1 Introduction
Human–computer dialogue design encompasses two broad issues. The first concerns the manner in which the user will interact with the machine, whether, for example, they will be presented with a menu of options, a command line interface, or some form of question–answer interaction. The second area concerns the attempt to emulate dialogue of the sort engaged in by human conversants, in which topics of mutual concern are discussed, and in which each conversant has substantive points to make. The interest of this area is in the content and form of the dialogue, rather than the manner of its physical implementation. It is the second area with which we are chiefly interested in this paper.
There are at least two approaches to dialogue design, including the computational approach and the philosophical approach. Conjoining these two, compared with linguistic approaches, in the ways we propose is original and timely. The computational approach focuses on specification of protocols, for example, in the context of multi-agent systems (MAS) communication. FIPA specifications (2002) offer a number of examples expressed using AUML (agent unified modeling language) interaction diagrams (i.e. www.auml.org). For the task at hand, the computational approach falls short in several respects. The focus in the computational approach has, for obvious reasons, been placed squarely upon the design of a clear and unambiguous means of multi-agent communication, rather than upon modelling real world communication. This means that MAS models are poorly suited to being extended to handle the intelligent mediation of human–human or even human–computer interaction. Further, there are also computational limitations to implemented systems. One of the most crippling of these limitations is that the finite state machine approach, which is common throughout models of agent communication, is inflexible and limited. The philosophical approach utilizes ‘Dialogue Game Theory’ models developed within the field of Informal Logic (IL) to prescribe how dialogue should be regulated. It is the second approach in which we are chiefly interested. In particular, we discuss the role of IL as a vehicle for the conduct of significant aspects of dialogue as so defined.
The paper is therefore organized as follows. We start with a brief overview of the field of IL and Dialogue Game Theory in particular. We then present a case for the use of IL within human–computer interaction (HCI), and an overview of current systems utilizing IL theories. Despite the strength of the case for IL, however, we argue that a number of important barriers need to be overcome if the potential of IL is to be fulfilled. The remainder of the paper outlines how current research, by the authors and others, is attempting to overcome these barriers.
2 Basics of informal logic
The field of IL can be seen as an attempt to develop tools that can analyse and evaluate the reasoning and arguments that occur in contexts such as education (e.g. Andrews, Reference Andrews1995; Jackson, Reference Jackson1998; Ravenscroft, Reference Ravenscroft2007), political debate, and legal proceedings (e.g. Walton, Reference Walton1989b; Fisher, Reference Fisher2000; Johnson & Blair, Reference Johnson and Blair2000). Johnson and Blair (Reference Johnson and Blair2000) identify 14 recent areas of research in IL and Groarke (Reference Groarke2002) suggests that three approaches to IL can be characterized, concerned with fallacy theory, rhetoric, and dialogue. The last of these is often referred to as ‘dialectics’ (e.g. Walton, Reference Walton1998) and has been of much recent interest to people working in human–computer dialogue and in artificial intelligence (AI; cf. Vreeswijk et al., Reference Vreeswijk, Brewka and Prakken2003; Reed & Grasso, Reference Reed and Grasso2007; Rahwan & McBurney, Reference Rahwan and McBurney2007). Gordon (Reference Gordon1996), for example, speaks of ‘computational dialectics’ as a new subfield of AI, and Bench-Capon and Dunne (Reference Bench-Capon and Dunne2007) claim that ‘argumentation has come to be increasingly central as a core study within Artificial Intelligence’. Further, Computer-supported collaborative argumentation (CSCA) has become an established subfield of computer-supported collaborative learning (CSCL).
A common approach within dialectics is to construct dialogue games (e.g. Hamblin, Reference Hamblin1971; Walton, Reference Walton1984; Mackenzie, Reference Mackenzie1990; Loui, Reference Loui1998). A dialogue game can be seen as a prescriptive set of rules, regulating the participants as they make moves in the dialogues. These rules legislate as to permissible sequences of moves, and also as to the effect of moves on participants’ ‘commitment stores’, conceived as records of statements made or accepted. Some dialogue games also include challenges and other locutions among commitments, for example, Mackenzie (Reference Mackenzie1979). Such dialogue games have received much recent interest from the AI community and much of the current paper will be concerned with them. A thorough review of dialogue games can be found in Prakken (Reference Prakken2006) who explicitly focuses on games for persuasion (our focus here is somewhat broader) and provides a very useful example dialogue to which he applies various dialogue game models.
One such game, for example, is ‘DC’ (Mackenzie, Reference Mackenzie1979), a system that has received much attention as a vehicle for interactive computer systems (e.g. Moore & Hobbs, Reference Moore and Hobbs1996; Amgoud et al., Reference Amgoud, Parsons and Maudet2000; Maudet & Moore, Reference Maudet and Moore2001; Yuan et al., Reference Yuan, Moore and Grierson2007a, Reference Yuan, Moore and Grierson2008; Reed & Wells, Reference Reed and Wells2007). The system allows five move types: statement (‘P’, ‘Q’, etc., and their truth functional compounds), withdrawals (e.g. ‘no commitment to P’), questions (‘is it the case that P?’), challenges (‘why is it held that P?’), and resolution demands (‘resolve whether P’). There are four rules regulating commitment stores, which are as follows. Stores are null at dialogue commencement. Statements by either participant are added to the store of each. A statement P in response to a challenge of Q results in both P and P→Q being added to each store. A challenge of P results in P being added to the store of the hearer, and P being removed from, and why P is being added to, the store of the maker of the move.
