans arerelated tangentially to one another. Parallel association uses amechanism Hobbs (1990) called discourse pivot. A discourse pivotforms a link between two otherwise unrelated topics. Discourse pivotincorporates some associations in the preceding text with those of theemergent topic, thus smoothing the transition from one topic toanother. Parallel association is similar to the RST LIST multi-nuclearrelation, which consists of two or more comparable text-spans, exceptthat it occurs across multiple conversational turns or asynchronousmessages. Other possible counterparts are the CONTRAST andANTITHESIS relations, in which there is some basis for comparison,but, in other respects, the differences override the similarities.Metatalk occurs when one text-span comments on another regardingthe objectives of the conversation (Hobbs, 1990). When this happens,the topic may shift to become a discussion about the discussion. Themain RST counterpart of metatalk is the EVALUATION relation, inwhich the satellite text-span assesses the situation presented innucleus text-span. However, metatalk is distinctive in that it assessesnot the content, but the form or process of the evaluated text-span.Chained explanation is a complex mechanism involving a series ofinterlinked explanations, with each new explanation displacing thetopic of its predecessor (Hobbs, 1990). Chained explanations mayoccur using a variety of relations in RST, such as ELABORATION,EVIDENCE, and INTERPRETATION. It may also incorporate elements ofparallel association and metatalk. Through a sequence of text-spanslinked recursively by these relations, the topic may rapidly shift towhere it has no relevance to its original subject.

Hobbs (1990) argued that parallel association, metatalk, andchained explanation account for most topic drift in conversation.These strategies permit speakers to alter the topic of conversationwithout resorting to overt breaks in continuity. An expectation for thisinvestigation was that the devices of topic drift in asynchronousdiscussions would be similar to those of spoken conversation asidentified by Hobbs (1990). That is, drift in asynchronous discussionwould be describable in terms of parallel association, metatalk, andchained explanation. To the extent that this is the case, it could be

askedwhether the phenomenon is problematic: topic drift is commonin conversation, with little if any harmful effect. However, in anasynchronous learning environment, students rely on asynchronousdiscussions to achieve their learning objectives. The delayed turn-around and reduced social presence in message exchange makesrecovery from topic drift difficult and sometimes unachievable(Harasim, 1990; Whittaker, Bellotti, & Gwizdka, 2006).

Finally, the comparison of interactional coherence in differingconference systems built on the earlier analysis of rhetorical structure,argumentation, and topic drift by considering whether there weredifferences in coherence that could be accounted for by differences inthe conference systems used. Whittaker (2003) and others haveobserved that the features of a computer conferencing environmentwill influence the nature of the interaction. Features of threadmanagement, for example, differ from one conferencing system toanother, and in systems lacking thread support, participants resort tovarious forms of reference in order to maintain the integrity of thediscussion (Kear, 2001; Pincas, 1999; Preece, 2000; Reed, 2001).

3.4. Barriers and limitations

The theoretical nature of the study imposed numerous barriersand limitations on the study. There were several issues associatedwith the use of rhetorical structure theory. These include the partialnature of RST as a theory of coherence, the role of subjectivity instructural analysis, and the possibility of multiple analyses for a giventext.While these issues do not invalidate RSTas a theoretical tool, theyimpose limitations on the certitude of any conclusion reached. Moresignificant to this dissertation research is the applicability of RST toasynchronous discussion: RST was designed for use with monologue,not dialogue or discussion. The following sections discuss these issuesin detail.

3.4.1. RST as a partial theory of coherenceRST formulates coherence as the ability to account for the presence

of the elements of a discourse by providing a plausible description ofthe relational structure of these elements (Mann & Taboada, 2005).While this yields a comprehensive model of a rhetorical structure,there are other aspects of coherence that RST does not address. It doesnot, for example, address the syntactic characteristics of coherence,developmental order, or holistic coherence (Mann, Matthiessen, &Thompson, 1992).

3.4.2. Subjective judgmentSubjective judgment is a necessary part of RST methodology

(Mann et al., 1992). For a text to be judged coherent, judgments aboutthe functions of the parts of a text and their relations are a necessarypart of the analysis. To this extent, RST relies on the analyst'sunderstanding of the language, culture, and subject matter of the text(Mann & Thompson, 1987). It is the claim of RST that such judgmentsare plausible rather than certain. The judgments comprising ananalysis achieve credibility by means of their internal cohesion—thatis, the structure arising from an analysis is essentially a localizedtheory of the text under analysis (Mann & Thompson, 1987, 1988;Mann et al., 1992; Moore & Wiemer-Hastings, 2003). As such, RST is amethodology for generating theory.

3.4.3. Differences in analysisDifferences in analysis in RST are attributable to a number of

sources, including boundary judgments, structural ambiguities,simultaneous analyses, differences between analysts, and analyticalerror (Mann & Thompson, 1987). Boundary judgments are aninevitable consequence of having to choose from among categories;borderline cases must be resolved in order for the analysis to proceed.Similarly, structural ambiguities are an inevitable and normal part ofRST because the language itself contains ambiguities. To insist

Fig. 2. A rhetorical network.

92 A. Potter / Internet and Higher Education 11 (2008) 87–97

otherwise would be to demand of the text greater precision than itcontains (Mann & Thompson, 1987). Simultaneous analyses occurwhen multiple relations are applicable to a single pair of text units.Unlike ambiguity, where the intended meaning may not be dis-cernible, simultaneous analyses occur when the rhetorical intentseems clear, but the intent seems to involve two dissimilar relations(Mann & Thompson, 1987).

