\documentclass{llncs}
\begin{document}
\title{Managing Large Scale On-line Discussions: Secrets of the Open
Meeting\thanks{The on-line meeting described herein was a collaboration of the
Artificial Intelligence Laboratory, the National Performance Review,
Lawrence Livermore National Laboratory and Mitre Corporation. Jonathan
Gill, Thomas Kalil, Randy Katz and Howard Shrobe provided more support.
Research was partly supported by the Defense Advanced Research Projects
Agency of the Department of Defense under contract number
MDA972-93-1-003N7.}}

\titlerunning{Secrets of the Open Meeting}
\author{Roger Hurwitz and John Mallery}
\institute{The Artificial Intelligence Laboratory \\
Massachusetts Institute of Technology}


\maketitle
%
\begin{abstract}

The possibility of communities steering themselves by large scale,
online meetings are limited by support system capabilities for managing
their discourse.  An asynchronous conferencing system the Open Meeting
supported a meeting in which thousands of U.S. government workers
discussed reengineering government services.  It enabled users to learn
and share opinions about proposed changes, because its design focused on
decomposing the inflows of comments, structuring multi-lateral
conversations and maintaining civility.  Access was provided over SMTP
and HTTP to a topically differentiated hypertext synthesized from an
object database, which was extended by users' comments. The comments
composed virtual conversations, structured by a discourse grammar that
constrained what types of comments could be attached in specific
contexts.  Civility benefitted from use of moderators and from
participants' having similar bureaucratic culture backgrounds.  The
results suggest that cultural diversity is a greater challenge than
numbers to enabling effective on-line democratic action.

\end{abstract}

 
Can the Internet invigorate the democratic practices of large
communities?  Will it become a new public sphere\footnote{The term
public sphere was brought to contemporary attention by Habermas
\cite{1}, who links the decline of public spaces for political
discussion to the rise of mass media.}, a space where people gather to
rationally discuss issues in the hopes of reaching consensus and
influencing policy makers?  Expectations for that have been raised by a)
the diffusion of on-line community forums and newsgroups, b) the power
of the Internet to quickly and cheaply gather dispersed individuals
around common interests, and c) its use by governments and political
activists to distribute information.  Parties, candidates and even
rebellions now have web sites and email.  As expressed in the ugly
buzzword {\it disintermediation}, digital networks can create direct
links between people and government, bypassing the mass media, interest
groups and other gate keepers.

Some argue that the Internet will not nurture a public sphere, which, in
any case, is an inappropriate model for democratic practice.  Its
earlier manifestations in Berlin salons and American civic associations
were unrepresentative and exclusivist; its rhetoric, abstractions and
self-evident truths assume an underlying homogeneous society, steered by
white, literate and propertied males\cite{2}.  On this critical view,
characteristics of on-line communications as well as the growing
diversity of voices on the Internet prevent the realization of a public
sphere, even one accessible to all and where difference is acknowledged.
These include a bias toward speedy, unreflected responses, the absence
of social context cues and hyperlinking which, pointing a reader
elsewhere, dissolves the authority of the text at hand.  On-line
political discussions will be situated, personal, and chaotic; more
likely to immobilize a community than help form its collective will.

This paper joins the debate by examining how the Open Meeting an
experimental network-based information system brought together many,
dispersed people to discuss government policies, which affected their
work lives.  The system's architecture and its theory of knowledge
representation have been described elsewhere\cite{3}; the present focus
is on how the system organized the large scale discussions, particularly
with regard to managing inflows, structuring interactions and
maintaining civility.  The achievements and shortcomings of the
experiment are examined for their relevance to issues of on-line
governance. Clearly just broadening people's access to one another and
to information does not generate participation, order discourse or
produce group decisions.  So it is important to show how it is possible
to get some of these results.

\section{Overview of the System}

% 

The Open Meeting system was initially designed and implemented in fall,
1994, to support an on-line meeting, in which several thousand United
States government workers conducted virtual discussions of recommended
changes in government operations and performance standards.  At the
system's core is a persistent object database, which coupled to a task
engine, supports email and web servers.  These present messages and
pages, with content generated on-the-fly by the database, in response to
users' inputs through queries, remote commands or interactive forms.
Inputs can include users' comments which are added to the database and
reconfigured as threaded conversations or annotation overlays.  The
recommendations for discussion had been prepared by the National
Performance Review (NPR),\cite{4} a unit in the Office of United States
Vice-President Al Gore, with the mission of "reinventing government."
It believed a large scale on-line meeting would

