Involving Intelligent Assistants in Active Human CommunicationDonald J. PattersonUniversity of California IrvineAbstractIntelligent assistants that support human communication needto respect the difficulty of understanding the context sur-rounding the interchange. Rather than attempting to directlycommunicate for a user, intelligent assistants should supportdecision making on the part of the involved parties so thatcomplex social negotiations are preserved. We describe anintelligent assistant that does this for instant messaging calledNomatic*Gaim.IntroductionThere are many examples of learning algorithms that sup-port communication exchanges between individuals. Spamfilters, grammar checkers and fundamental TCP packet rout-ing algorithms are some examples. There are few examples,however, of AI algorithms or intelligent assistants that usersrely on during real-time communication. An excellent ex-ample of an application that would benefit from such supportis instant messaging (IM).IM was originally architected in a world in which userswent online in at most two places, home and work. With thearrival of IM on cell phones, the typical IM-enabled distinc-tions of being online/offline no longer give enough informa-tion for communication partners to understand whether ornot it is appropriate to initiate a communication.Within the old paradigm of desktop computers, even inthe presence of online status indicators, availability nego-tiation consumed 13% of IM communications (Handel &Herbsleb 2002). As users are increasingly mobile and in-creasingly “always online” it is reasonable to expect that thispercentage will increase.As a result it is necessary to give users assistance in ne-gotiating communication availability. A classic AI solu-tion would be to craft an algorithm that would automaticallylearn and then set the availability status of a user. We claim,however, that this is the wrong approach. Deciding whetheror not one is available is a complex social negotiation thatis best handled by people. This paper proposes using an in-telligent assistant to communicate context information in anIM side-channel on behalf of a user to reduce the need toengage in discussions about availability.Copyrightc 2007, American Association for Artificial Intelli-gence (www.aaai.org). All rights reserved.Palen & Dourish frame the problem of determining avail-ability as one of boundary negotiation (Palen & Dourish2003). It is a constant subconscious process that people doon a regular basis that requires complex calculations that areunlikely to be successfully done by an intelligent assistantanytime soon. In contrast people do this without a second-thought. The problem can be exemplified by the questionof whether one should interrupt someone with a phone callduring a movie. In most cases the right answer would seemto be “no,” unless the call is for a doctor, unless it’s abouta billing question not an emergency, unless the doctor isn’tactually watching a movie, but is just picking up her son,etc. Making a good decision is easy as an informed person,but crafting an effective algorithm for the same task seemsdaunting.Palen & Dourish might argue that one of the reasons whyit is difficult for intelligent agents to participate in automaticcommunication on behalf of users is because it puts them inthe position of negotiating boundaries: a place where mean-ing and nuance dominate. In such a situation, current gener-ations of assistive agents have no place. Instead they shouldbe supporting the decision making process so that users cannegotiate the boundary more efficiently.A better approach is to provide the users with informationabout the context of the parties so that that information canbe used as part of the boundary negotiation. Knowing thatthe doctor is at the movie theater might be all that a callerneeds to know in order to make a good decision about inter-rupting her. By giving the responsibility of reporting contextto an intelligent assistant, the human boundary managementnegotiation is supported but not replaced. In the case of IM,we propose Nomatic*Gaim as an example of such a solu-tion.Nomatic*GaimNomatic*Gaim is an open-source, multi-operating system,multi-protocol IM client that is based on the open-sourceproject gaim. It has been enhanced to support more complexpresence indicators than just online/offline. Instead it takesan unusual approach by revealing the current location of theuser on the presence/status line.The input to Nomatic*Gaim is a set of Wi-Fi accesspoints (APs) that can be seen from the user’s current com-puting platform. The APs need not be “open” in thesense that they provide Internet connectivity without a pass-word, they simply need to be broadcasting their existence,as most do, for Nomatic*Gaim to use them. After find-ing a collection of APs that are visible in the environ-ment, Nomatic*Gaim looks up their position in a databaseof known APs and infers the user’s current location usingbeacon positioning algorithms (e.g., (LaMarca et al. 2005;Letchner, Fox, & LaMarca 2005)). From this data, the user’sapproximate latitude and longitude can be obtained.It would be possible to simply list the latitude and longi-tude on the presence status indicator of an IM client. How-ever, although this is a great deal of information, it is notvery useful in determining whether or not it is appropriate tointerrupt a user for an IM chat. Enter the intelligent assistant.Instead, the SSIDs of the APs, denoted as a, the currentday and time, t, and the latitude and longitude, x, are usedas observed features in a machine learning algorithm thatuses a trained model, M, to obtain a guess about the opti-mal semantic label, L∗, to use to describe a user’s currentplace (versus position, see (Hightower 2003)). This placebecomes valuable information that can then be used by bud-dies on a user’s IM buddy list to decide whether or not toinitiate a communication with an individual. Rather than de-ciding availability we instead support the IM users in mak-ing the decision themselves. We use the formulation belowto account for positioning error in the neighborhood aroundthe true position, x0:L∗= argmaxL∈P lacesZP (L|a, t, x0, M)P (x0|x)dx0(1)The training data comes from two places. First the bestdata is going to come from users themselves. We hypoth-esize that the amount of time currently spent negotiatingavailability in IM is enough incentive for a user to enterplace information into