Given the susceptibility of RST to differences in boundary judgments,structural ambiguities, and simultaneous analyses, analytical discre-pancies may occur. Even so, Mann and Thompson claim thatdiscrepancies occur infrequently (Mann & Taboada, 2005; Mann &Thompson, 1987; Mann et al., 1992), and this claim is supported by theliterature. den Ouden, van Wijk, Terken, and Noordman (1998) studiedthe reliability of segmentation and structuring of relations and found ahigh degree of consistency among the analysts studied. Marcu,Amorrortu, and Romera (1999) developed a statistical method formeasuring agreement in rhetorical structures and found that analystsachieved high levels of agreement in defining structures. Theirinvestigation also suggested that divergence was more likely to occurwhen the analysts were unfamiliar with the subject matter of the text.Similarly, den Ouden (2004) found high levels of reliability relative toother structural analysis methods and suggested that this may be bestaccounted for by the explicitness of the definitions used and the laborintensive nature of the analysis.

Finally, there is the issue of analytical error. While the possibility oferror can never be ruled out, the occurrence of analytical error decreaseswith experience (O'Brien, 1995), and the rigorous definitional basisdefined for RST makes errors less likely, so long as the methodology isstrictly observed (Mann & Taboada, 2005; Mann & Thompson, 1988;O'Brien,1995; Taboada, 2004). The definition of each relation includes aset of constraints, and these constraints define not only the relationshipbetween the nucleus and its satellite, they also place constraints on thenucleus itself (Mann,1984). These constraints served as signposts to theanalyst, reducing, but not eliminating, the likelihood of analytical error.

3.5. Delimitations

The study was based on the analysis of transcripts of two selectedasynchronous discussions that occurred in masters courses offered atNova Southeastern University Graduate School of Computer andInformation Science between 2003 and 2005. Additional publiclyavailable transcripts were used to support the study. Delimitationsinclude the following:

1. The asynchronous discussions studied were from courses that weretaught entirely online using a combination of online conferencing,email, andWeb-based resources. The generality of the results of thestudy is limited to comparable courses offered in an online format.

2. The conferencing software used in the study included AllaireForums andWebCT. It was anticipated as part of this study that thefeatures of these products would affect the characteristics ofinteractional coherence in asynchronous discussions. To the extentthat this was the case, it also follows that discussions taking placeusing other products may have unique characteristics. Therefore,the generality of the results may be limited to discussions using theAllaire Forums and WebCT products.

3. The analytical approach used was fundamentally theoretical, andtherefore the conclusions reached are of a plausible, not definitivenature. As described in the section on barriers and limitations, anRST analysis is a plausible explanation for the relational structure ofa set of discourse elements (Mann & Taboada, 2005). RST is amethodology for generating theory. Any inferences drawn from anRST analysis are thereby qualified.

4. The reliability of the results reached in this studywas limited towhatcould be provided by tools used. As described in the section onbarriers and limitations, several studies foundhigh levels of reliability

with the RSTmethodology (denOuden, 2004; denOuden et al.,1998;Marcu et al., 1999). However, no direct measure of reliability wasincorporated in this study. Again, the level of certitude applicable tothe results of this study is of a theoretical nature.

5. The study relied on RST for its validity. Although RST continues tobe used by many investigators (den Ouden, 2004; Taboada, 2004;Taboada & Mann, 2006a,b; Wolf & Gibson, 2005), not allresearchers embrace the theory. For example, Knott and Dale(1994) and Kehler (2002) have criticized RST for failing to provide adefinitive set of relations.

4. Results

The number of threads per discussion ranged between 6 and 14, andthe average number of interactions per thread ranged fromas fewas 1 toas many as 90. Not only did thread sizes differ, but also the tendency ofparticipants to engage in threaded interactions varied from group togroup. In the Allaire Intuitiveness group, less than one-third of themessages belonged to a thread; in the STS debate, almost all messageswerepart of a thread.Onlya small subsetof RST relationswasused in theinteractions. The most frequently used relations were ELABORATION,ANTITHESIS, CONCESSION, EVALUATION, and EVIDENCE. Elaborationwas the preferred relation. The frequent use of EVALUATION suggeststhat messages might be likely to pass judgment on one another. ThatELABORATIONwas themost frequently used relation is not, from anRSTperspective, surprising. This is consistent with previous RST research(Mann et al., 1992). The prevalence of ELABORATION suggests an overalllevel of discourse more constructive than contentious.

More fundamental than patterns of relation usage were thedynamics of inter-message structures. The rhetorical structure evolvedand took shape over the course of the discussion. Although, at the basiclevel of analysis, any given interaction between two messages could bemodeled using RST, thiswas not the case for complex threads. This is notconsistent with conventional rhetorical structure theory, where RSTstructures resulting from static documents are acyclic tree-shapedstructures; the rhetorical networks representing asynchronous threadswere frequently cyclic. This is possible because in an asynchronousenvironment, anymessagemay respond to anyof its predecessors at anytime, and it may do so using any rhetorical relation, without regard forany preexistent structural commitments.

