AI Magazine Volume 16 Number 4 (1995) (© AAAI) Workshop Report

but it has taken some time to put the The Workshop on full measure of their ideas together and make explicit pronouncements. Computational Dialectics There were two notable forerunners of AI workshops that explicitly con- templated argument. The last was run R. P. Loui successfully by Sergio Alvarado, who was able to interest a much larger number of researchers in natural lan- guage processing in the 1991 AAAI Spring Symposium Series. In contrast, the workshop in Seattle had a clear bias toward AI and law and knowl- ■ The 1994 Workshop on Computational attend. Gilad Zlotkin (Hebrew Univer- edge representation. The population Dialectics was held during the 1994 sity), Sandra Carberry (University of National Conference on AI. At issue shift is natural: In the intervening Delaware), Gerhard Brewka (Universi- years, formal models of argument were the following ideas: (1) dialectic ty of Vienna, Austria), Jon Doyle (Mas- have grown precise. They are much has something to do with computation, sachusetts Institute of Technology), (2) AI has something new to contribute more precise than at any time in the Anne Gardner, L. Thorne McCarty to the understanding of dialectic, and few thousand years that Western (Rutgers University), and Kevin Ashley (3) dialectic approaches to long-stand- intellectuals have studied argument as (Pitt) also contributed their intellectu- ing AI problems permit new progress. a logicolinguistic phenomenon. The al weight to the proceedings. This article outlines the significant AI and law community has felt the research presented at the conference. At issue were the following ideas: impact of the developments and has (1) dialectic has something to do decided for some time to take the lead with computation, (2) AI has some- toward further developments. The om Gordon (GMD) had the idea thing new to contribute to the organizers of the last Nonmonotonic for an American Association for understanding of dialectic, and (3) Reasoning Workshop (Brewka) and Artificial Intelligence (AAAI) dialectic approaches to long-standing T the next Knowledge Representation workshop on computational dialec- AI problems permit new progress. and Reasoning Conference (Doyle) tics, and the intriguing phrase is his. Dialectic is at the core of models of were present, and their presence is It is to his credit that the idea has rational inquiry that are honest evidence that AI and law will be fol- earned a further hearing in a German about procedural rationality. It is no workshop that he is currently organiz- accident that Herbert Simon and lowed, if not joined, in its pursuits. ing with colleagues on the continent. Dialectic has had a tough time For the workshop in Seattle, Washing- being respectable in this century. In ton, at AAAI-94, our committee its more logically sober guises, it is included Johanna Moore (Pitt) and Dialectic … is basically basically the idea that rational inquiry Katia Sycara (Carnegie Mellon Univer- is best achieved through largely adver- sity). It is to their credit that they the idea that rational sarial discourse. Dialectic is an old agreed to be on the committee, per- inquiry is best achieved idea that simply will not disappear. It ceiving the fundamental nature of the is the idea of structured linguistic workshop’s topic and the broadness of through largely interactions proceeding according to a its implications in time. adversarial discourse. protocol. The term computational Senior colleagues Hajime Yoshino dialectics was meant to describe an (Meijigakuin University) and Layman area of activity in AI that considers Allen (University of Michigan) were the language and protocol of systems that mediate the flow of messages able to attend: Both have records of Nicholas Rescher, collaborating in interdisciplinary leadership, where between agents constructing judg- the early 1960s, would shortly there- law and formal systems meet, and ment, agreement, or other social after discover the idea of procedural their input is always valued. Trevor choice to recognize or achieve an out- rationality (Simon) and a compact Bench-Capon from England (Imperial come in a fair and effective way. logical description of dialectic College), Daniel Poulin from Quebec Significant papers were presented (Université de Montréal), and perhaps (Rescher). Why should there be by Brewka and McCarty: Brewka gave half of the other attendees made the debate instead of just unpacking the a formal reconstruction of Rescher’s trip to Seattle just for the workshop, declared logical entailments of non- formal theory of disputation within although other scheduled speakers, monotonic or defeasible reasons? the framework of default reasoning. If notably Jaap Hage from The Nether- The answer has to be that there is Rescher is indeed the high mark in lands (University of Limburg) and search, and the answer makes no the history of understanding dialec- Arthur Merin from Germany (Univer- sense unless search is limited. Simon tic, which now seems incontrovert- sity of Stuttgart), were unable to and Rescher both knew this answer, ible, then Brewka represents the first