There are six dialogue rules, specifying the protocol under which dialogue takes place, as follows. Participants may utter individual permitted locutions in turn. Mutual commitments may not be uttered. The question ‘P?’ must be answered by ‘P’, ‘not P’, or ‘no commitment P’. The challenge ‘why P?’ must be responded to by a withdrawal of ‘P’, a statement not under challenge by its speaker, or a resolution demand of any commitments of the hearer which immediately imply ‘P’. Resolution demands may be made only if the hearer is committed to an immediately inconsistent conjunction of statements, or withdraws or challenges an immediate consequent of his commitments. A resolution demand must be followed by withdrawal of one of the offending conjuncts, or affirmation of the disputed consequent.
We argue that dialogue games developed within the field of informal logic, such as the one just outlined, tend to have a number of attractive properties for human–computer dialogue.
3 A case for the use of informal logic in human–computer dialogue
The essential argument for the adoption of a dialogue game framework is that the games purport to be models of ‘what is fair and reasonable in argument and criticism’ (Walton, Reference Walton1985; cf. Prakken & Sartor, Reference Prakken and Sartor1996; Loui, Reference Loui1998; Bouwer, Reference Bouwer1999). Given this, then constraining both computer and user to such a game will, if the game is valid, yield—or at least encourage—‘fair and reasonable’ dialogue, and this presumably is what is sought especially in human–computer dialogue and educational dialogue. And since it has been shown (Amgoud et al., Reference Amgoud, Parsons and Maudet2000) that the DC model sketched above, for example, can cater for Walton's six basic types of dialogue, namely persuasion, negotiation, inquiry, deliberation, information-seeking, and eristic dialogues (Walton & Krabbe, Reference Walton and Krabbe1995; Walton, Reference Walton1998, Reference Walton2000), this would represent a considerable broadening of human–computer interaction, especially when considered with the work of Ravenscroft et al. (Reference Ravenscroft, Sagar, Baur and Oriogun2008), which has extended the usability of dialogue games to realize semi-natural and informally logical dialogue interaction.
Further, such a model has a number of desirable properties from the computational perspective. The model's rules can be used to establish the legality of user input, and to assist in formulating a response by restricting attention to the set of legal moves, thus guiding the dialogue beyond a single question/answer interaction. The rules of the model are typically quite tight (cf. Reed, Reference Reed1998): they can prescribe as few as three types of responses to an incoming move type, thus easing the task of checking for legality, and reducing both the search problem for the computer's rejoinder and the strategic issue of selection between competing alternatives. Thus, as argued by Hulstijn (Reference Hulstijn2000), dialogue game rules can be regarded as ‘recipes for joint action’ between the dialogue partners.
The games’ concern with modelling commitment rather than belief (Walton, Reference Walton2000; Lewin, Reference Lewin2000) is also attractive from a computational perspective. Commitments are incurred during the dialogue, in line with clearly stated commitment rules, and the resulting ‘commitment store’ is seen not as a psychological model of memory, but rather as akin to a publicly inspectable set of statements recorded on a slate (Hamblin, Reference Hamblin1970; Mackenzie, Reference Mackenzie1979). This separation of commitment from belief has the advantage that it obviates the need for sophisticated models of belief, knowledge, and sincerity in the context of dialogue. Further, the explicit nature of the commitment function aids computational utilization, and since at any stage the number of entries in a store is at most twice the number of moves made to that stage in the dialogue, combinatorial explosions of commitments cannot occur (Mackenzie, Reference Mackenzie1990). Neither do the models require consistency of commitments—inconsistency is allowed, after which the other dialogue participants may demand retraction of one of the sources of the inconsistency (Prakken, Reference Prakken2000).
The characterization of dialogue moves is also advantageous from the computational point of view, for several reasons. For one, the content of a move is often restricted to one ‘locution’, that is, a statement together with a statement operator (e.g. assert, question, and withdraw). This has the computational advantage of avoiding complexities such as deciding on a practical length for turns (Clark & Schaefer, Reference Clark and Schaefer1989), or formulating a turn-length control policy (Frohlich & Luff, Reference Frohlich and Luff1990), and difficulties involving moves spanning more than one speaker turn (Reichman, Reference Reichman1985). Further, the restricted range of move types allowed by the games makes it possible to use a menu scheme for user input and enables relatively straightforward checks on the legitimacy of input.
Perhaps the main computational attraction, however, is the clarity and expressiveness provided by the moves. Their pre- and post-conditions relate purely to explicit moves and to participants’ (inspectable) commitment stores, so that complications concomitant upon implicit moves never arise. Neither is there any requirement to attempt the difficult task of divining the intentionality of the maker of the move (cf. Reichman, Reference Reichman1985; Mackenzie, Reference Mackenzie1990), given the moves’ clearly defined function within the game. Intentionality is, in effect, embedded within the move (Ravenscroft & Pilkington, Reference Ravenscroft and Pilkington2000). The difficulty of interpreting pragmatic content is, therefore, largely overcome (Pilkington, Reference Pilkington1992).
Further, support for their computational utilization is also forthcoming from researchers in the field of IL itself. Thus, for example, Walton (Reference Walton1998):
One of the primary uses of the new dialectic is to show people how to avoid common errors and fallacies in arguments. But the influence that will lead more directly to the acceptance of the new dialectic as a normative model of argumentation is its adoption in the field of computer science, especially in Artificial Intelligence, in the field of human-computer interactions (developing user-friendly computers), and in the use of computers to assist group deliberations and argumentative dialogues.
4 Current uses of informal logic in human–computer dialogue
There is, then, a strong prima facie case for the use of IL within human–computer dialogue. Although Carbogim et al. (Reference Carbogim, Robertson and Lee2000) argue that ‘only a few argument systems have been deployed in real, complex domains’, much evidence of increasing interest in the use of IL can be found in the literature.