Fig. 2 shows an example of a rhetorical network. In this graph theinitial message (M37) has elicited several responses. The first of theseadopted a concessionary relationship to the original message. Hereone can see that the satellite–nucleus schema has been applied,resulting in the original message being satellite to the response. Thisfirst response was followed by another (M39), one that summarizedthe two preceding messages. This second response was a nucleus totwo satellites, and in both cases the nucleus–satellite schema has been

Fig. 4. An EVIDENCE interaction in the Allaire intuitiveness discussion.

Fig. 3. An argument in the STS discussion.

93A. Potter / Internet and Higher Education 11 (2008) 87–97

applied. The final response of the thread (M40) evaluated the originalmessage, and it too used the nucleus–satellite schema.

The structural dynamics of a thread could take two forms:convergent and divergent. In a convergent structure, a messageresponded to multiple predecessors. Because the respondent messagecould employ any RST relation to any predecessor, including relationsusing a nucleus–satellite schema, the respondent message could be asatellite to multiple predecessors. In a divergent structure, a messagereceived multiple responses. Here again, the responses could employany RST relation. When the responses used relations based on satellite–nucleus schema, the original message then became satellite to multipleresponses. Since the RST uniqueness constraint permits a satellite tohave only one nucleus, such convergent and divergent structures werenot conformant to RST. In addition, the analysis found that any messagecould refer to any previous message, regardless of the presence of otherintervening messages. This resulted in the loss of structural adjacency.When combined with non-uniqueness, the discussions could not berepresented using conventional RST diagrams. This realization moti-vated the development of rhetorical networks.

4.1. Argumentation

Inter-message argumentative structures used only ANTITHESIS,CONCESSION, and EVIDENCE argumentative relations. The JUSTIFYand MOTIVATION relations were not used. ANTITHESIS was the mostfrequently used relation, followed closely by CONCESSION andEVIDENCE. Much of the use of ANTITHESIS and CONCESSION occurredin the STS group. These relations accounted for about 85% ofargumentative interactions within the STS group. EVIDENCE was thepreferred inter-message argumentative relation in the Allaire group,while EVIDENCE and CONCESSIONwere the same in theWebCTgroup.

Since the ANTITHESIS and CONCESSION relations were used forexpressing disagreement whereas EVIDENCE was useful for expres-sing agreement, it would seem the dominant mode of interaction inthe STS discussion is one of disagreement—or as Hert rathermildly putit, the discussion revealed “a heterogeneity of goals among theparticipants” (Hert, 1997, p. 329).

Fig. 3 shows an example of an argument in the STS discussion. Thisexample is taken from a larger thread of 54 messages. As can be seen

in the rhetorical network, the epicenter of the argument is messageSTS-OCT-M92-Fuller, which initiated the dispute by taking exceptionto four previously posted messages and provoked a variety ofresponses, all of them negative.

Half of the inter-message relations in this argument wereargumentative. Of these, Antithesis was the relation of choice,followed by CONCESSION, and trailed by EVIDENCE. Interestinglythough, EVALUATION and ELABORATION were used as often asANTITHESIS and CONCESSION. The presence of EVALUATION andELABORATION was due mainly to their use the messages posted insupport of Traweek. These messages formed an area of agreementwithin the overall argument that was largely unconnected from theother messages. Heath, Stockdale, and Reid provided encouragementto Traweek, who did battle against Fuller. However, it would be amistake to assume that the use of these relations signaled agreement,their agreement was on their disagreement with other unspecifiedmessages in the thread. Thus engaged in metatalk, these messages arean example of topic drift, which will be discussed later.

In the WebCT and Allaire asynchronous learning environments,inter-message argumentationwas generally constructive and agreeable,tending to rely on the EVIDENCE relation. An illustrative example of thiswas found in theAllaire Intuitiveness thread. Theoverall structure of thisshort thread is shown in Fig. 4. The thread consisted of a single nucleusfollowed by two satellites. Both satellites related to the nucleus using theEVIDENCE relation. The nucleus argued that intuitiveness involvedknowing what to do without being given instruction:

I think intuitiveness can be defined as the next logical step to takewithout further instruction. This makes a website or softwareprogram easy to follow and learn. Nothing is easier than to navigate

94 A. Potter / Internet and Higher Education 11 (2008) 87–97

in a program/website that is built with intuitiveness in mind; ofcourse, what is logical to some might not be logical to others.

The next two messages followed up on this claim with additionalevidence. The first response argued that knowing what to do withoutbeing given instruction would come more easily if it could be relatedto a frame of reference. The next response expanded the themefurther, and argued for the elimination of other possible interpreta-tions of what intuitiveness is. Thus, in contrast to the STS discussion,this example suggests that it is possible to interact argumentativelywhile remaining constructive and agreeable.

However, not all argumentation in the Allaire and WebCT groupswas agreeable. There were several inter-message disagreements,usually employing the CONCESSION relation. For example, in onethread in the WebCT Usability discussion, student P31 identifiedseveral elements of user interface learnability, and noted that whenbroken down into its constituents, learnability was easier to evaluate:

Learnability is a critical attribute for a new interface or application.Today's user doesn't have time to waste trying to learn the ins andouts of a new system. Preece sees learnability as a measure of howeasy a system is to learn to use (2002). Learnability may further bedescribed in quantifiable areas such as familiarity, consistency,predictability, simplicity, and generalizability (Usability 101, 2003).Familiarity is the extent to which a piece of software or interfacebuilds upon the user's prior knowledge of similiar applications orinterfaces. Consistency points to the attribute of only having to learna taskwithin the application once and allows the user to perform thesame task in the same manner every time. Breaking downlearnability into sub areas makes it easier to evaluate.