Copyright © 1995, American Association for Artificial Intelligence. All rights reserved. 0738-4602-1995 / $2.00 WINTER 1995 101 Workshop Report direct step forward in nearly two ble concept probably evolved into his classes. Graphic support is provided decades. Of course, these two decades famous idea of an open-textured term for diagramming arguments interac- were active years for the nonmono- to be defined perhaps using defeasi- tively. Like much of the work pre- tonic reasoning community that ble rules (which can be traced to the sented at the 1991 symposium, it makes Brewka’s work possible. Brewka philosopher Ludwig Wittgenstein by extends the idea of Toulmin diagrams uses the precision of his revised way of philosopher Friedrich Wais- for arguments into the world of default framework to answer ques- mann). However, McCarty took the graphic user interfaces. It is ambitious tions that Rescher did not or could opportunity to assail AI’s work on work in part because all the research not answer. I disagreed with Doyle defeasible reasoning. He said that that builds on the GIBIS (Conklin) and over the long-term importance of pre- defeasible rules were trivial devices AQUANET (Marshall et al.) experiences cisely answering exactly these ques- that could be eliminated in favor of points to the average user’s unwill- tions; I felt that variations of the two kinds of negation. As devices for ingness to frame arguments formally. model remain to be explored at a rule-based programming, they pro- In addition, it insists that the argu- higher level of abstraction rather than vide no advance over intuitionistic ments to be diagrammed be scientific to obsess over one variation’s details. logic programming (says McCarty). arguments. Surely, scientific argu- Still, a full literature search of cita- They are trivial, that is, when com- ments have their own special logic. tions of Rescher’s 1977 monograph, pared to the defeasibility of open-tex- Cavalli-Sforza has for a while been Dialectics, reveals no useful formal tured concepts, the logic of which interested in Toulmin’s own attempts extension or clarification of the logi- remains unanalyzed (says McCarty, to apply his work on argument to cal system prior to Brewka. The downplaying the work of case-based specialized forms of reasoning, such Argentine constructive logician Gino reasoning researchers, especially his as legal, decision-theoretic, or scien- Roetti is the only person who has colleagues Edwina Rissland, Ashley, tific. BELVEDERE is significant because it told me he read the book in 1977, and David Skalak). Ironically, the AI made good of Toulmin, with possible saying he knew then that is was an and law luminary, McCarty, like the ramifications in the interactions of important little book but that it was deontic legal logician Carlos Alchour- users with scientific databases; it is perhaps ahead of its time. Was an ron, chooses to attack a legal inven- perhaps more significant in this understanding of it really impossible tion (defeasible reasoning) on techni- respect than it is successful as a tool until AI had had its nonmonotonic cal aesthetics. for teaching young scientists. reasoning excursions? The issue of open texture was Freeman gave a synopsis of her dis- McCarty’s paper was as electrifying made more pointed by a paper pre- sertation work, which also included a as it was an object of impassioned sented by Pierre St-Vincent. He and graphic system for diagramming disagreement. Although primarily Poulin gave a nondialectic, nondefea- arguments based on Toulmin. Her historical and not essentially con- sible treatment of open-textured con- system provides an automated rea- cerned with the logic of dialectic, it cepts. These Canadians wanted to soner for additional support of the was recognized at once for its impor- analyze open texture as a normed or user. Although her representational tance. McCarty noted that it was the fuzzy concept, in the manner that scheme seeks to mix dialectic ideas greatest Anglo-American philosopher might be appropriate at a meeting with measures of plausibility, the of law, H. L. A. Hart, who introduced about uncertainty in AI. A role for audience found nothing objection- the logicolinguistic world to the term dialectic remained in St-Vincent’s able in her work. This reaction shows defeasible (Hart imported it from approach: Dialectic provided for an me that the research program that English contract law). These origins exchange of messages between advo- seeks to import graphic user interface are perhaps of greater interest to the cates interested in the contrary aims versions of Toulmin’s ideas to AI and nonmonotonic reasoning communi- of enlarging the scope of a concept computer-supported collaborative ty because this community has a and restricting it. This paper would work (CSCW), a paradigm present in stake in a better understanding of be an excellent place for fuzzy logi- 1991 and mentioned in the recent defeasible reasons; still, it is clear that cians to approach the issues raised by Association of Computing Machinery dialectic presupposes defeasible rea- this workshop. computing-survey article on CSCW, sons on which to construct argu- These papers, instead of pushing