\begin{itemize}
\item Increase the visibility and support for these recommendations;
\item Enhance their prestige through association with new technology;
\item Gather reports on promising practices in line with them;
\item Bypass organizational boundaries that keep employees in
different organization from discussing common work problems; 
\item Overcome geographical barriers that confine discussions on
government to the capital; 
\item Demonstrate the possibility of on-line/ electronic governance.  
\end{itemize} 
% 
As the last point suggests, in the NPR
view, government workers comprised a community whose opinion needed to
be articulated and mobilized.  The NPR project team\footnote{Led by
Larry Koskinen and Andy Campbell} for the meeting was guided
by the metaphor of a New England town meeting, where any member can of
the community can
speak to issues of common concern. It envisioned an on-line
environment analogous to a town hall, that would include chat rooms for
an assembly hall, committee rooms and water cooler.\footnote{The
Clinton-Gore team used face-to-face "townhall meetings" 1992
presidential campaign and later in the Clinton I administration to get
public attention for its issues and positions, while bypassing the
national mass media.  This populist strategy was highly successful with
citizens and voters, but soured the team's relationships with the
journalists, editors and owners of the media.}

The trajectory toward a community activation system changed with the
involvement of the Open Meeting developers, members of an MIT group,
whose research concerned intelligent networks that route and enhance
messages, based on understanding of their contents and the organizations
through which they flow.  At the time, the research group was planning a
web based system that would manage public participation in government
inquiries on proposed changes in regulations.  The initial plan called
for interested parties to attach their views as annotations to the
proposals, differentiated according to the types of comments being made,
e.g., question, statement of support.  Officials could then retrieve
these comments by target and type from a textbase of review documents,
using the system, in effect, as an annotation server.  Alternatively,
the comments on a document could be routed directed to officials whose
profiles indicated their in that document.  By enabling the public and
officials to receive and reply to each other's comments, the system
would support virtual discussions composed of threads of typed comments.
Through the prism of this plan, the federal workers appeared like
individual stakeholders responding to inquiries on proposed changes in
regulations affecting their individual interests, work rules and
performance standards.  They were primarily sources of information about
the benefits, feasibility and acceptability of the proposals.

To encourage broad participation and knowledge sharing, regardless of
organizational differences, we consequently advised that the meeting
focus on eleven NPR reports, each of which concerned reinventing an
operating system, like human program development or information
management, found in all departments and agencies and subject to the
same rules and standards.  These could provide common grounds for
discussions across organizational boundaries, but recommendations for a
specific department were unlikely to draw workers from other departments
into a discussion.  We also advised dropping the town hall format for
several reasons:

\begin{itemize} 
\item a reluctance to assume the participants already
recognized themselves as a community;
\item the desired dispersion of participants over time zones made
asynchronous communications more feasible; 
\item chatrooms, MUDs and MOOs would not scale for the desired number of
participants; 
\item email would not support the GUIs which could suggest a town hall
setting, and email was mandatory because relatively few workers then had
web browsers.  

\end{itemize}


The group proposed instead a system similar to the public comment system
sketched above, as an application layer for Comlink, an enhanced document
distribution system, developed primarily by Mallery and used for the
indexing, publication and retrieval of White House electronic releases
on the Internet.\cite{5}.  The principal features and functionalities of
the resulting system included

\begin{itemize}

\item {\it An extensible, persistent object database} whose instances
include texts and users.  The text instances include hypertext links and
indices, and so support extension of an initial textbase configurable as
hypertext; arbitrarily constrained, on-the-fly construction of local
hypertext; and views of the hypertext, restricted by arbitrary indices,
like quality ratings, security classifications, etc.  The user objects
can include slots for interests and preferences, enabling subscription
to and routing of texts.

\item {\it Dynamically generated web pages} which facilitate
representation and navigation of the online discussions and
collaborations.  Regions of the resulting hypertext can be isolated;
pointers to comments and annotations descending from an initial text can
be displayed in an outline or tree, each with an icon for the type of
the comment.  This permits both overviews of the link structure and
direct access to those comments which interest the user.

\item {\it Interactive form processing} to support user queries and
comments, hypertext traversal, and search interfaces for retrieval by
indices.

\item {\it Message threading} that constrains input according to
conversation or discourse grammars.  Grammar means a set of (context
sensitive) rules that specify the types of texts or comments that can be
linked to another text or comment, according to its type.  By type, we
mean remarks that constitute recognizable moves or roles in discourse or
communicative interaction, e.g., remark types in a debate include claim,
challenge, rebuttal, question, answer, etc.  A grammar thus represents a
procedural order, often with possible branching.


\item {\it Subscription to arbitrary node:} A retrieved text was wrapped
in a form which could be used to request emailing of comments
that are later attached to the text or any of its descendants.  Since
the particular text as a database object is automatically identified in
the form, subscription is transparent:  The user does not have to
specify the node.