One major use concerns the mediation of argument. Mediation in a legal context has been particularly important (cf. Bench-Capon, Reference Bench-Capon2000; Carbogim et al., Reference Carbogim, Robertson and Lee2000). Gordon (Reference Gordon1994), for example, has developed a model of civil pleading, the ‘pleadings game’, where the plaintiff and defendant confront each other. Bench-Capon (Reference Bench-Capon1998) argues that ‘the rule-governed environment of a dialogue game can provide the necessary structured context for a quasi-courtroom argument’, and develops ‘Toulmin Dialogue Game (TDG)’, a dialogue game based on Toulmin's argument schema, for this purpose. This game can be used to mediate discussions between human participants and seeks to ensure that the argument resulting from the dialogue has an appropriate (Toulmin-based) structure (Bench-Capon, Reference Bench-Capon1998). Similarly, the dialectical argumentation system ‘DART’ has been used to model legal reasoning and argumentation (Freeman & Farley, Reference Freeman and Farley1996). The model captures arguments both as supporting explanations—connecting claims with appropriate supporting data—and as dialectical process—‘alternating series of moves made by opposing sides for and against a given claim’ (cf. Prakken and Sartor, Reference Prakken and Sartor1996, p. 166). The DART model incorporates a ‘burden of proof’ concept, enabling it to support decisions ranging from ‘sceptical’ to ‘credulous’. Similarly, Prakken and Sartor (Reference Prakken and Sartor1996) develop a dialogue game to assess conflicting arguments in legal reasoning. Leenes et al. (Reference Leenes, Lodder and Hage1994) argue for using a dialogue game model of legal justification and develop the dialogue game DiaLaw (Lodder & Herczog, Reference Lodder and Herczog1995; Lodder, Reference Lodder1998). They also argue that dialogue games requiring students to construct legal arguments provide a useful means of teaching legal skills. Again, Prakken (Reference Prakken2008) presents a formal dialogue game for adjudication dialogues. The model reconciles logical aspects of burden of proof induced by the defeasible nature of arguments with dialogical aspects of burden of proof as something that can be allocated by explicit decisions on legal grounds.
Grasso et al. (Reference Grasso, Cawsey and Jones2000) outline a system designed to change the attitudes of its users in the domain of health promotion. The system is based on informal argumentation theories (Walton, Reference Walton1989b), on the grounds that the theory is able to capture ‘everyday arguments and the way they are used to change opinions and values’ (p. 1080). The system's dialogue manager can be armed with a differing range of dialogue games according to prevailing circumstances. Atkinson et al. (Reference Atkinson, Bench-Capon and McBurney2005) propose a dialogue game protocol for practical reasoning, and have constructed a system (PARMA) operationalizing the protocol to allow two human participants to argue with each other (cf. Atkinson, Reference Atkinson2005).
Uses for IL theories have also been proposed as enhancements to intelligent help systems (Pilkington, Reference Pilkington1992; Moore & Hobbs, Reference Moore and Hobbs1996) and expert system explanation generators (Bench-Capon et al., Reference Bench-Capon, Dunne and Leng1991, Reference Bench-Capon, Dunne and Leng1992; Moore & Hobbs, Reference Moore and Hobbs1996). Walton (Reference Walton1998) describes ‘Negotiator Pro’, an expert system, built on dialectical principles, that offers advice on the conduct of bargaining negotiations and elsewhere (Walton, Reference Walton2000) argues that ‘the growing field of expert systems provides a natural application for dialogue theory…sequences of questions and replies—dialogue in short—is a vitally important aspect of the implementation of any expert system’ (p. 330).
Perhaps the main use of IL to date, however, has been within technology-enhanced learning (TEL) systems. As long ago as the mid-1980s, Girle (Reference Girle1986) was expressing a concern that computer-assisted learning systems may be based on assumptions about interactive communication that could put the interaction into a very narrow and inflexible mode. This concern is reflected in Laurillard's concern that ‘too often the multimedia products on offer to education use the narrative mode, or unguided discovery, neither of which supports the learner well nor exploits the capability of the medium’ (Laurillard, Reference Laurillard1995: 184; Retalis et al., Reference Retalis, Pain and Haggith1996; Cumming & McDougall, Reference Cumming and McDougall2000; Yuan et al., Reference Yuan, Moore and Grierson2008). Given the history of its use to improve critical thinking skills (Walton, Reference Walton2000), IL is seen as a means of widening the communications channel and thus surmounting this concern with undue didacticism.
One approach is to have the TEL system engage its student in educational debate on controversial issues such as capital punishment and abortion. Vreeswijk (Reference Vreeswijk1995) has designed ‘IACAS’, an interactive argumentation system enabling disputes between a user and the computer. Yuan et al. (Reference Yuan, Svansson, Moore and Grierson2007b) have applied the argument game (Wooldridge, Reference Wooldridge2002: 153–154) and the abstract argumentation system (Dung, Reference Dung1995) and constructed a computer game ‘Argumento’. Argumento enables human–agent, agent–agent, and human–agent to exchange abstract arguments, and is being extended to exchange concrete arguments as well (Yuan & Schulze, Reference Yuan and Schulze2008). The game is expected to be used as an assistive tool for students who are studying argument games and abstract argumentation. User evaluation suggests that the game is both challenging and entertaining (Yuan et al., Reference Yuan, Svansson, Moore and Grierson2007b). Another raft of projects in Digital Dialogue Games (DDG; see www.interloc.org) conducted by Ravenscroft et al., that is mentioned later, has developed a number of tools (e.g. CoLLeGE, Academic Talk, and InterLoc) that have been adopted and evaluated in a broad range of contexts (e.g. see Ravenscroft, Reference Ravenscroft2007 for a review). These have laid particular emphasis on how dialogue game features are rendered within interfaces to realize human and machine locutions within collaborative learning interactions.
