Butterfly: A Conversation-Finding Agent for Internet Relay Chat Neil W. Van Dyke, Henry Lieberman, Pattie Maes Media Laboratory Massachusetts Institute of Technology 20 Ames Street #El 5-305 Cambridge, MA 02139, USA { nwv, lieber, pattie} @media.mit.edu ABSTRACT The Internet enables groups of people throughout the world to interact to discuss issues, get assistance, learn, and socialize. However, when there are thousands of loosely defined groups in which a user could potentially participate, the problem becomes finding the groups of most interest. In this paper we focus on the domain of Internet Relay Chat real-time text messaging, and describe a “social butterfly” agent called Butterfly that samples available conversational groups and recommends ones of interest. We discuss Butterfly’s motivation, usage, real- world design constraints, implementation, and results. Finally, we introduce work in progress on a multi-agent approach that has grown out of our experience with Butterfly. Keywords Agents, Internet, conversation, information filtering INTRODUCTION The Internet allows groups of people with similar interests to interact with each other, with little regard to geographic location. Popular group interaction media that are layered atop the Internet substrate include electronic mailing lists, Usenet newsgroups, and real-time text chat systems. We focus in this paper on Internet Relay Chat (IRC) [7], a major real-time textual group messaging system, although the general approach we describe is applicable to other Internet-based group media. Conversational groups on IRC are defined by channels (“chat rooms”), most of which have both regular and participants and drop-in visitors. Users explicitly join channels in which they wish to participate, and any message sent to a channel is seen by all users joined to it. Channels are created on demand by any IRC user, and each channel exists until the last participant has left. Permission to make digital or hard copies ofall or part oftllis work for personas or c~assruom use is granted without fee procidrd that copies arc mt n&e or distributed Ibr prolit Or commcrclal XlLalltage k3lld that copies bear this notice and the full citation on the tirst page. ‘f0 COPY otherwise. to republish, to post on servers or to redistribute to k3s, requires prior specific permiSsion an&or a Tee. IL’I 99 Redondo Beach CA liSA Copyright ACM 1999 l-581 13-098~8/99/01...$5.00 There are typically over ten thousand IRC channels at any given moment,’ each identified by a short string name that often does not give much indication as to its content. Each channel can optionally have a one-sentence topic string that states the intent of the channel, although in practice the topic string is often not used for this purpose. There is no hierarchy or organizing mechanism to aid a user in finding channels. Thus, a user interested in a certain topic is reduced to trying likely channel names and manually searching through a list of thousands of channels’ names and topic strings, guessing at the content of each. We propose augmenting the user’s interface to IRC with a software agent that deals with the information overload of ten thousand potential conversations. Our first such prototype agent, called Butterfly, samples the content of all of the channels and makes recommendations using a keyword-based model of interest. Since it is deployed in an existing real-world environment, Butterfly’s design has been influenced by privacy concerns and architectural properties of the IRC infrastructure. USAGE SCENARIO Butterfly’s usage is perhaps best illustrated by an example. First, the user sends Butterfly an IRC message containing keywords for some of the user’s interests. User: “I am interested in agents, communityware, ecommerce, sociology, and HCI.” Butterfly immediately returns a message recommending a channel that it remembers contained some of the keywords: Bu&QZy: “OK. You may be interested in channel #VirSoc (communityware, sociology).” ’ In a representative sample of the IRC network used for evaluation, EFnet, there were 38,367 connected users in 15,951 channels, distributed among 56 servers. The actual base of users is much larger, as most do not remain connected continuously. 39Butterfly also remembers the keywords for the user’s interests, and later when it discovers one of them being used in a channel, it notifies the user: Butterfly: “You may be interested in channel #hahvahd (ecommerce).” The pseudo-natural language interface is described in more detail later in this paper. INTEREST AND CONTENT MODELS The user interest model presently used by Butterfly is based on a simple term vector with positive and negative weights. For an example of differing weights in the user interest profile: User: “I am interested in sprinting.” User: “I’m very interested in Aikido, but not in Seagal.” User: “Who am I?” Butterfly: “The interest profile I have for you is: aikido(200) seagal(-50), sprinting( 100)” The current version of Butterfly uses fixed constants for normal interest, great interest, and negative interest weights, although more variable weights could be supported if the interest profile were learned, such as by watching the user’s Web browsing behavior [5]. Channel content is also represented as a term vector, with the weights corresponding to frequency of occurrence. Relevance of a channel to an interest profile is determined by summing the weights of the dot product of the two vectors. If the sum is over some threshold, then the channel is recommended. The incremental building and decay of the channel content term vectors is tied to the channel-sampling behavior. CHANNEL SAMPLING Our decision to have