\item {\it Moderator tools} enabling submissions to be handled
before being posted to public view. These include a) a moderator's view,
b) virtual queues for routing submissions on a topic to its moderator,
c) forms for attaching ratings and view restrictions, c) form letters
accepting and rejecting submissions.  With these, the Open Meeting can
support moderated as well as unmoderated online conferences.
\end{itemize}

\section{Meeting the Challenges}

This technology better met the NPR goals than chatrooms or list-serves,
but it put more emphasis on the meeting as interactive knowledge
building than the NPR team originally planned and had less regard for
the meeting as community activation.  Its deployment was shaped by our
recognition of the major challenges that a large on-line meeting
presented to the system's potential for organizing texts and building
knowledge: Managing the fan-in of comments on the proposals, structuring
the interactions of participants and maintaining civility among them.

\subsection{Fan-in} 

In hierarchically structured communication flows, viz., 1-to-many/
many-to-one, a request for comments to a community can produce a stream
of responses likely to overwhelm the source of the request, but miss
other interested parties.  Directing the response stream to all members
of the community in a many-to-many pattern just compounds the problems
of message volume (gain) and recipients' discrimination of messages that
interest them.  A strategy to meet this challenge is the decomposition
of both the request and response streams, according to a taxonomy of the
issue space.  This was done for the on-line meeting by considering the
NPR reports on the operating systems as subdomains and their respective
proposals as categories in a reinventing government domain, and
requiring that a comment be subsumed by a proposal.  A comment either
had to target a specific proposal or another comment descending from it.
Instead of being asked "what are your views on reinventing government,"
a participant was in effect asked:  "Should government offices be issued
credit cards for making small purchases?" by the presentation of that
proposal.  Users could find proposals and later comments on them that
interested them by drill down through the hypertext o rhrough a
taxonomic listing.  They could subscribe to receive comments added to
such a proposal or, even more finely grained, to a node subsumed by it.

Decomposition of the domain was facilitated by NPR having prepared each
of the reports according to the same strategic-plan like format: an
Executive Summary described the problems in the respective system, a set
of Recommendations proposed how to fix them, one or two Actions were
described for effecting each Recommendation, and Appendices added some
details of the implementations.  These parts were easily reconfigured
into a hypertext, using a standard node architecture, that replicated
the structure across the main branches of the hypertext (the individual
reports), simplifying implementation and providing a consistent user
interface.  Eleven such nodes, one for each operating system, were
linked to a root document which presented the plan and procedures of the
meeting.  The standard node eventually had links to the parts of the
report, including a link to each recommendation, and to other relevant
documents, including a short Overview of the system problems, official
examples of practices that satisfied a recommendation, and newsletters,
which summarized the ongoing discussions of the proposals.

The decomposition is based on the idea that knowledge can be
hierarchically localized in the sense that understanding in one branch
or area of a domain requires knowledge of more general domain principles
but not of fine grained propositions in other branches.  The
decomposition was intended to reduce the number of people in a
conversation and the circulation of their comments, enhance their
interest and focus and make the conversational threads easier to follow.
The resulting topology was small circles of lateral communication, i.e.,
several dozen to several dozen, with isolation of conversations from one
another being overcome in practice by users' visiting several
conversations.  Systematic, updated overviews of the discussions
regarding an operating system were provided by the newsletters.  These
were hand-made summaries by the moderators, but automatic summarization
or gisting which exploits the hypertext links is an important focus of
work to extend the system.

\subsection {Structuration} 

The challenge of devising an interaction structure for the meeting had
two sides:  First, the need to enable orderly, yet relatively
unconstrained conversations, through which knowledge and information
about the proposals could be built; second, the need to produce a
coherent textbase which would index each comment according to its type
of move in a conversation. The messages posted to the meeting were
imagined as both pieces of ongoing conversations about particular
proposals and parts of a self-extending, distributed hypertext,
contributed by multiple sources, including NPR.  However conventional
messages threaders, such as subject, target ("in reply to"), submission
time, were deemed insufficient to structure or represent a set of
messages as a sequence of conversational or rhetorical.

Indexing the intentions of the moves is important:  First, on the view
that conversations are composed of interlocking moves or speech
acts\cite{6}, on-line communications are problematic, because they have
fewer inflection, non-verbal and setting cues which speakers in
face-to-face communications exploit to make and recognize speech acts,
their intentions, and the expected responses.  When the move types of
posted comments are indicated, a user can know what types are
appropriate responses to a target she has chosen, and by identifying the
type of her comment, she can indicate the type of response she expects.
Second, when comments are threaded through their targets in hypertext
display, the tagging clarifies the relationships between otherwise
opaque texts (especially when the text itself is not displayed but only
pointed at by a link), and provides an overview of the flow of
intentions and expectations through the entire conversation.  Finally
indices are needed so for a given target, all replies of the same
attitude, agreement for example, can be displayed together, retrieved
together or more easily checked for redundancy.