Mackenzie's (Reference Mackenzie1979) DC system has been used for competitive debate, and its applicability has been tested in educational discourse contexts (Moore Reference Moore1993, Reference Moore2000; Moore & Hobbs, Reference Moore and Hobbs1996; Moore, Reference Moore2000; Maudet & Moore, Reference Maudet and Moore2001). The DC system has been further evaluated through a two-agent system, and this motivated the development of a further system DE (Yuan et al., Reference Yuan, Moore and Grierson2003) which has been used as the underlying model for a human–computer debating system (Yuan, Reference Yuan2004; Yuan et al., Reference Yuan, Moore and Grierson2007a, Reference Yuan, Moore and Grierson2008). In the debating system, the computer offers the student a choice of positions on a controversial issue, takes up the position diametrically opposed to that of the student, and engages the student in debate on that issue. This, it is held, may foster the student's debating skills and level of critical awareness, and make him more aware of the substantive issues involved (Moore & Hobbs, Reference Moore and Hobbs1996; Retalis et al., Reference Retalis, Pain and Haggith1996; Quignard and Baker, Reference Quignard and Baker1997; Bouwer, Reference Bouwer1998, Reference Bouwer1999). The educational benefit of such a system seems clear, in that the importance of discussion and debate in education is frequently stressed at the primary (e.g. National Curriculum Council, 1990a), secondary (e.g. National Curriculum Council, 1990b), and tertiary (e.g. Garrison, Reference Garrison1991) levels. The argument is given extra force by Laurillard's mention of an ‘emphasis on whole-class teaching … at the expense of opportunities for discussion and interaction’, and her claim that ‘reflection is too often neglected in the teaching–learning process’ (Laurillard, Reference Laurillard1995: 182). Empirical evaluation of the debating system in use suggests that it can be used to help the students practise argumentation (Yuan et al., Reference Yuan, Moore and Grierson2008).
Other work goes beyond debate, to incorporate other interaction styles. Lewin (Reference Lewin2000) has implemented a dialogue manager, based on ‘conversational game theory’, for task-oriented instructional dialogue. Pilkington et al. (Reference Pilkington, Hartley, Hintze and Moore1992) have used dialogue games to implement a computer-mediated argumentation system called DIALAB, and in a similar vein, Burton et al. (Reference Burton, Brna and Pilkington2000) use dialogue games to implement CLARISSA, a computer-modelling laboratory for investigating collaboration. Pilkington (Reference Pilkington1998) (cf. Pilkington & Parker-Jones, Reference Pilkington and Parker-Jones1996) demonstrates enhancements to a medical simulation-based learning system brought about partly by the adoption of dialogue games. Two types of dialogue games are identified, an inquiry dialogue with asymmetrical participant roles and a more collaborative game-generating cognitive conflict and reflection (Pilkington & Parker-Jones, Reference Pilkington and Parker-Jones1996; Pilkington & Mallen, Reference Pilkington and Mallen1996; Pilkington, Reference Pilkington1998).
Ravenscroft et al. have shown how dialogue games can support conceptual change in science (e.g. Ravenscroft & Pilkington, Reference Ravenscroft and Pilkington2000) and also foster inclusive, engaging, and reasoned learning dialogues of a more generic nature (e.g. McAlister et al., Reference McAlister, Ravenscroft and Scanlon2004; Ravenscroft & McAlister, Reference Ravenscroft and McAlister2006; Ravenscroft et al., Reference Ravenscroft, Wegerif and Hartley2007). For example, Ravenscroft and Pilkington (Reference Ravenscroft and Pilkington2000) used a dialogue game framework to facilitate a ‘structured and constrained dialectic’, which in turn aided the student in enhancing explanatory domain models in ways that led to conceptual development concerning the physics of motion. Their dialogue game framework was able to simulate the tutorial tactics of an expert tutor within a ‘collaborative argumentation’ approach to ‘learning as knowledge refinement’. Thus, the game involved asymmetrical participant roles, the computer being a ‘facilitating tutor’, and the student the ‘explainer’. The framework has been implemented in a prototype system ‘CoLLeGE’ (Computer-based Lab for Language Games in Education). Empirical studies have shown the effectiveness of the dialogue game framework (Ravenscroft, Reference Ravenscroft2000; Ravenscroft & Matheson, Reference Ravenscroft and Matheson2002). More recently, Ravenscroft and his colleagues have developed the Academic Talk and subsequent Interloc tools that mediate dialogue games that realize engaging and ‘reasoned’ discourse and critical thinking through ‘live’ peer interaction (McAlister et al., Reference McAlister, Ravenscroft and Scanlon2004; Ravenscroft et al., Reference Ravenscroft, McAlister and Baur2006; Ravenscroft, Reference Ravenscroft2007). These approaches have been particularly successful in realizing attractive, usable, and engaging dialogue games that link to related digital practices, for example, with widespread and related learning and social software technologies (Ravenscroft, Reference Ravenscroft2008; Ravenscroft et al., Reference Ravenscroft, Sagar, Baur and Oriogun2008). A recent project (Ravenscroft et al., Reference Ravenscroft, McAlister and Sagar2009) has involved evaluations of this DDG approach with over 350 students and 10 tutors. This showed that the DDG approach and InterLoc tool was easy and intuitive to use; popular with, and valued by, tutors and students, where in all cases the tutors/institutions were keen to continue to use it; perhaps most notably, it succeeded in providing a unique way to stimulate critical and collaborative thinking among students, which was evidenced through student and tutor appraisals and analyses of the dialogues; and the generated content or Collaborative Thinking Texts (CTTs) were used in varied, and sometimes unanticipated, ways by students and tutors. In brief, as Ravenscroft et al. (Reference Ravenscroft, McAlister and Sagar2009) point out, ‘the DDGs stimulated and supported learners to think critically on the web and linked this to learning and pedagogical practices’.
