1995 draft for AI MAGAZINE Report REPORT ON THE COMPUTATIONAL DIALECTICS WORKSHOP Report on The Workshop on Computational Dialectics Seattle, Washington July 1994 R. P. Loui Washington University Dialectic is the fancy word for debate. AI contributes to the logic and processing of argument and uses ideas of argument in its models of communication; as it continues to do this, the computational study of dialectic, like the computational study of argument, is inevitable. Dialectic has had a tough time being respectable in this century. In its more logically sober guises, it is basically the idea that rational inquiry is best achieved through largely adversarial discourse. Dialectic is a very old idea that simply will not disappear. It is the idea of structured linguistic interactions proceeding according to a protocol. The term "Computational Dialectics" was meant to describe an area of activity in AI which considers the language and protocol of systems that mediate the flow of messages between agents constructing judgement, agreement, or other social choice, to recognize or achieve an outcome in a fair and effective way. Tom Gordon had the idea of a AAAI workshop on "Computational Dialectics," and the intriguing phrase is his. It is to his credit that the idea has earned a further hearing in a German workshop that he is presently organizing with colleagues on the continent. For the workshop in Seattle, our committee included Johanna Moore and Katia Sycara. It is to their credit that they agreed to be on the committee, perceiving the fundamental nature of the workshop's topic and how broad its implications will be in time. Senior Colleagues Hajime Yoshino and Layman Allen were able to attend: both have records of interdisciplinary leadership, where law and formal systems meet, and their input is always valued. Trevor Bench-Capon from England, Daniel Poulin from Quebec, and perhaps half of the other attendees made the trip to Seattle just for the workshop, although other scheduled speakers, notably Jaap Hage from the Netherlands and Arthur Merin from Germany were unable to attend. Gilad Zlotkin, Sandra Carberry, Gerhard Brewka, Jon Doyle, Anne Gardner, L. Thorne McCarty, and Kevin Ashley also contributed their intellectual weight to the proceedings. At issue were (1) the idea that dialectic has something to do with computation, (2) the idea that AI has something new to contribute to the understanding of dialectic, and (3) the idea that dialectical approaches to longstanding AI problems permit new progress. Dialectic is at the core of models of rational inquiry that are honest about procedural rationality. It is no accident that Herbert Simon and Nicholas Rescher, communicating in the early sixties, would shortly thereafter discover the idea of procedural rationality (Simon), and a compact logical description of dialectic (Rescher). Why should there be debate instead of just unpacking the declared logical entailments of non-monotonic or defeasible reasons? The answer has to be that there is search, and the answer makes no sense unless search is limited. Simon and Rescher both knew this answer, but it has taken some time to put the full measure of their ideas together and make explicit pronouncements. There had been two notable forerunners of AI workshops that explicitly contemplated "argument." The last was run successfully by Sergio Alvarado, who was able to interest a much larger number of researchers on natural language processing in the 1991 Spring Symposium. In contrast, the workshop in Seattle had a clear bias toward AI and Law and knowledge representation. The population shift is natural: what has happened in intervening years is that formal models of argument have grown very precise. They are much more precise than at any time in the few thousand years that Western intellectuals have studied argument as a logico-linguistic phenomenon. The AI and Law community has felt the impact of the developments and has decided for some time to take the lead toward further developments. The organizers of the last Non-Monotonic Reasoning Workshop (Brewka) and the next Knowledge Representation and Reasoning Conference (Doyle) were present, and their presence is evidence that AI and Law will be followed, if not joined in its pursuits. Significant papers were presented by Brewka and by McCarty. Brewka gave a formal reconstruction of Rescher's formal theory of disputation within the framework of default reasoning. If Rescher is indeed the high mark in the history of understanding dialectic, which now seems incontrovertible, then Brewka represents the first direct step forward in nearly two decades. Of course, those two decades were active years for the non-monotonic reasoning community which makes possible Brewka's work. Brewka uses the precision of his revised default framework to answer questions that Rescher did not or could not answer. I disagreed with Doyle over the long-term importance of precisely answering exactly these questions; I felt that there remain variations of the model to be explored at a higher level of abstraction instead of obsessing over one variation's details. Still, a full literature search of citations of Rescher's 1977 monograph, DIALECTICS, reveals no useful formal extension or clarification of the logical system, prior to Brewka. The Argentine constructive logician Gino Roetti is the only person who has told me he read the book in 1977, saying he knew then that it was an important little book, but that it was perhaps ahead of its time. Was an understanding of it really impossible until AI had had its non-monotonic reasoning excursions? McCarty's paper was as electrifying as it was an object of impassioned disagreement. Although primarily historical and not essentially concerned with the logic of dialectic, it was at once recognized for its importance. McCarty noted that it was the greatest Anglo-American philosopher of law, H.L.A. Hart, who introduced the logico-linguistic world to the term "defeasible" (Hart imported it from English