The indices can be obtained by asking contributors to identify the types
of moves they are making.  This self-indexing is similar and not
necessarily more inhibiting than the widespread practice of putting
"smilies" and other cues into email messages to reduce their
ambiguity.\footnote{Multi-media messages will not entirely eliminate
the problem, especially when speakers do not know each other or are from
different cultures.}.  Self-indexing of utterances is also a standard
linguistic practice, especially when there is a probability of
misinterpretation or a particular type of move is privileged, in the
sense of having a priority on getting attention, e.g., "I have a
question?"\footnote {Self-indexing is often used in political discourse
to emphasize the intention of certain utterances.  One famous example is
Zola's attack on the French government for its cover-up in the Dreyfus
affair: {\it J'accuse}.}

We also imagined that the sequences of moves and the discussions they
created would be constrained by a context sensitive discourse grammar, a
set of moves and rules that restricts their possible combinations by
specifying the moves that can be legally attached to each move.  These
rules abstract the notions of appropriateness and expectation that are
associated with the move types in a sequence.  Since they have
institutional as well as logical bases, a grammar encapsulates a quasi
normative order or a procedure employed for a social purpose, not just
idealized linguistic phenomena\footnote {People are often upset because
discourse grammars work to mitigate the threats to face -- people's
power and dignity -- that is inherent in social interactions.
Violations of the grammar are consequently implicit attacks on
face\cite{7}.} The sequences which we thought most likely to occur and
best to support on-line were multi-lateral arguments for and against the
proposals.  Our model of argument was influenced by modern dialectical
theories, which understand argument as an ordered, interactive process
whose utterances are constrained to speech acts for making and
supporting claims, challenges and rebuttal.\cite{8,9,10}.  The highly
restricted move sets of the dialectical argument models were attractive
because they barred the flames, digressions, emotional discharges and
similar failings of newsgroups.  However, we anticipated that
discussions constrained by that grammar would be more conflictual and
address fewer possibilities than if the grammar also permitted
alternatives, examples, information seeking questions and answers.
Because the Open Meeting was intended to stimulate interest and
accumulate information about government reinvention, the teams selected
moves that could lead to consensus while permitting differences of
opinion:  Agreement (reason for), Disagreement (reason against),
Question, Answer, (propose an) Alternative, Qualification ("yes,
but"), or (report a) Promising Practice. The root document explained
these types, presented a distinctive icon for each, and asked users to
frame comments according to them.

The grammar for the meeting was additionally shaped by pragmatic
institutional as well as logical constraints.  To avoid the scent of
electronic plebiscite, polling, voting and simple endorsements were not
implemented.  Agreements and Disagreements had to provide reasons.
Because the executive summaries, the overviews based on them, the
appendices and the example promising practices were approved by NPR
leadership, they were in a sense above debate, at least by workers at
the meeting and no comments could be attached to them.  Finally, the
attachment of alternatives to questions, answers or promising practices,
and the attachment of anything but answer to question were considered
illogical and excluded.

The discourse grammar is weakly enforced in the Open Meeting system by
using it to drive the dynamic reconfiguration of input interfaces.
Given the user's choice of target, the server presents an input
interface which limits the choice of comment types to just those which
can be "legally" attached to the target's type.  Of course, this does
not prevent a user from flouting the rules, by identifying a comment as
a permitted move, but actually making one that is not.  A common case of
flouting in political meetings is the use of a turn in the question
period to assert one's own position, sometimes eliciting the rebuke:
"That's not a question?"   In general, to prevent potentially disruptive
flouting requires natural language processing capability that
understands the intention of utterances -- not just recognizes forms.
Absent robust natural language processing technology for this, human
moderators are needed to verify that the contributor has not flouted.
As noted above, moderators for the meeting did verify that Agreements
and Disagreements gave reasons, but verified no other move types,
because of the low potential for disruptive flouting.

The resulting virtual discussions in the Open Meeting differed from
face-to-face, multi-lateral discussions in not having an inherent
linearity, that is, forward movement toward conclusions.  In
face-to-face conversations, discourse contexts or foci of attention are
successively closed to further comments and are only reopened with
difficulty\cite{11}.  A speaker might present a reason against a
proposal, with the reason then becoming the context of subsequent debate
among speakers.  At some point, however, discussion returns from this
stack to the top level and moves to another point, opening another
context.  There will be no backtracking to the first context, unless
information or conclusions are premises that have been contradicted in
the current context.   In the virtual discussion on a proposal, all the
contexts (and their contexts recursively) remained open to further
comment, even if recent comments were in the same context.  The
discussion could therefore develop as a set of statement and reply
sequences on different points, a long sequence elaborating one point
(analogous to breadth versus depth search) or a mix of both approaches.
This openness partly reflects the primary orientation of the meeting
toward collecting information and building knowledge rather than moving
a community toward a consensus.