5 Major barriers
Thus far, we have shown that there is increasing evidence of the use of dialogue games within computer dialogue systems. However, there are, we argue, five potential and yet significant barriers to the use of dialogue games in this role. For the remainder of this paper, we examine these barriers in turn and, by discussing current research work, explore routes to overcoming them.
5.1 Barrier 1: lack of flexibility
Since in the main the games purport to be normative, prescriptive models of dialogue, they tend to represent idealized rather than everyday interactions (Walton, Reference Walton1998) and their prescriptions may impose an unduly high cognitive load on the would-be user. There is some evidence, however, that mastery of the systems’ stipulations is not unduly problematic (Moore & Hobbs, Reference Moore and Hobbs1996), particularly in a computational environment (Ravenscroft, 2000; Yuan et al., 2008). A further concern is that such dialogues as can be generated through a dialogue game may be rather stilted (e.g. Bench-Capon, Reference Bench-Capon1998). For example, most dialogue game systems tend to restrict questions to the yes/no variety and to questions asking for grounds of commitments (‘challenges’), and to insist that they be answered directly. This may not always represent a reasonable approach to dialogue (Walton, Reference Walton1998), and work in the logic of questions could be profitably considered. It seems likely, however, that many such weaknesses can be overcome in a computational environment. For example, arranging for key points of the computer's dialogue contribution to be represented as hypermedia nodes (Moore, Reference Moore2000) enables a user to clarify points from the computer's dialogue contributions and thus to bypass the restriction on yes/no questions. More generally, we argue (Maudet & Moore, Reference Maudet and Moore2001) that enriching the dialogue with varying forms of media may solve the problem of the potentially stilted nature of the dialogue. This point has been vindicated by recent research by Ravenscroft et al. through their development of the InterLoc tool (e.g. see Ravenscroft et al., Reference Ravenscroft, Sagar, Baur and Oriogun2008), which supports multimedia and multimodal dialogue games that have proved to be attractive and engaging to users in addition to being configurable in ways that evolve with related requirements and ambitions in the context of students developing digital literacies.
In any event, such problems represent weaknesses with the dialogue models per se combined with the complexities of language use, rather than with their computational utilization. As such, they can be overcome by suitable amendments to the models. For example, Yuan et al. (Reference Yuan, Moore and Grierson2003) use a two-agent dialogue simulation system to facilitate evaluation of Mackenzie's (Reference Mackenzie1979) DC system in preventing fallacious arguments, in the light of which they develop a further system DE as the underlying model for their debating system. Similarly, Loui details four game systems ‘exhibiting a range of considerations that have been raised in the design of defeasible reasoning systems’ (Loui, Reference Loui1998: 22), the InterLoc dialogue games adapt and extend some of DC's moves to facilitate a ‘richer and more flexible dialogue game’ (Ravenscroft et al., Reference Ravenscroft, McAlister and Baur2006), and Prakken (Reference Prakken2001, Reference Prakken2005) adopts some of the DC and permissive persuasion dialogue (PPD) (Walton & Krabbe, Reference Walton and Krabbe1995) moves and defines various systems enabling more than one move in a turn, postponing of replies, and backtracking. Ravenscroft and Pilkington (Reference Ravenscroft and Pilkington2000) also talk of ‘ … providing future designers with a toolkit for investigating dialogue games to suit their own applications’ (p. 294) and have operationalized this in recent DDG projects (see www.interloc.org) that now provide a dialogue game authoring tool along with means for their practical application to realize learning interactions among small groups (InterLoc5) within ecosystems for cyber-argumentation (Ravenscroft & McAlister, Reference Ravenscroft and McAlister2008). The two-agent dialogue simulation system developed in (Yuan et al., Reference Yuan, Moore and Grierson2003) can be enhanced to provide IL researchers with a tool to evaluate their dialogue models and strategies. Indeed, computational use of the models can be expected to help clarify the models themselves (Krabbe, Reference Krabbe2000), so that research in dialectics and in human–computer dialogue works to each other's mutual advantage (Hitchcock, Reference Hitchcock2000; Maudet & Moore, Reference Maudet and Moore2001).
A further approach, complementary to that of extending the dialogue games, is to provide for the synthesis of multiple dialogue game types into a single dialogue. Walton suggests that ‘the problem of how to formally represent … functional embeddings of dialogues has not yet been solved. It is by no means a purely philosophical problem, and also represents a real problem for the development of computer dialogue systems’ (Walton, Reference Walton2000: 338). However, approaches have been investigated that allow diverse dialogue types to be modelled as games within a larger structure reflecting the global coherence of dialogue (Reed, Reference Reed1998; Maudet & Evrard, Reference Maudet and Evrard1998; Maudet & Moore, Reference Maudet and Moore2001; cf. Hulstijn, Reference Hulstijn2000; McBurney & Parsons, Reference McBurney and Parsons2002). On this model, extended dialogues are seen as potentially consisting of sequences, embeddings, and other combinations of dialogue games, which can be distinguished from each other through topic (different games of the same type), aim (different game types, same topic), or both (different game types and different topics).