contract law). This is perhaps of greater interest to the non-monotonic reasoning community, since that community has a stake in a better understanding of defeasible reasons; still, it is clear that dialectic presupposes defeasible reasons upon which to construct arguments that can stand in opposition. McCarty raised the question of why Hart never used his neologism, "defeasibility", in his work again. There was superficial evidence that Hart even disavowed his early paper that introduced the idea of a defeasible concept, and the modern lore of jurisprudence holds that he did disown the paper and its ideas. For Hart, the early idea of a defeasible concept probably evolved into his famous idea of an open-textured term to be defined perhaps using defeasible rules (which can be traced to Wittgenstein via Friedrich Waismann). But McCarty took the opportunity to assail AI's work on defeasible reasoning. He said that defeasible rules were trivial devices that could be eliminated in favor of two kinds of negation. As devices for rule-based programming, they provide no advance over intuitionistic logic programming (says McCarty). They are trivial, that is, when compared to the defeasibility of open-textured concepts, the logic of which remains unanalyzed (says McCarty, downplaying the work of case-based reasoning researchers, especially his colleagues Edwina Rissland, Kevin Ashley, and David Skalak). Ironically, the AI and law luminary, McCarty, like the deontic legal logician Carlos Alchourron, chooses to attack a legal invention (defeasible reasoning) on technical aesthetics. The issue of open texture was made more pointed by a paper presented by Pierre St-Vincent. He and Daniel Poulin gave a non-dialectical, non-defeasible treatment of open-textured concepts. These Canadians wanted to analyze open-texture as a normed or fuzzy concept, in the manner that might be appropriate at an Uncertainty in AI meeting. A role for dialectic remained in St-Vincent's approach: dialectic provided for an exchange of messages between advocates interested in the contrary aims of enlarging the scope of a concept, or else restricting it. This paper would be an excellent place for fuzzy logicians to approach the issues raised by this workshop. These papers, instead of pushing models of dialectic forward, involved debating the current understanding of defeasible reasoning and dialectic. The papers of Violetta Cavalli-Sforza and Dan Suthers, and Kathy Freeman were counterpoise. Cavalli-Sforza reported on Belvedere, "an environment for practicing scientific argumentation." Belvedere is a tutoring system for high school and junior high school science classes. Graphical support is provided for diagramming arguments interatively. Like much of the work in the 1991 symposium, it extends the idea of Toulmin diagrams for arguments into the world of graphical user interfaces. It is ambitious work. This is in part because all of the research that builds on the gIBIS (Conklin) and AQUANET (Marshall et al.) experiences points to the average user's unwillingness to frame arguments formally. The Cavalli-Sforza-Suthers work is ambitious also because it insists that the arguments to be diagrammed be scientific arguments. Surely scientific arguments have their own special logic. Cavalli-Sforza has for a while been interested in Toulmin's own attempts to apply his work on argument to specialized forms of reasoning, such as legal, decision-theoretic, or scientific reasoning. Belvedere is significant as a making good of Toulmin, with possible ramifications in the interactions of users with scientific databases; it is perhaps more significant in this respect than it is successful as a tool for teaching young scientists. Kathy Freeman gave a synopsis of her dissertation work, which also included a graphical system for diagramming arguments based on Toulmin. Her system provides an automated reasoner for additional support of the user. Although her representational scheme seeks to mix dialectical ideas with measures of plausibility, there was nothing objectionable in her work to this audience. To me, this shows that the research program that seeks to import GUI versions of Toulmin's ideas to AI and CSCW (Computer-Supported Collaborative Work), a paradigm present in 1991 and mentioned in the recent ACM Computing Surveys article on CSCW, is a research program that remains on strong footing. Bench-Capon and P.H. Leng's paper was self-explanatory: "Developing heuristics for the argument based explanation of negation in logic programs." The paper extends work presented at a series of Expert Systems conferences: "Interacting with knowledge-based systems through dialogue games," (1991), "Using Toulmin's argument schema to explain logic programs," (1991) "A dialogue game for dialectical interaction with expert systems," (1992) and "Argument based explanation of the British Nationality Act as a logic program" (1993). The premise in all of this work is that negation-as-failure is inadequate as a report to a user. In its stead, a dialogue game is played in which the automated reasoner uses its recently "completed" search to defeat a player who seeks to establish the contentious proposition (against the advice of the program). The user's stubbornness as a player in such a game defines the amount of information that is required in way of "explaining" the result: the more stubborn the use, the longer the dialectical game, the larger the "explanation" generated. Understanding the interaction as dialectic, as a two-player adversarial game of responses, is novel. It is what computational dialectics brings to AI's foundations, with ramifications in the design of systems. I believe that Layman Allen's paper, Gordon and Brewka's paper, and my own could also be regarded as novel. Professor Allen gave the rules of an "A-Hohfeldian game" which some might properly regard as a Wff'n'Proof for logical statements using the turn-of-the-century, heretofore underappreciated, largely deontic ontology of the legal scholar Hohfeld. The