Unfortunately, the outline form used on the web page to represent the
links in the hypertext obscures the time dimension and possible focus or
convergence in the discussion.  This regime recursively indents the
pointer to a comment with its move type icon under the pointer to its
target, with comments at the same level listed in order of submission.
As a result some comments at a deep level that were more recent than
other comments at a more shallow level can higher on the page.  Since
all submissions are time stamped, it is trivial to present the hypertext
in ways that could encourage or represent convergence, such as hiding or
closing contexts that have received no comment for an arbitrary time and
color coding contexts according to how recent are the last comments in
each.

\subsection {Civility and Moderation}

The design included several means of maintaining civility and the
quality of interactions.  Users were prodded to self-reflection and some
restraint on impulsive commenting by having to identify comment type and
summarize the content in the a subject line.  A user was also unable to
disguise her identity, because had to submit the comment on a form that
was mailed to her email address\footnote{This method is no longer
useful, due to the proliferation of free mail services.} In addition,
moderators were used to minimize the posting of low quality, redundant
and inappropriate comments.

The moderators' tools included virtual queues to allocate work, review
forms, form letters to users, and a constraint-based view system.  These
were used to overview all submissions to the meeting, access unreviewed
and pending submissions on a topic, rate a submission's quality and
decide its status.  Options for the last function included accepting the
submission for exposure to users, rejecting it or returning it for
revision with a form letter explaining the reason, and deferring a
decision to another moderator.  The moderators in the on-line meeting,
volunteers recruited by NPR and assigned to specific nodes, were trained
to reject submissions that were low quality, redundant, obscene,
personal attacks, whistleblowing or commercial solicitations. By
checking on the review form the reason for a rejection, a moderator
automatically returned the submission to the user with the appropriate
letter of explanation.  Much of this labor can be further reduced by the
development of rule and neural network based programs that can
automatically detect some prohibited content types.

Moderation exploits the database support of views, since accepting a
comment merely included it in the public view. As implemented,
moderators could see all submitted documents with their review status
and quality ratings, while the general user saw or received only those
that had been accepted for public display.  More generally, views are
displays generated by constraints that determine what gets shown to
whom, display of the textbase and can be apportioned by arbitrary
criteria, e.g., security classifications, organizational roles.  A role,
like moderator, then can be defined as a capacity to make certain moves
towards others, based on the information the role holder gets or sees.
Information is thus the context for interactions and power relations in
the group.

The discourse grammar and the moderator reviews were constraints on what
was said during the meeting, but there were factors that diminished
their potential for arbitrary suppression of voices by the moderators.
First the grammar itself was fairly transparent and was consistent with
the announced purpose of the meeting as a discussion.  Second, the
moderators for the meeting were peers of the users, shared an
interest in a meaningful discussion of the recommendations, and had to
give reasons for rejecting a submission.   Nevertheless, it is
worthwhile comparing the risks that having rules of order and moderator
authority poses for democratic practice with other means of control in
on-line forums.

\begin{itemize}

\item {\it Unmoderated forums}, bulletin boards, mailing lists, etc.,
depend on the civility and common sense of their users.  They are
plagued by digressions and flame wars which drive away users,
effectively suppressing diverse opinions.  Efforts by some users as
self-appointed "thread police" may be counter-productive. An option to
improve control is collective moderation in which all participants rate
the value, relevance and civility of the posted comments, with users
receiving low scores being automatically barred.  Certified registration
could foil attempts by exiles to post under assumed names and also
determine who is a participant.

\item {\it Moderated forums} involve posted guidelines and
moderators who usually preview (sometimes review) submissions to block
(remove) from posting those violating the guidelines.  The costs are a)
possible abridgment of free speech rights\cite{14}, b) moderators seeing
themselves as proprietors of the forum, c) moderators' interpretations
of the guidelines are opaque regarding how and what they judge are
violations.  Users, after all, do not see what is not posted.  The use
of content analytic as filters on submissions would reduce the need for
moderators and enable collective reviews of rules or examples of
violations more fine-grained than the initial guidelines. This leaves
open the questions of who writes the rules or chooses the egregious
examples and how intelligent do the filters need be.  