is on strong footing. ments that can stand in opposition. models of dialectic forward, involved Bench-Capon and P. H. Leng’s McCarty raised the question of why debating the current understanding paper was self-explanatory: develop- Hart never used his term, defeasiblity, of defeasible reasoning and dialectic. ing heuristics for the argument-based in his work again. There was The paper of Violetta Cavalli-Sforza explanation of negation in logic pro- superficial evidence that Hart even and Dan Suthers and that of Kathy grams. The authors extend Bench- disavowed his early paper that intro- Freeman were counterpoised. Capon’s work presented at a series of duced the idea of a defeasible con- Cavalli-Sforza and Suthers reported expert system conferences: “Interact- cept, and the modern lore of jurispru- on BELVEDERE, “an environment for ing with Knowledge-Based Systems dence holds that he did disown the practicing scientific argumentation.” through Dialogue Games (with P. E. paper and its ideas. BELVEDERE is a tutoring system for high S. Dunne and P. H. Leng, 1991), For Hart, the early idea of a defeasi- school and junior high school science “Using Toulmin’s Argument Schema

102 AI MAGAZINE Workshop Report to Explain Logic Programs” (with D. represent which arguments are pre- had just produced the nifty mono- Lowes and A. M. McEnery, 1991), “A ferred over their opposing arguments. graph Rules of Encounter (The MIT Dialogue Game for Dialectical Inter- Some people will think their analysis Press, 1994), in which the economic action with Expert Systems” (with P. is backward, but this is the novelty. approach to achieving agreement is E. Dunne and P. H. Leng, 1992), and Past authors have imported the used to derive interesting results in “Argument-Based Explanation of the axioms of utility into the analysis of distributed AI. Compared to their British Nationality Act as a Logic Pro- multiple nonmonotonic extensions methodology, the emerging meth- gram” (with F. P. Coenen and P. (M. Wellman and J. Doyle, “Impedi- odology of computational dialectics Orton, 1993). The premise in all this ments to Universal Preference-Based seems troubling. Rosenschein and work is that negation as failure is Default Theories, AI Journal 49(1): Zlotkin prove interesting theorems inadequate as a report to a user. In its 97–128). The idea here is to use the about games; computational dialecti- stead, a dialogue game is played in machinery of defeat among argu- cians give rules for interesting games. which the automated reasoner uses ments to replace real-valued utility its recently completed search to when decision making requires mak- defeat a player who seeks to establish ing trade-offs. Instead of saying that the contentious proposition (against money and time can be mapped to the advice of the program). The user’s utilities, say that an argument for a stubbornness as a player in such a decision based on consequences of game defines the amount of informa- one kind (one attribute) defeats an Understanding the tion that is required to explain the argument for a different decision interaction as result: the more stubborn the use, the based on consequences of some other dialectic, as a longer the dialectic game, the larger kind (other attribute). It is a qualita- the explanation generated. Under- tive approach to multiattribute utility. two-player standing the interaction as dialectic, The idea seems initially sound, adversarial game as a two-player adversarial game of although it is perhaps pushed too far. responses, is novel. It is what compu- Preferences among arguments are of responses, tational dialectics brings to AI’s foun- sometimes based on superficial syntac- is novel. dations, with ramifications in the tic features, such as specificity among design of systems. defeasible reasons. These preferences It is what I believe that Allen’s paper, Gordon might be used purely for notational computational and Brewka’s paper, and my own convenience. If so, then the availabili- dialectics brings to could also be regarded as novel. Allen ty of only one kind of preference gave the rules of an A-Hohfeldian might be representationally impover- AI’s foundations, game, which some might properly ished rather than merely clever. with ramifications regard as a Wff’n’Proof (a 1960s board My paper, “Argument and Arbitra- game that taught logic) for logical tion Games,” sought to include an in the design statements using the turn-of-the-cen- argument game within the various of systems. tury, heretofore underappreciated, games of negotiation. The approach largely deontic ontology of the legal of game theory to negotiation is to scholar Wesley Hohfeld. The work characterize it as a series of proposals. was jointly undertaken with Charles Sometimes settlement is based on Saxon. The Allen-Saxon revision of reason, although as Sycara notes, Providing rules for a game without Hohfeld’s ontology first saw good use sometimes the reason is merely that justifying or analyzing these rules in an expert system reported at the one person is in a better bargaining might be the right thing to do and 1991 First International Workshop on position than the other. It is unfortu- might be no