P37 responded to this by acknowledging the value of thelearnability factors, but asserted that the “Ten-Minute Rule” waspreferable because if a user cannot figure out to use an application in10 min, he or she would be unlikely to be satisfied with theapplication:

While I agree with the additional factors that you include inunderstanding learnability, I like the Ten-Minute Rule described inour text. Unless the application is complex, the user should be ableto learn how to use the system in under 10 minutes. Frankly, I findthat I have behaved much that way in my own experience. Unless Iwant to sit down and become an expert I prefer to see progressand some results; fast….

Fig. 5. Lateral associat

The instructor followed up, summarizing the Ten-Minute Rule asan ideal for providing users with rapid and effortless functionality.There were further elaborations in this thread, but no furtherargumentation. This was typical of the learning environments. Unlikeargumentation in the STS discussion, argumentative interactions werenot sustained. In arguments involving disagreement, the threads wereshort-lived.

4.2. Topic drift

Topic drift does not seem to occur as a matter of chance. Parti-cipants use topic drift devices to adapt discussion to a preferred topic.To this extent, topics do not drift so much as they are pushed andpulled. A consequence of this is that threads often begin with a strongresearch-based opening message, but descend to anecdotes andpersonal commentary.

Just as Hobbs (1990) found in spoken conversation, three devicesaccounted for topic drift in asynchronous discussion: parallelassociation, metatalk, and chained explanation. Using parallel asso-ciation, participants leveraged previous discussion as opportunitiesfor posting messages about favorite subjects. The analysis revealedthat participants accomplished this using several types of parallelassociation, including lateral association, subtopic escalation, pedago-gical pivot, and redirection.

Lateral association is an association between the main topic of amessage and its response. In subtopic escalation, the respondent to amessage responds to a subtopic within the previous message, withoutacknowledging the primary topic. Pedagogical pivot entails a deliberateintervention by the instructor to shift the topic into alignment withlearning objectives. In topic redirection, the respondent dismisses thepreviousmessage andproposes anewapproach. Redirection is similar topedagogical pivot, except that thewriter carrying it out is not the courseinstructor. Redirectionwas seen only in the STS discussion. Examples ofeach of these subcategories are given in Fig. 5. In the discussions studied,subtopic escalation occurred in instances of parallel association, chainedexplanation, and metatalk. Lateral association and redirection werefound in instances of parallel association and chained explanation.Pedagogical pivot occurred only in parallel association.

Chained explanations commonly used subtopic escalation.Responses focused on explaining a subtopic within a previousmessage, and this subtopic would then become subject to a series ofchained explanations. An example of this occurred in the WebCTUsability discussion, shown in Fig. 6.

ion subcategories.

Fig. 6. Subtopic escalation in chained explanation.

95A. Potter / Internet and Higher Education 11 (2008) 87–97

In the first message of this thread, student P46 student offered apersonal view of user interface flexibility, and then corrected this witha more detailed definition based on research. This was amplified withan example based on the Microsoft Paint application and theobservation about the flexibility of Microsoft Windows applicationsin general.

Responding to these observations, student P37 challenged thenotion thatWindows programs are flexible, citing as examples of earlyversions of the Microsoft FrontPage Web authoring product. Thisresponse ignored the central ideas of the previous message andfocuses entirely on the example. The next message continuedwith thediscussion of Microsoft FrontPage, as did the next two responses.Finally, the instructor commented on the original message, in amanner that recovered the original topic. However, there were nofurther contributions to the thread. This is consistent with the notionthat discussion participants seek to manipulate the topic to areas theyare comfortable with. Discussing the shortcomings of applicationsthey were familiar with was easy, but developing the concept offlexibility would have been challenging.

There were no instances of metatalk in the Allaire and WebCTdiscussions. Metatalk was used several times in the STS discussion,usually to voice disagreement with ongoing discussion or to expresssolidarity with others who were in disagreement. Topic recovery wasused several times in the STS discussion, but was seldom used in theAllaire andWebCT discussions. The salient relations for topic recoverywere ANTITHESIS, CONCESSION, and ELABORATION. When used withANTITHESIS and CONCESSION, topic recovery expressed dissatisfac-tion with the current topic.

4.3. Conference system comparisons

The conferencing systems used for the discussions were similar intheir features, but the discussions differed, particularly in their use ofthreading. In one group, less than half of the messages were threaded,with the remainder posted as singletons. In other groups most of themessages were in threads.

The comparative study examined the rhetorical structures used byparticipants in the Allaire andWebCTconferencing environments. RSTrelation use in the Allaire and WebCT messages was similar. In bothforums, Elaboration was most commonly used. Other frequently usedRST relations were Concession, Background, and Evidence. The overalluse of argumentative structures in individual messages was roughlyequivalent between the Allaire and WebCT groups. In both of thesegroups, the argumentative relations most frequently used wereAntithesis, Concession, and Evidence. Although the Allaire group

tended to be more argumentative than the WebCT group, the inter-message argumentative relation most frequently used in Allaire wasEvidence. In theWebCT group, the use of argumentative relations wasroughly evenly distributed between Evidence and Concession.

In the Allaire group, less than half of the messages belonged to athread, with the remainder being posted as singletons. In contrast,most of themessages in theWebCTgroupwere in threads. Further, theAllaire threads were quite short, averaging about three messages inlength, with the longest thread containing 10 messages. Many of thethreads consisted of a single interaction between two messages. TheWebCT threads were somewhat longer, with an average of fivemessages per thread.