\item {\it Adding Roles to Discourse Grammars} is a means to reduce the
potential for aberrant, disruptive interpretations of the situation,
that can be enacted within the constraints of the grammar.  A cheap
implementation of this idea is providing more directions to users in the
introduction to the meeting, including the roles they might imagine
themselves playing and there motivations in these roles.  A more
complete fix would be the support, similar to that given moderators, of
roles that are either help the group pursue the discussion or manage
itself.  For example, a resident expert with access to specialized
sources may be able to settle questions of facts that arise in the
discussions.  In contrast, a mediator could handle the problem of two
participants using the forum as a private battleground by opening a back
channel to them with suggestions to resolve their differences, or just
remove their exchanges from the view of other participants. Increasing
the organization of the meeting also increases disparities in power and
social relations, since these are based on the information and moves
available to the respective users.

\end{itemize}

\section{Discussion}

\subsection{Use Patterns} Announced over specialized mailing lists and
bulletin boards for government workers, the meeting attracted the
interest and participation of workers from a wide and geographically
dispersed range of federal government offices and military
installations.  By its beginning in early December, 1994, over 4200
people had returned an on-line registration survey.  90\% were
government workers, with over 60\% located outside the Washington, DC
area, including in all fifty states and twenty countries besides the
United States.  They represented 40 executive branch agencies and
departments, with the large and more technically oriented organizations,
viz., Defense, NASA, Interior, tending to have the most representatives.
They were typical of government workers with respect to age (average in
early 40s) and experience, but significantly more senior, educated, and
technically oriented (60\% vs. 25\% in supervisory capacities; 47\% vs.
15\% with MA or more; 66\% vs. 11\% in information systems or
engineering).  Women were under-represented (30\% vs. 42\%), but their
rate of registration was double the estimate percent female of Internet
users at the time (15\%).  These distributions indicated that interest
and network access were key factors for participation. The registration
rate of women, despite their lower levels of employment in technical
positions, and the large registration at Housing and Urban Development
were due to special outreach efforts by NPR personnel to these groups.
In contrast, few Treasury and National Security Agency personnel
registered because these offices had firewalls blocking access to
outside networks.

Over 3000 users accessed the meeting during its two week, but only about
2000 went beyond the introductory page and welcoming pages to a standard
node and only 1000 attended the discussion threads.  Several factors
accounted for the attrition: End-of-the-year and holiday chores left
workers little time for the meeting; they found email traversal of the
hypertext too cumbersome and the texts themselves overwhelming, and may
user were interested only in information and not discussions.

The distribution of web accesses by file types shows two patterns among
users accessing a standard system node.  About two-thirds of these users
surfed the meeting material in several reinvention topics, looking
primarily at the overviews and newsletters when available.  Users in the
second group tracked online discussions on several proposals apiece, in
one or two reinvention topics, through both retrieval from the hypertext
interface (16 comments read on average), and subscription to mailings of
comments posted to the thread (as indicated by the concentration of
subscriptions on those topics and proposals which were most accessed).
Those who commented were necessarily in this group and similarly focused
in their comments; with few exceptions, they posted to just one or two
discussions.  There was relatively little use of the tools in the
meeting environment for search directly by reinvention topic, comment
type, government agency referenced and other indices.  On the whole,
these patterns suggest an orientation toward the meeting as a source of
general, rather than specific information, and an interest on the part
of the in-depth users in what their peers had to say on a few specific
issues.

290 different individuals contributed 1013 comments that were accepted
for posting.  The ratio of one contributor for five actual lurkers at
the discussions is much higher than the 1 to 10 or 20 ratio variously
estimated for newsgroups. Half the comments were identified as Agreement
and 15\% were Disagreements. The relatively large number of Questions
(167), Alternatives (106) and Promising Practices (72) indicates the
contributors' willingness to seek and share new information at the
meeting.  The near absence of the cognitively complex Qualification (3)
supports the contention that cyberspace does not nurture the reflection
desired in a public sphere, although a sampling of comments found that
some qualifications were misidentified as agreements.  Possibly these
users did not want their comments to be construed as unsupportive.  As a
group, the commenters were positive toward the proposals and tried to
present constructive ideas, some of which NPR leadership culled for
further examination.

Web users were generally satisfied with the system, email users
complained about its clumsiness for traversing hypertext, and both
groups complained about an overwhelming amount of text to read.  These
responses, particularly the last, were anticipated.  We believed that,
consistent with their bureaucratic training, people attending the
meeting would want to read the "manual" (the reports) before commenting
on the proposals, so we wanted to reduce the texts to much shorter
pieces.  Because, the NPR team lacked the authority and personnel to do
the editing, the overviews were produced as a means for users to get
acquainted quickly with reinvention topics.