different (in the Deontic Logic. In our Seattle work- nate that Arthur Merin was unable to abstract) from what people do when shop, the work took a turn toward give his paper along with mine they design AI programs. Describing formal games, renewing Allen’s inter- because the two seemed to be in line such discourse games leads to neither est in producing the rules of adversar- with each other. The work has just theorem nor consensus. It just ial logic games. Allen hinted that begun, but I am already perched on describes a regimen for the interac- rules for a fully disputational dis- the limb that holds that future tion of a society of minds. Clearly, course game, that is, his ideas for a economists’ models of negotiation this regimen is important, but there model of dialectic disputation, were will integrate discourse and reason. must be some way to evaluate the in the works. Cheap talk and precedent are simply merit of the resulting models. Com- Gordon and Brewka gave a paper too important as parts of what can be putational dialecticians can defend about decision making. The prefer- considered negotiation. their results and develop the kinds of ence ordering required for evaluating Zlotkin gave a synopsis of his work taxonomy that logicians have for trade-offs in decision making was and his reaction to the thrust of the their wares. They can also inspire assimilated into the ordering used to workshop. He and Stan Rosenschein better programs, as seen in the pre-

WINTER 1995 103 Workshop Report

sentations of Cavalli-Sforza and Suthers and Bench-Capon and Leng. More work needs to be done. We had our share of students com- pleting dissertations: George Fergu- Computation son (under James Allen at Rochester), Jennifer Chu-Carroll (under Carberry at Delaware), Vincent Aleven (under Ashley at Pitt), and Jeremy Werth- eimer (with John Mallery at MIT). &Intelligence Although each has, in the past, writ- ten papers and programs that were centrally relevant to the theme of the Collected Readings workshop, all remained surprisingly quiet. This problem is a major failing of a workshop. Although the topic EDITED BY GEORGE F. L UGER and its discussion heavily favored law and logic interdisciplinarians, there is no future for a field that does not promote the development of its his comprehensive collection of 29 readings covers artificial intel- future scholars. ligence from its historical roots to current research directions and practice. With its helpful critique of the selections, extensive bibli- The biggest disappointment was T that Gerard Vreeswijk could not ography, and clear presentation of the material Computation and Intelli- gence will be a useful adjunct to any course in AI as well as a handy refer- attend because his dissertation work ence for professionals in the field. neatly defines the area. His work seeks explicitly to model dialectic as a The book is divided into five parts. The first part contains papers that discourse game in logic. He is now present or discuss foundational ideas linking computation and intelli- taking the same paradigm of compu- gence, typified by A. M. Turing’s “Computing Machinery and Intelli- gence.” The second part, Knowledge Representation, presents a sampling tational dialectics and addressing dis- of the numerous representational schemes—by Newell, Minsky, Collins tributed AI problems of coordination and Quillian, Winograd, Schank, Hayes, Holland, McClelland, Rumel- within it. I see Vreeswijk’s remaining hart, Hinton, and Brooks. The third part, Weak Method Problem Solving, on the other side of the Atlantic focuses on the research and design of syntax based problem solvers, Ocean as a metaphor. Much of the including the most famous of these, the Logic Theorist and GPS. The best work on this topic will likely fourth part, Reasoning in Complex and Dynamic Environments, presents remain on the continent, where the a broad spectrum of the AI communities’ research in knowledge-inten- sive problem solving, from McCarthy’s early design of systems with temperament is more intellectual, the “common sense” to model based reasoning. The two concluding selec- logical traditions are more varied, tions, by Marvin Minsky and by Herbert Simon, respectively, present the and the relevance of paradigm to recent thoughts of two of AI’s pioneers who revisit the concepts and con- application is better welcomed. Like troversies that have developed during the evolution of the tools and tech- AI and law, the conditions for leader- niques that make up the current practice of artificial intelligence. ship exist outside the United States. We were lucky in Seattle that the ISBN 0-262-62101-0 American conference was able to take 700 pp., index. $35.00 softcover the lead for a year. The AAAI Press • DISTRIBUTED by The MIT Press Massachusetts Institute of Technology Cambridge Massachusetts 02142 R. P. Loui is associate professor of com- To order call toll free: puter science, affiliate associate professor (800) 356-0343 or (617) 625-8569. of legal studies, and research affiliate of MasterCard and VISA accepted. the Center for Semantic Control and Opti- mization at Washington University in St. Louis, Missouri.

104 AI MAGAZINE