Although the threads in the Allaire group were relatively short,there were occurrences of topic drift. Over half of these were chainedexplanations, and one-third were parallel associations. The WebCTgroup, however, preferred parallel association; only about one-third ofits topic drift devices were chained explanations. Both groups madeextensive use of subtopic escalation. Pedagogical pivot occurred morefrequently in the Allaire group than in the WebCT group. The Allairegroup seemed to compensate for lower levels of threaded interactionby means of informal intertextuality.

In the Allaire and WebCT discussions, the depth of reference rarelyexceeded more than one message. The only occasions when the depthexceeded this were in messages from the instructor, in which shesought to elaborate, evaluate, or summarize previous discussion.

The study also considered overlapping threads as a possible sourceof interactional incoherence. Overlapping threads occur when themessages comprising multiple threads are intermixed with oneanother in their presentation to the user, such that the user is left todistinguish which message responds to which (Herring, 1999; Pincas,1999). Because both Allaire andWebCT provide strong thread support,messages are organized as threaded structures. Consequently, therewere no instances of overlapping threads in the two conferencingsystems.

5. Conclusion

This research sought to develop an enhanced understanding ofinteractional coherence in asynchronous learning environments.Rhetorical structure theory (RST) was used to analyze discussionsfrom two computer conferencing systems and an email debate. Theconferencing systems were Allaire Forums and WebCT. The emaildebate took place on a list devoted to the topic of science, technology,and society (STS). The research included an assessment of theapplicability of RST for analysis of asynchronous discussions, an

96 A. Potter / Internet and Higher Education 11 (2008) 87–97

examination of the use of argumentative rhetorical relations inasynchronous discussions, an analysis of topic drift, and a comparativestudy of interactional coherence in the WebCT and Allaire computerconferencing systems.

The application of RST to asynchronous discussions showed thatdiscussions evolve and take shape onmultiple levels. At the basic levelof analysis, any given interaction between two messages can bemodeled using RST, and under some circumstances extended threads,consisting of a series of interactions, may conform to the constraints ofrhetorical structure theory. However, at a more complex level thethread structures presented significant challenges. Principally, this isbecause, for any given interaction, the incipient structure is at thediscretion of the respondent, without regard for any preexistentstructural commitments, and, moreover, any given message may belinked to any other message, provided the two messages were notcomposed concurrently. A message may at anytime be coerced intobecoming a satellite to some new message.

The dynamic character of discussions has important implicationsfor the application of RST. A single instance of coercion, where onemessage is coerced by another into becoming a satellite, is well withinthe RST constraints of completeness, connectedness, uniqueness, andadjacency. However, a problem arises when multiple messages coercesome other message into becoming a satellite. In this case, the coercedmessage becomes satellite to multiple nuclei. This violates theprinciple of uniqueness. Consequently, for the inter-message analysisit was necessary to relax the uniqueness requirement. In addition, anymessage may refer to any previous message, regardless of thepresence of other intervening messages. A result of this is a loss ofstructural adjacency. When combined with non-uniqueness, thediscussions cannot be represented using conventional RST diagrams.This realization motivated the development of a variant of RST calledrhetorical networks.

The investigation of argumentation included an identification ofargumentative structures, an examination of these structures andtheir dynamics, and comparison of argumentation patterns in thediscussion groups. Inter-message argumentative structures used onlyANTITHESIS, CONCESSION, and EVIDENCE argumentative relations.ANTITHESIS and CONCESSION was used extensively in the STS group,suggesting the dominant mode of interaction in the STS discussion isone of disagreement. EVIDENCE was the preferred inter-messageargumentative relation in the Allaire group, while EVIDENCE andCONCESSION were evenly distributed in the WebCT group. In the STSdiscussions, arguments were sustained over large numbers ofmessages and involved numerous participants. Disagreements in theAllaire and WebCT arguments involved only a few participants andextended for only a few interactions.

The investigation of topic drift sought to determine whether thedevices of topic drift in asynchronous discussions are similar to thoseof spoken conversation. The topic drift devices included parallelassociation, chained explanation, and metatalk as defined by Hobbs(1990). The investigation considered whether the use of these deviceswould manifest themselves in RST analysis. The analysis suggestedthat topic drift does not occur as a matter of chance. Participants usedthe devices of topic drift to adapt the discussion to a topic of pre-ference. Using parallel association and chained explanation, partici-pants leveraged previous messages as opportunities for postingmessages about favorite subjects.

The comparative study examined the rhetorical structures used byparticipants in the Allaire and WebCT environments in an effort todiscover how differences in these systems lead to differences ininteractional coherence. RST relation use in the Allaire and WebCTmessages was similar. The overall use of argumentative structures inindividual messages was also roughly equivalent. Although the Allairegroup tended to be more argumentative than the WebCT group, theinter-message argumentative relation most frequently used in Allairewas EVIDENCE, indicating that the argumentation tended to be

supportive rather than disputative. In both the Allaire and WebCTgroups, threads tended to be short, with many threads consisting of asingle interaction of two messages.