\subsection{Discourse Management} 

The on-line meeting demonstrated that virtual discussions localized to
specific topics in a domain could attract people across organizational
boundaries, remain coherent and produce possibly valuable ideas.   What
factors account for the success?   Can they be replicated in other
venues and with other groups?

The meeting provided individual self-esteem incentives for access and
participation, felicitous levels of discussion and effective discourse
management.  Users knew they could acquire, share and be seen to have
information or ideas.  Any positive motivation which they as information
workers had toward these activities was strengthened by the salience of
the discussion domain and the opportunity to investigate a new
information technology.  "Reinventing government" was frequently
mentioned by the media at the time in the context of downsizing and
hence a possible threat to their employment, but fewer than half the
registered reported knowing much about it or even having seen an NPR
report.  The meeting also offered concerned workers possibly their only
opportunity to influence the implementation of the recommendations.
This incentive was strengthened for some workers by their belief the
Vice-President would personally review the discussions\footnote{In
addition to contributing a letter of welcome to the meeting, the
Vice-President Gore did post two questions.}.  More group oriented
motivations, like helping to form a group or speaking with one's peers,
may have motivated some users to revisit or track discussions.  In
informal, post-meeting evaluations, these users reported having a "we"
feeling when they discovered people from other organizations who were
contending with the same problems.

The focus on individual recommendations for common operating systems
placed the discussions at a level close to that of concrete operations,
without requiring users to know terminology or conditions of a
particular organization. It was consequently easy for users to draw on
their own experiences in reading and responding to the recommendations
and comments.  Because the focus also limited the attraction and
participation for any particular discussion, potential contributors
could have realistic expectations that their opinions would be noticed
and could influence the discussion.  This sense of being critical
encourage some users to expend the intellectual energy and time to
comment, rather than just lurking.  Note the meeting was arranged to
reduce other costs of participation by enabling attendance from their
desktops and at any time convenient for them during work hours -- with
the blessings of the Vice-President. Indeed the potential for the
Internet to decrease the transaction costs of participation is a major
reason to imagine it can help revive a public sphere.

The presentation format for the recommendations and the structure
provided by the discourse grammar were sufficient cues for participants
to recreate the collective policy or program reviews, that many were
practiced in their work.  This result could be expected, given their
background organizational cultures and bureaucratic roles, but was not
inevitable.  Since roles were under-specified, and the discourse moves
were general, the situation was open to interpretations.  One
participant, a line worker and union steward, submitted scores of
questions, agreements and disagreements, which emphasized the need for
the government as employer to improve working conditions.  He seemed to
be following a script or personal role of questioning authority and only
stopped, when a moderator suggested by private email communication that
he could get more attention by posting fewer comments.  The incident
highlights that discourse grammars do not constitute a universal
pragmatics, that is sets of moves (speech acts) and rules which are
understood similarly by all and which construct processes that are
transparent to everyone.

As implied above, the second source of success was the people who
attended the meeting. Most had training that made them ideal
participants:  They were information workers, socialized to a culture of
meetings, with a professional or managerial interest in what was being
said and pre-existing familiarity with how to say it.  Such attributes
can also be found among the information workers, who have some
experience of collaboration over networks.  But while their rapidly
increasing numbers make them politically significant in the United
States\cite{12} these workers hardly constitute the universe, and
newsgroups demonstrate that discussions of political and social issues
among networked information workers are no more rational and orderly
than those of the general population\cite{13}.  On-line public spheres
open to politically and culturally heterogeneous participants are
consequently likely to have problems of governability, beyond the
capacities of current types of controls.


\subsection{The Open Meeting as a Public Sphere}

How well did the Open Meeting realize the ideal of a public sphere?  How
good a model is it for an on-line democratic institution?  One answer is
that the participatory process through which a community becomes an
active agent on an issue requires members' access to relevant
information, their deliberation, decision, privacy, commitment to the
decision and time.  Privacy here refers to protection of participants
from intimidation, insult, defamation or similar means intended to chill
participation in public discussions.  Time refers to the examination of
a significant issue and the expression of all interested parties' views
on it taking a considerable amount of time -- weeks or months.  By these
standards, the Open Meeting does half the job; it successfully provides
a platform for accessing information and participating in orderly
discussions.