work was jointly undertaken with Charles Saxon. The Allen-Saxon revision of Hohfeld's ontology first saw good use in an expert system reported at the 1991 First International Workshop on Deontic Logic. In our Seattle workshop, the work took a turn toward formal games, renewing Layman Allen's interest in producing the rules of adversarial logical games. Allen hinted that rules for a fully disputational discourse game, that is, his ideas for a model of dialectical disputation, were in the works. Gordon and Brewka gave a paper about decision-making. The preference ordering required for evaluating tradeoffs in decision-making was assimilated to the ordering used to represent which arguments are preferred to their opposing arguments. Some people will think their analysis is backwards, but this is the novelty. We have seen past authors importing the axioms of utility to the analysis of multiple non-monotonic extensions (e.g., Wellman and Doyle). The idea here is to use the machinery of defeat among arguments to replace real-valued utility when decision-making requires making tradeoffs. Instead of saying that money and time can be mapped to utilities, say that an argument for a decision based on consequences of one kind (one attribute) defeats an argument for a different decision based on consequences of some other kind (other attribute). It is a qualitative approach to multi-attribute utility. The idea seems initially sound, although as Henry Prakken observes, it is maybe pushed too far. Preferences among arguments are sometimes based on superficial syntactic features, such as specificity among defeasible reasons. These might be used purely for notational convenience. If so, then the availability of only one kind of preference might be representationally impoverished rather than merely clever. My paper, "Argument and arbitration games," sought to include an argument game within the various games of negotiation. Game-theory's approach to negotiation is to characterize it as a series of proposals. Sometimes settlement is based on reason, although as Sycara notes, sometimes the reason is merely that one person is in a better bargaining position than the other. It is unfortunate that Arthur Merin was unable to give his paper along with my paper, since the two seemed to be in line with each other. The work is just begun, but I am already perched on the limb that holds that future economists' models of negotiation will integrate discourse and reason. "Cheap talk" and precedent are simply too important as parts of what can be considered to be negotiation. Zlotkin gave a synopsis of his work and his reaction to the thrust of the workshop. He and Rosenschein had just produced the nifty monograph, RULES OF ENCOUNTER, in which the economic approach to achieving agreement is used to derive interesting results in DAI. Compared to their methodology, the emerging methodology of computational dialectics seems troubling. They prove interesting theorems about games; we give rules for interesting games. Providing rules for a game without justifying or analyzing those rules might be the right thing to do, and might be no different (in the abstract) from what people do when they design AI programs. But this needs to be defended. Describing such discourse games leads to neither theorem nor consensus. It just describes a regimen for the interaction of a society of minds. Clearly this is important, but there must be some way of evaluating the merit of the resulting models. Computational dialecticians can defend their results and develop the kinds of taxonomies that logicians have for their wares. They can also inspire better programs, as seen in Cavalli-Sforza-Suthers and Bench-Capon-Leng. More work needs to be done. We had our share of students completing dissertations: George Ferguson (under James Allen), Jennifer Chu-Carroll (under Carberry), Vincent Aleven (under Ashley), Jeremy Wertheimer (with John Mallery). Although each has in the past written papers and programs that were centrally relevant to the theme of the workshop, all remained surprisingly quiet. This is a major failing for a workshop. Although the topic and its discussion heavily favored law and logic interdisciplinarians, there is no future for a field that does not promote the development of its future scholars. The biggest disappointment was that Gerard Vreeswijk could not attend, since his dissertation work neatly defines the area. His work seeks explicitly to model dialectic as a discourse game in logic. He is now taking the same paradigm of computational dialectics and addressing DAI problems of coordination within it. I see Vreeswijk's remaining on the other side of the Atlantic Ocean as a metaphor. Much of the best work on this topic will likely remain on the continent, where the temperament is more intellectual, the logical traditions more varied, and the relevance of paradigm to application better welcomed. Like AI and Law, the conditions for leadership exist outside of the US. We were lucky in Seattle that the American conference was able to take the lead for a year. Alvarado, S., ed. Proc. AAAI Spring Symposium on Argumentation and Belief, 1991. Conklin, J. ``gIBIS: a tool for all reasons,'' J. American Soc. for Information Science 40, 1989. Hart, H.L.A. ``Ascription of responsibility and rights,'' in A. Flew, ed. Logic and Language, Oxford, 1951. Malone, T. and Crowston, K. ``The interdisciplinary study of coordination,'' ACM Computing Surveys 26, 1994. Marshall, C. et al., ``Aquanet: a tool to hold your knowledge in place,'' Proc. Hypertext 1991. Rescher, N. Dialectics, SUNY Albany Press, 1977. Rosenschein, J. and Zlotkin, G. Rules of Encounter, MIT, 1995. Simon, H. ``The logic of heuristic decision making,'' in Rescher, N., ed. The Logic of Decision and Action, Pittsburgh, 1966. Simon, H. The Sciences of the Artificial, MIT Press, 1969. Toulmin, S. The Uses of Argument, Cambridge, 1958. Vreeswijk, G. Studies in Defeasible Argumentation, doctoral dissertation, Vrije University Amsterdam, 1993.