This research has yielded important implications for learningtheory, natural language processing, and knowledge representation.By demonstrating the applicability of RST, argumentative analysis, andtopic drift analysis to asynchronous discussion, this research providesa framework and a terminology for fine-grained analysis of interac-tional coherence. By showing the applicability of RST to asynchronousdiscussion, this study has offered evidence that essay assessmenttechnology could be developed for evaluating the quality of onlinediscussions. The development of rhetorical networks as a graph theoryfor representing the semantics of asynchronous interaction could leadto a richer knowledge representation technology for inter-agentcollaboration.

These implications have, in turn, identified new directions forfuture research. The insights in intertextuality, argumentation, topicdrift, and the structural dynamics of asynchronous discussion indicatethat additional research is needed in the theory and practice ofasynchronous discussion strategy. In NLP, further research is necessaryto develop the ability to apply essay assessment technology toasynchronous discussions. Additional research is needed to refineRSTas a knowledge representation technology for use by collaborativemulti-agent systems.

References

Arbaugh, J. B. (2004). Learning to learn online: A study of perceptual changes betweenmultiple online course experiences. The Internet and Higher Education, 7(3),169−182.

Azar, M. (1999). Argumentative text as rhetorical structure: An application of rhetoricalstructure theory. Argumentation, 13(1), 97−114.

Baccarini, E. (1991). Rational consensus and coherence methods in ethics. GrazerPhilosophische Studien, 40, 151−159.

Blanchette, J. (2001). Questions in the online learning environment. The Journal ofDistance Education, 16(2), 37−57.

Brennan, S. E., & Ohaeri, J. O. (1999). Why do electronic conversations seem less polite?The costs and benefits of hedging. Proceedings, International Joint Conference onWork Activities, Coordination, and Collaboration (WACC '99) (pp. 227−235). NewYork: ACM Press.

Brisch, K. H. (2002). Treating attachment disorders: From theory to therapy (K. Kronenberg,Trans.). New York: Guilford Press.

Cech, C.G., & Condon, S. L. (2004, January 5–8). Temporal properties of turn-taking andturn-packaging in computer-mediated communication. Paper presented at the37th Hawaii International Conference on System Sciences (HICSS'04), Big Island,Hawaii.

Chute, A. G. (2003). From teletraining to e-learning and knowledge management. In M. G.Moore, &W. Anderson (Eds.),Handbook of distance education (pp. 297−314). Mahwah,NJ: Lawrence Erlbaum Associates.

Crystal, D. (2001). Language and the Internet. Cambridge: Cambridge University Press.Dalziel, C. (2003). Community colleges and distance education. In M. G. Moore, & W.

Anderson (Eds.), Handbook of distance education (pp. 663−672). Mahwah, NJ:Lawrence Erlbaum Associates.

den Ouden, H. J. N. (2004). Prosodic realizations of text structure. Unpublished doctoraldissertation, University of Tilberg, Tilberg, the Netherlands.

den Ouden, H. J. N., van Wijk, C. H., Terken, J. M. B., & Noordman, L. G. M. (1998).Reliability of discourse structure annotation. IPO Annual Progress Report, 33,129−138.

Doherty, P. B. (1998). Learner control in asynchronous learningenvironments.ALNMagazine,2(2) Retrieved October 25, 2005 from http://www.sloan-c.org/publications/magazine/v2n2/doherty.asp

Dringus, L. P., & Terrell, S. (1999). The framework for DIRECTED online learningenvironments. The Internet and Higher Education, 2(1), 55−67.

Dusek, V. (1998). Where learned armies clash by night. Continental Philosophy Review,31(1), 95−106.

Engelbrecht, E. (2005). Adapting to changing expectations: Post-graduate students'experience of an e-learning tax program. Computers & Education, 45(2), 217−229.

Farrell, R. (2002). Summarizing electronic discourse. International Journal of IntelligentSystems in Accounting, Finance and Management, 11(1), 23−38.

Fjermestad, J., Hiltz, S. R., & Zhang, Y. (2005). Effectiveness for students: Comparisons of“in-seat” and ALN courses. In S. R. Hiltz, & R. Goldman (Eds.), Learning togetheronline: Research on Asychronous networks (pp. 39−80). New York: LawrenceErlbaum.

Garrison, D. R., & Kanuka, H. (2004). Blended learning: Uncovering its transformativepotential in higher education. The Internet and Higher Education, 7(2), 95−105.

Greenfield, P. M., & Subrahmanyam, K. (2003). Online discourse in a teen chatroom:New codes and new modes of coherence in a visual medium. Journal of AppliedDevelopmental Psychology, 24(6), 713−738.

97A. Potter / Internet and Higher Education 11 (2008) 87–97

Grice, H. P. (1975). Logic and conversation. In P. Cole, & J. Morgan (Eds.), Syntax andsemantics (pp. 41−58). New York: Academic Press.

Harasim, L. (1990). Online education: An environment for collaboration and intellectualamplification. In L. Harasim (Ed.), Online education: Perspectives on a newenvironment (pp. 39−64). New York: Praeger.

Harasim, L., Hiltz, S. R., Teles, L., & Turoff, M. (1995). Learning networks: A field guide toteaching and learning. Cambridge, MA: MIT Press.

Henri, F. (1992). Computer conferencing and content analysis. In A. Kaye (Ed.), Colla-borative learning through computer conferencing: The Najaden papers (pp. 117−136).Berlin: Springer-Verlag.

Henri, F. (1995). Distance learning and computer-mediated communication: Interactive,quasi-interactive, or monologue? In C. O'Malley (Ed.), Computer supportedcollaborative learning (pp. 145−161). Berlin: Springer-Verlag.