As implemented for the meeting, the major shortcomings of the system
appear in regard to decision and developing a community spirit that
would facilitate commitment by participants to a decision.  As noted,
there were reasons for not implementing a decision mechanism, and such a
mechanism, arguably, might not be needed, since argument grammars have
an implicit decision mechanism. They terminate when challenges to a
claim are exhausted, a challenge cannot be countered or the opposing
sides agree to disagree. These endings are more in keeping with a
consensus seeking process based on force of reason alone. Nevertheless,
few communities really have the time for exhaustive arguments and few
scenarios allow decisions to wait for them.  Realistic, effective
support of on-line processes in which communities clarify issues and
take stands has to include event-creating moves like cloture and voting
procedures.  These moves are event creating; they change the phase of a
deliberative process by ruling out some previously permitted moves and
activating other ones.  More generally, an effective Open Meeting system
will need a library of tested procedures, represented by discourse
grammars, with projected outcomes and identified risks, against which
conveners of community meetings or collaborations can measure their
needs.

As noted earlier, the strategy of decomposing large meetings into small
discussions has the potential to isolate users and block their sight of
the large community.  A low cost means of meeting this problem is an
interface display of the distribution of interests over the various
topics and proposals, as measured by user visits or comments.  Users
would be able to gauge where their own discussions fit in the
distribution of the community's attention and to directly access popular
or other spots, if they so wanted.  Slightly more expensive is the
automatic generation of cross-links, based on statistical information
retrieval methods and patterns of semantic links, that point users to
other discussions with themes or information similar to their own.  As
well as discovering cross cutting themes or issues to integrate the
knowledge produced in the discussions, this technique can also be used
to discover construct wider circles of participants with similar
problems.\footnote{The second idea follows from \cite{15}, an
implemented application of agent technology, using cluster analysis of
email to find individuals with similar interests.  Inversely, cluster
analysis of the distribution of participants over the discussions may
suggest cross links that are not apparent at the text level.}

By showing under certain condition, suitable mechanisms can produce a
successful, widely-attended discussion of issues, the Open Meeting was a
useful step toward building democratic practices in networked
communities.  However, because of the importance of shared
organizational/ cultural backgrounds for the success of these
discussions, it is more realistic to think about multiple public
spheres, each serving a different subcommunity.  Each group could
develop its own discourse grammar, and agents in each sphere would seek
cross-links to other groups by similarities in issues discussed.  While
that procedure does not create consensus, it might promote mutual
awareness.
  
\begin{thebibliography}{15}

\bibitem{1} Habermas, J.: The Structural Transformation of the Public
Sphere. M.I.T., Cambridge, MA (1989), first published as Strukturwandel
der Offentlichkeit. Luchterhand, Berlin (1962).

\bibitem{2} Knapp, J.: Essayistic Messages: Internet Newsgroups as an
Electronic Public Sphere. In: Porter, D. (ed.): Internet Culture.
Routledge. New York London (1997).

\bibitem{3} Hurwitz, R., Mallery, J.: The Open Meeting: A Web-Based
System for Conferencing and Collaboration. Proceedings of The Fourth
International Conference on The World Wide Web. Boston (1996) 19-36
http://www.w3j.com/1/ hurwitz.349/paper/349.html


\bibitem{4} http://www.npr.gov

\bibitem{5} http://www.whitehouse.gov/WH/publications/html/Publications.html

\bibitem{6} Searle, P.G., Searle, J.: Speech Acts. Cambridge U. Press,
Cambridge (1970).

\bibitem{8} Toulmin, S.: Uses of Argument. Cambridge U. Press, Cambridge (1958).

\bibitem{10} Rescher, N.: Dialectics: A Controversy-oriented Approach to the
Theory of Knowledge. State U. of New York Press, Albany (1978)
 
\bibitem{9} Eemeren, F. v., Grootendorst, R.: Speech Acts in Argumentative
Discussions. Foris Publishers, Dordrecht, Holland (1984).

\bibitem{7} Brown, P., Levinson, S.: Politeness : Some Universals in
Language Usage.  Casmbridge University Press, Cambridge (1987).

\bibitem{11} Grosz, B., Sidner, C.: Attention, Intentions and the Structure
of Discourse. Computational Linguistics 12 (1986) 175-204.

\bibitem{12} Winograd, M., Buffa, D.: Taking Control: Politics in the
Information Age. Holt, New York (1996).

\bibitem{13} Schneider, S.: Expanding the Public Sphere through
Computer-mediated Communication: Political Discussion about Abortion in
a Usenet Newsgroup. Ph.D. dissertation, Political Science Dept., M.I.T.
(1997)

\bibitem{14} Docter, S., Dutton, W.: The First Amendment Online: Santa
Monica's Public Electronic Network. In Tsagarousianou, R., Tambini, D.,
Bryan, C. (eds.): Cyberdeomocracy: Technology, Cities and Civic Networks
(1998).

\bibitem{15} Foner, L.: Yenta: A Multi-Agent, Referral Based Matchmaking
System. The First International Conference on Autonomous Agents Marina
del Rey, California (1997). 

\end{thebibliography}
          

\end{document}