5.2 Barrier 2: computer strategy
Dialogue games tend to be very sparse if seen as functional specifications (Moore & Hobbs, Reference Moore and Hobbs1996). Although this is advantageous in some respects from the computational point of view, the cost of the models’ simplicity is a reliance on the strategic wisdom of the participants, for example, to maintain relevance. In strategic knowledge lies the potential for a link between the logic of argumentation, as modelled by dialogue games, and the semantics of discourse. Without this link, it might be suggested, dialogue games will add nothing to currently available computer dialogue systems. For it is possible for locutions to be syntactically correct and semantically meaningful, yet inappropriate at the pragmatic level (cf. Poesio & Traum, Reference Poesio and Traum1998; Dessalles, Reference Dessalles1998; Pasquier & Chaib-draa, Reference Pasquier and Chaib-draa2005). Further, one possible weakness of dialogue game models is that they tend to be largely neutral vis-à-vis focus and relevance considerations, and this places a further burden upon participants’ strategic reasoning. Indeed, the ability to capture this wisdom could be seen as the crucial aspect of computational use of dialogue games.
For the practical use of IL in HCI, therefore, there must be a suitable link to semantics. It should be noted that any link to the discourse semantics forged via such strategic decisions is largely an extra-game consideration: all that the IL model does is to legitimize a set of move types given the prevailing circumstances and occasionally give some indication of the semantic possibilities. One might think that the model would indicate what is legally available, a strategic choice would be made from this range, and the choice would then be ‘filled up’ with semantic content. Such a view would be too simple, however, for the strategic decision is likely to rely heavily on the available content, and a strategic decision may be needed between alternative contenders for the content (different supporting evidence, for example). In any computerized system, some interplay between strategic and semantic components is therefore needed.
Considerable research effort has therefore been devoted to the development of suitable computational strategies. For a dialogue involving competitive debate, for example, we argue that there are three levels of decisions to consider (Pilkington et al., Reference Pilkington, Hartley, Hintze and Moore1992; Moore & Hobbs, Reference Moore and Hobbs1996; Maudet & Moore, Reference Maudet and Moore2001; Yuan, 2007a). At level 1, the issue is whether to retain or change the current focus. At level 2, the decision for the computer is whether to try to ‘demolish’ the user's position, or to seek to ‘build’ its own position. In a similar manner, Freeman and Farley (Reference Freeman and Farley1996) picture side-1 building support for its claim, and side-2 seeking to refute side-1's arguments via ‘undercutting’ or ‘rebutting’. At level 3, the decision involves which method to adopt in fulfilment of the objectives set at levels 1 and 2.
For dialogue types other than debate, other strategies may be appropriate. Grasso et al. (Reference Grasso, Cawsey and Jones2000), for example, adopt, for their nutritional advice-giving system, schemas derived from Perelman and Olbrechts-Tyteca's (Reference Perelman and Olbrechts-Tyteca1969) ‘New Rhetoric’, and Ravenscroft and Pilkington (Reference Ravenscroft and Pilkington2000) have ‘a repertoire of legitimate tactics available for addressing common conceptual difficulties’ (p. 283) that can be selected based on a strategic pedagogy developed through modelling effective tutor behaviour (Ravenscroft, Reference Ravenscroft2000). Amgoud and Maudet (Reference Amgoud and Maudet2002) suggest ‘meta-preferences’, such as ‘choose the smallest argument’, to drive the choice of move, and Freeman and Farley (Reference Freeman and Farley1996) delineate ordering heuristics as guidelines for selecting argument moves. Prakken (Reference Prakken2001, Reference Prakken2005) suggests, in line with his dialogue framework, that the dialogue focus and dialectical relevance can be maintained by restricting dialogue participants to replying and backtracking to the previous move. Yuan et al. (Reference Yuan, Svansson, Moore and Grierson2007b) implement a probability–utility-based strategy, which enables a software agent to compute all the possible dialogue sequences and then select a legal move with the highest probability of winning an abstract argumentation game.
5.3 Barrier 3: group dialogues
Although the trend is starting to shift, dialogue game research to date has focused predominantly on two-party games. An important development is to relax this assumption and enable multiple participants. This is an important issue in current interactive systems, given the importance of group work in education (e.g. Cumming & McDougall, Reference Cumming and McDougall2000) and the growing interest in CSCL in general (Steeples et al., Reference Steeples, Unsworth, Bryson, Goodyear, Riding, Fowell, Levy and Duffy1996; Hoadley, Reference Hoadley1999; Stahl, Reference Stahl2006) and CSCA in particular (e.g. Veerman et al., Reference Veerman, Andriessen and Kanselaar2002; Ravenscroft, Reference Ravenscroft2007; Ravenscroft et al., Reference Ravenscroft, Wegerif and Hartley2007). Further, the burgeoning of social software practices including widespread collaborative argumentation provides a fertile landscape for the application of the next generation of dialogue game technologies (Ravenscroft & McAlister, Reference Ravenscroft and McAlister2008). Thus, building on these contemporary practices, Ravenscroft et al. have produced tools for group dialogue games and performed significant evaluations of their acceptability and effectiveness (see Ravenscroft, Reference Ravenscroft2007; Ravenscroft et al., Reference Ravenscroft, McAlister and Sagar2009).