Herring, S. C. (1999). Interactional coherence in CMC.Journal of Computer-MediatedCommunication, 4(4) Retrieved January 9, 2004 from http://www.ascusc.org/jcmc/vol4/issue4/herring.html

Herring, S. C. (2001). Computer-mediated discourse. In D. Schriffin, D. Tannen, & H. E.Hamilton (Eds.), The handbook of discourse analysis (pp. 612−634). Malden, MA:Blackwell Publishing.

Hert, P. (1997). The dynamics of on-line interactions in a scholarly debate. TheInformation Society, 13(4), 329−360.

Hew, K. F., & Cheung,W. S. (2003). Evaluating the participation and quality of thinking ofpre-service teachers in an asynchronous online discussion environment: Part I.International Journal of Instructional Media, 30(3), 247−262.

Hewitt, J. (2001). Beyond threaded discourse. International Journal of EducationalTelecommunications, 7(3), 207−221.

Hiltz, S. R., & Wellman, B. (1997). Asynchronous learning networks as a virtualclassroom. Communications of the ACM, 40(9), 44−49.

Ho, C. -H., & Swan, K. (2007). Evaluating online conversation in an asynchronouslearning environment: An application of Grice's cooperative principle. The Internetand Higher Education, 10(1), 3−14.

Hoadley, C. M., & Enyedy, N. (1999). Between information and communication: Middlespaces in computer media for learning. Proceedings of the Computer Support forCollaborative Learning (CSCL) 1999 Conference (pp. 242−251). Mahwah, NJ: LawrenceErlbaum Associates.

Hobbs, J. R. (1990). Topic drift. In B. Dorval (Ed.), Conversational organization and itsdevelopment (pp. 3−22). Norwood, NJ: Ablex Publishing.

Hovy, E., & Maier, E. (1993). Parsimonious or profligate: How many and which discourserelations?Marina del Rey, CA: Information Sciences Institute, University of SouthernCalifornia.

Hulstijn, J., Dignum, F., & Dastani, M. (2004, July 19). Coherence constraints for agentinteraction. Paper presented at the AAMAS 2004 Workshop on Agent Communica-tion (AC2004), Columbia University, NY.

Jones, Q. (2000). An empirical investigation of boundaries to virtual public discoursestructure.CHI '00 extendedabstracts onhuman factors in computing systems (pp. 69−70).New York: ACM Press.

Jones, Q., Ravid, G., & Rafaeli, S. (2001). Empirical evidence for information overload inmass interaction. CHI '01 extended abstracts on human factors in computing systemsNew York: ACM Press.

Kear, K. (2001). Following the thread in computer conferences. Computers & Education,37(1), 81−99.

Kehler, A. (2002). Coherence, reference, and the theory of grammar. Stanford, CA: CSLIPublications.

Kelleher, J., Costello, F., & van Genabith, J. (2005). Dynamically structuring, updating andinterrelating representations of visual and linguistic discourse context. ArtificialIntelligence, 167(1–2), 62−102.

Kempler, D. (2004). Neurocognitive disorders in aging. Thousand Oaks, CA: SagePublications.

Knott, A., & Dale, R. (1994). Using linguistic phenomena to motivate a set of coherencerelations. Discourse Processes, 18(1), 35−62.

Kramer, C. (2001). Success in on-line learning. Scarborough, Canada: Thomson DelmarLearning.

Lenk, U. (1998). Discourse markers and global coherence in conversation. Journal ofPragmatics, 30(2), 245−257.

Lindblom, K. (2001). Cooperating with Grice: A cross-disciplinary metaperspective onuses of Grice's cooperative principle. Journal of Pragmatics, 33(10), 1601−1623.

Mann, W. C. (1984). Discourse structures for text generation. Proceedings of the 22ndConference on Association for Computational Linguistics (pp. 367−375). New York:ACM Press.

Mann, W. C., & Taboada, M. T. (2005, October). An introduction to rhetorical structuretheory (RST). Retrieved December 19, 2005, from http://www.sil.org/~mannb/rst/rintro99.htm

Mann, W. C., & Thompson, S. A. (1987). Rhetorical structure theory: A theory of textorganization (Technical Report No. ISI/RS-87-190). Marina del Rey, CA: University ofSouthern California, Information Sciences Institute (ISI).

Mann, W. C., & Thompson, S. A. (1988). Rhetorical structure theory: Towards afunctional theory of text organization. Text, 8(3), 243−281.

Mann,W. C., Matthiessen, C. M. I. M., & Thompson, S. A. (1992). Rhetorical structure theoryand text analysis. InW. C.Mann, & S. A. Thompson (Eds.),Discourse description: Diverselinguistic analyses of a fund-raising text (pp. 39−78). Amsterdam: John Benjamins.

Marcu, D., Amorrortu, E., & Romera, M. (1999). Experiments in constructing a corpus ofdiscourse trees. In M. Walker (Ed.), Proceedings of the ACL Workshop on Standards

and Tools for Discourse Tagging (pp. 48−57). New Brunswick, NJ: Association forComputational Linguistics.

Moore, J. D., & Wiemer-Hastings, P. (2003). Discourse in computational linguistics andartificial intelligence. In A. C. Graesser, M. A. Gernsbacher, & S. R. Goldman (Eds.),Handbook of discourse processes Mahwah, NJ: Lawrence Erlbaum.