An IL model designed explicitly for supporting group discussion offers suggests, in our view, two major advantages in this context. One major benefit of an IL model is its ability to provide a regulatory framework for interactions within collaborative discussions. Means of suitably controlling the evolving discussion are required (cf. Okamoto & Inaba, Reference Okamoto and Inaba1997). And given that, as suggested earlier, IL models purport to be models of ‘fair and reasonable’ dialogue, the case for their adoption as the regulatory framework seems clear. In a similar manner, Finkelstein and Fuks (Reference Finkelstein and Fuks1990) use a dialogue games model as the basis for a system for providing automated support for groups collaborating on the development of software specifications (cf. Finkelstein, Reference Finkelstein1992; Burton et al., Reference Burton, Brna and Treasure-Jones1997; Bouwer, Reference Bouwer1998). Similarly, the Zeno Argumentation Framework (Gordon & Karacapilidis, Reference Gordon and Karacapilidis1997; Gordon et al., Reference Gordon, Angi, Gernot and Oliver2001) is an Internet-based environment for supporting argumentation, negotiation, and group decision-making, based on IL principles. The system is conceived as an ‘intelligent support system for human mediators’, which ‘transforms Issue-Based Information System (IBIS) from a lifeless method to organize and index information into a playing field for stimulating debate’.
A second major benefit of IL models to computer-supported collaboration is that, by providing a computationally tractable model of dialogue, the model makes it possible in principle for a computational agent to participate in the dialogue. This is advantageous in a number of ways. CSCL work is often of a discursive nature (Simon, Reference Simon1997) and the ability of the computational agent to play the role of ‘devil's advocate’ is potentially of educational value (cf. Retalis et al., Reference Retalis, Pain and Haggith1996). This is especially the case, perhaps, in contexts in which the human participants all agree but it is felt educationally advantageous for them to critically explore their shared view. Further, it may be the case in asynchronous computer conferences that propositions posed by one participant evoke no response (Hewitt & Teplovs, Reference Hewitt and Teplovs1999; Cunningham-Atkins et al., Reference Cunningham-Atkins, Powell, Moore, Hobbs and Sharpe2004) and that discussion is therefore stymied. A computational agent could potentially provoke discussion in such circumstances, and play the role of symmetrical participant or asymmetrical facilitator or expert. A further advantage of computational participation is that it affords participants the possibility of their own ‘private’ discussion with the agent. This might be used for rehearsal and practice prior to entering the group discussion (perhaps to resolve any ‘intra-agent conflict’ (Amgoud & Maudet, Reference Amgoud and Maudet2000)), and/or reflection and analysis after a group discussion. The facility may be particularly useful for people reluctant to enter the group discussion or for people with a social disability, such as autism, which restricts their participation (cf. Moore et al., Reference Moore, McGrath and Thorpe2000). A final advantage of computational participation is that it would enable a number of computers to hold discussions with each other, in a multi-agent system (Amgoud et al., Reference Amgoud, Parsons and Maudet2000) and, given claims concerning the educational benefits of vicarious learning from the dialogue of others (Cox et al., Reference Cox, McKendree, Tobin, Lee and Mayes1999; Stenning et al., Reference Stenning, McKendree, Lee and Cox1999; Lee, Reference Lee2006), the resulting transcripts might make educationally valuable study material.
Challenging research issues arise, however, when seeking to apply dialogue games to group discussions. In the event that the group discussion involves two teams being formed, the issue of how team commitment stores (CS) should be updated is complex. The usual arrangement is ‘de facto commitment’—a participant has to explicitly withdraw from his commitment store those statements of his interlocutor to which he is not prepared to commit (Mackenzie, Reference Mackenzie1979). The position is more complex in a team situation: for different theses can be expressed within a team and this raises the issue of how the other team's CS should be updated. A more radical position on the role of commitment within group discussion has been adopted by Ravenscroft and his colleagues (e.g. see Ravenscroft, Reference Ravenscroft2007, for a review). The dialogue games performed through their InterLoc tool do not have any explicit shared CS that are displayed separately from the dialogue itself. Their emphasis is on mediating and scaffolding legitimate ‘thinking conversations’, where the IL of the dialogue is contained through the rules about legitimate move patterns rather than through implied and explicitly displayed commitment. In other words, the underlying logic subserves the play of dialogue instead of driving it.
A further issue with the team arrangement concerns turn taking. Many dialogue games reduce the turn-taking problem to the following equation: one move=one turn. Grasso et al. (Reference Grasso, Cawsey and Jones2000) express some concern about such an arrangement, and in the context of a set of participants, in particular, the definition may need to be refined: for it may not be realistic to allow only one move for the whole team, or to oblige one move for each player of the team. Some dialogue games support more complex move/turn structures (Walton & Krabbe, Reference Walton and Krabbe1995), although this approach often leaves open the problem of determining what, in a given case, constitutes a turn. A possible solution could be to introduce an explicit turn-taking move, in line with Bunt's ‘dialogue-control-acts’ (Bunt, Reference Bunt1994; cf. Poesio & Traum, Reference Poesio and Traum1998). Ravenscroft et al. (Reference Ravenscroft, Sagar, Baur and Oriogun2008) have introduced a straightforward turn-taking model that is not just acceptable to users but also stimulates them to ‘listen’ and reflect.
In the event of genuine ‘polylogues’, dialogues in which each player wants to achieve a different goal and teams do not therefore form, it is not currently clear what alterations will be needed to the model (cf. Traum, Reference Traum2004). This, in fact, is an example of what we believe to be a fundamentally important research issue, namely the ramifications of computer dialogue research for the field of IL itself. The crucial point, we argue, is that the computer environment can act as a test-bed in which the dialectical theories can be evaluated and refined. Walton (Reference Walton1998: 29) argues: ‘the formal systems of dialogue that have proliferated in recent times appear potentially useful, but … they are too diffuse, too multiple and too abstract’. Further, a computational test-bed is likely to provide a useful facility for rationalizing the proposed models and making them less abstract. One useful approach might be to allow two or more computer systems to run with a proposed system in dialogue with each other, and to study the results. As Amgoud and Maudet (Reference Amgoud and Maudet2000) point out: ‘conversation simulation between computational agents is more and more considered as an important means to get empirical results about dialogue structures and behaviours’ (cf. Hitchcock, Reference Hitchcock2000; Krabbe, Reference Krabbe2000; Dignum & Vreeswijk, Reference Dignum and Vreeswijk2004).