Moran, C. (1991). We write, but do we read? Computers and Composition, 8(3), 51−61.Neale, S. (2004). Paul Grice and the philosophy of language. Linguistics and Philosophy,

15(5), 509−559.O'Brien, T. (1995). Rhetorical structure analysis and the case of the inaccurate,

incoherent source-hopper. Applied Linguistics, 16, 442−482.O'Donnell, M. (2000). Intermixing multiple discourse strategies for automatic text

composition. Revista Canaria de Estudios Ingleses (RCEI), 40, 115−134.Pena-Shaff, J. B., & Nicholls, C. (2004). Analyzing student interactions and meaning

construction in computer bulletin board discussions. Computers & Education, 42(3),243−265.

Phillips, B. (2000). Should we take turns? A test of CMC turn-taking formats. In: CHI '00extended abstracts on human factors in computer systems (pp. 341−342). New York:ACM Press.

Phillips, G., & Santoro, G. (1989). Teaching group discussion via computer-mediatedcommunication. Communication Education, 38(2), 151−161.

Pincas, A. (1999). Reference in online discourse. TESL-EJ, 4(1). Retrieved September 24,2004 from http://www-writing.berkeley.edu/TESL-EJ/ej13/a1.html

Preece, J. (2000). Online communities: Designing usability, supporting sociability. NewYork: John Wiley and Sons.

Reed, D. (2001, January 3–6 2001). ‘Making conversation’: Sequential integrity and thelocal management of interaction on Internet newsgroups. Paper presented at the34th Annual Hawaii International Conference on System Sciences (HICSS-34), Maui,Hawaii.

Rovai, A. P. (2003). In search of higher persistence rates in distance education onlineprograms. The Internet and Higher Education, 6(1), 1−16.

Rovai, A. P. (2007). Facilitating online discussions effectively. The Internet and HigherEducation, 10(1), 77−88.

Sacks, H., Schegloff, E. E., & Jefferson, G. (1974). A simplest systematics for theorganization of turn-taking for conversation. Language, 50(4), 696−735.

Schallert, D. L., Lissi, M. R., Reed, J. H., Dodson, M.M., Benton, R. E., & Hopkins, L. F. (1996).How coherence is socially constructed in oral and written classroom discussions ofreading assignments. In D. J. Leu, C. K. Kinzer, & K. A. Hinchman (Eds.), Literacies forthe 21st Century: Research and practice, 45th yearbook of the national readingconference (pp. 471−483). Chicago: The National Reading Conference.

Severinson Eklundh, K., & Rodriguez, H. (2004, January 5–8). Coherence andinteractivity in text-based group discussions around Web documents. Paperpresented at the 37th Hawaii International Conference on System Sciences, Islandof Hawaii.

Simner, J., & Pickering, M. J. (2005). Planning causes and consequences in discourse.Journal of Memory and Language, 52(2), 226−239.

Stent, A. J. (2000, June). Rhetorical structure in dialog. Paper presented at the 2ndInternational Natural Language Generation Conference (INLG'2000), Patras, Greece.

Stent, A. J., & Allen, J. F. (2000). Annotating argumentation acts in spoken dialog (TechnicalReport No. 740). Rochester, NY: University of Rochester.

Taboada, M. T. (2004). Building coherence and cohesion: Task-oriented dialogue in Englishand Spanish. Amsterdam: John Benjamins.

Taboada, M. T., & Mann, W. C. (2006). Applications of rhetorical structure theory. Dis-course Studies, 8(4), 567−588.

Taboada, M. T., & Mann, W. C. (2006). Rhetorical structure theory: Looking back andmoving ahead. Discourse Studies, 8(3), 423−459.

van der Meij, H., de Vries, B., Boersma, K., Pieters, J., & Wegerif, R. (2005). Anexamination of interactional coherence in email use in elementary school. Com-puters in Human Behavior, 21(3), 417−439.

Van der Pol, J., Admiraal, W., & Simons, P. R. J. (2006). Context enhancement for co-intentionality and co-reference in asynchronous CMC. Artificial Intelligence andSociety, 20(3), 301−313.

Van Dijk, T. A. (1977). Text and context: Explorations in the semantics and pragmatics ofdiscourse. London: Longman.

Walker, M. A. (1996). The effect of resource limits and task complexity on collaborativeplanning in dialogue. Artificial Intelligence, 85(1–2), 181−243.

Weller, M. (2002). Learning on the Net: The why, what, & how of online education.London: Kogan Page.

Westfall, P. J. L. (2003). Distance education in the U.S. Air Force. In M. G. Moore, & W.Anderson (Eds.), Handbook of distance education (pp. 631−640). Mahwah, NJ:Lawrence Erlbaum Associates.

Whittaker, S. (2003). Theory and methods in mediated communication. In A. C.Graesser, M. A. Gernsbacher, & S. R. Goldman (Eds.), Handbook of discourse processes(pp. 243−286). Mahwah, NJ: Lawrence Erlbaum Associates.

Whittaker, S., Bellotti, V., & Gwizdka, J. (2006). Email in personal informationmanagement. Communications of the ACM, 49(1), 68−73.

Wolf, F., & Gibson, E. (2005). Representing discourse coherence: A corpus-basedanalysis. Computational Linguistics, 31(2), 249−287.

Woo, Y., & Reeves, T. C. (2007). Meaningful interaction in web-based learning: A socialconstructivist interpretation. The Internet and Higher Education, 10(1), 15−25.