5.4 Barrier 4: natural language
Argument and debate are home to some of the most complex and sophisticated examples of language use: they go far beyond what current natural language processing might hope to tackle. This barrier, though, also represents a key opportunity for IL. Where NLP is an unrealistic goal, IL offers a route to some structured headway—without precluding subsequent NLP contributions. The debating system (Yuan et al., Reference Yuan, Moore and Grierson2007a), for example, operates on the basis of a predetermined (albeit expandable) set of propositions. The Hartley and Hintze mediating system (Hartley & Hintze, Reference Hartley and Hintze1990) operates on strings, Yuan et al.'s (Reference Yuan, Svansson, Moore and Grierson2007b) abstract argument game operates on a graph with a set of visual arguments, Ravenscroft and Pilkington's (Reference Ravenscroft and Pilkington2000) CoLLeGE system adopts a ‘menu and template scheme’, and in the Grasso et al. system, input and output ‘consists at the moment of sentences in a first-order language’ (Grasso et al., Reference Grasso, Cawsey and Jones2000: 1097).
One can speculate that much of the difficulty in unregulated discussions concerns semantic shifts, and that this would be largely ruled out by the propositional logic of the dialogue games. Further, the models provide a valuable service at the propositional logic level, for example, by keeping track of commitments and pointing up inconsistencies and consequences of extant positions. In addition, it may be that attempting to work within the confines of propositional logic will turn out to be revealing about what Walton (Reference Walton1989a) sees as the contested ground between semantics and pragmatics.
5.5 Barrier 5: representation
The systems we have outlined face two distinct representational problems. The first is in representing the argument content that is contributed by users or offered by a system. The second is in representing the structure of the protocol that governs the dialogue game as it unfolds. We may, for example, build a system that is to use the protocol of DE to argue about capital punishment, but how are we to store the arguments about capital punishment, and how are we to store the rules of DE? To date, computational dialectic systems have approached the first problem with a wide range of idiosyncratic solutions, and have avoided the second problem entirely by building the logic of a game into the program structure itself. Idiosyncrasy in representation is problematic because resources cannot be shared or reused between the many systems that exist. Building game logic into a program means that a system is unable to play more than one game, and cannot even easily be rewritten to play a different game.
The Argument Interchange Format (AIF), is a nascent agreement in the argumentation community for an approach to simultaneously solving both problems (Chesnevar et al., Reference Chesnevar, McGinnis, Modgil, Rahwan, Reed, Simari, South, Vreeswijk and Willmott2006). Applications of the AIF to date have focused on monologic argument representation—that is, on the first of our two representational problems. The results from these early investigations have demonstrated the power and flexibility of the AIF, particularly in semantically rich environments (such as Semantic Web services (Rahwan et al., Reference Rahwan, Zablith and Reed2007)). AIF systems such as ArgDF (www.argdf.org) might reasonably be expected to provide input to systems of computational dialectics. The challenge lies in building interchangeable representations of dialogue systems and (most difficult of all) tying those representations into the representations of monologic argument. Thus, for example, DE's rule RCHALL involves the troublesome imc from Mackenzie's original presentation of DC—one of the ways of responding to a challenge, ‘Why P?’, is with a resolution demand of any of the challenger's commitments that immediately imply P. Representing this rule is not currently possible in the AIF because ‘immediate implication’ is something that could be computed only by a procedure over the monologic part (i.e. the arguments and CS of the interlocutors). More troublesome yet, it is difficult to see how the AIF, as it stands, could tie together in a principled way the result of a resolution demand specified by RCHALL and the update to the monologic structure. Implemented tools such as those described above may provide a concrete basis from which to develop the next iteration of the AIF.
6 Summary
This paper has outlined recent work in applying dialectical theories developed within the field of IL to dialogue involving people and computers, that is, ‘computational dialectics’ (Gordon, Reference Gordon1996). We have highlighted five barriers to be addressed if IL theories are to be successfully applied to human–computer dialogue, and reviewed current and potential approaches being taken by the authors and others to overcoming these barriers. Specifically, barrier 1, a lack of flexibility, is being addressed by: enriching the dialogue with varying forms of media; experimenting with amendments to the IL models; providing for the incorporation of multiple dialogue game types into a single dialogue; and the development of usable and configurable interfaces. Substantive strategies are being developed, for both players of individual dialogue games and for facilitating dialogues involving multiple games, to address barrier 2, the issue of strategies being needed to operationalize IL models. Barrier 3 concerns the facilitation of group dialogue. For team-based dialogue, this is being addressed by amending the IL models to allow for ‘de facto commitment’ and explicit turn-taking moves or deemphasizing the requirement for shared CS as the drivers of reasoned conversational practice. Means of catering for polylogues remains an open question. The issue of natural language processing (barrier 4) falls outside the remit of IL as such, being rather a matter of computational linguistics, but a variety of practical ‘ways round’ the problem are being adopted in the meantime. Turning to the final barrier—how to reuse system argument contents and dialogue game models, the AIF appears to be providing a promising approach to simultaneously solving both problems; challenging issues however still remain in the representation of dialogue games.
Much remains to be done, then, but the potential pay-off in terms of expanding human–computer communication is enormous. Further, there is great scope for an interesting and fruitful interplay between research within IL on the dialogue models per se and research on their computational utilization. The hope is that this paper will consolidate the work in computational dialectics and also move this interplay forward.
