SOaP: Social Agents Providing People -with Useful Information Angi Voss, Thomas Kreifelts GMD-FIT Institute for Applied Information Technology Schlol3 Birlinghoven D-53754 Sankt Augustin, Germany (angi.voss, thomaskreifelts )@gmd.de ABSTRACT We describe how a system of communicating software agents can help people in finding useful information. The agents operate on the Web because it constitutes an almost universal source of information. Search engines on the Web provide information with high recall and tolerable precision, but they can only be as good as the queries asked. Agents can exploit and supplement search engines by pro- viding information without being asked precise questions or without explicit questions at all. Agents can do that with little overhead for the users by exploiting bookmark collec- tions as sources of self-descriptions on the Web, by record- ing queries, results, and assessments, and by sharing all this among anonymous users as well as groups of users known to each other. Keywords Software agents, recommender systems, collaborative fil- tering, multi-agent systems, World Wide Web. THE WEB AS A RESOURCE The Web, with its increasing stock of up-to-date informa- tion, is becoming the premium source of information for a growing number of people. Putting one’s queries to the Web is becoming a preferred procedure over browsing an encyclopaedia, looking up something in a book, conducting a retrieval in a library, .asking a colleague, or performing even more expensive researches. Thus, any data that reflect a person’s queries to the Web, and his or her assessment of the results returned, is a useful mirror image of that person’s interests and knowledge. It is our goal to exploit this kind of data in order to provide people with useful information. What information can we supply that search engines cannot? Today’s search engines are quite exhaustive, and we Pem&ssion to make digitalil~nrd copies of ail or part of this material for per;onnl or clmsroom use is granted without fee provided that die copies are not mode or distributed for profit or comniercird advantage, Ule copy- ri&t notice, the title of tbe publication and its date oppmr, sod UO~ is given Ulnt copyrigbt is by penaission of tbe ACM, Inc. TO copy otherwise. to republish, to post on servers or to redistribute to list.% requires specific pemiission and/or fee. GRollP 97 Phoenix Arizom USA Copyrigbt 1997 ACM O-89791-897-5/97/1 I..s3.50 expect that they will keep up with the growing Web. Pub- licly available search engines are not as precise as one would expect, but they do filter the information to some degree of precision. Precision could be improved with tech- niques from information retrieval, but this will not be our focus. If the output of a search engine is too copious, the Web user nowadays has to reformulate the query. This may actually be a problem because one does not yet know what to look for exactly. Search on the Web is often exploratory. Increasing precision would constrain the chance of finding some nuggets. Another problem with search engines is the unknown quality of the& results. There is no editor controlling the quality of the material published on the Web. Although there are lots of users who inspect the available material and form their opinion of what they see, search engines do not receive and process this feedback. An established strategy among humans relies on word of mouth. If you are looking for help concerning a particular problem, or if you are exploring a new area and need a suitable introduction, it is a good and established strategy to ask your colleagues or .other experts you know for their recommendations. If you trust them, and if they understand your problem or are aware of your background, then you have a good chance of obtaining relevant and high quality information. In order to mimic this procedure, we propose to take into account user feedback. This may include simple signs of approval or disapproval, as well as explicit assessments in terms of ratings and annotations. In doing so we are well aware that assessments may be of limited value if their context is not known. In our case, the context consists of the persons who made the judgement, their specific goals and tasks, and other circumstances which cannot be formally captured. In order to cope with missing context, we support the for- mation of groups of users, assuming that context is sufficiently focused within such groups. By providing infor- mation originating from selected groups, we also hope to improve trust in quality. 291To summarize, we intend to use the Web as a substrate where people express the topics which they are interested in and the information they consider as relevant, and to sup- port the sharing of this information between users and especially among groups of users. Before we explain how we capture and process this information, and what support functionality we can offer, we will motivate our use of soft- ware agents. SOCIAL AGENTS AS WORKHORSES The acceptance of a computer system crucially depends on the overhead it introduces. An ideal system would not impose any additional overhead on its users. A tolerable system rewards overhead with sufficient benefits. We strive for a system which can be used without overhead, and where users can increase their investment if they feel it worth while. To realize such a system, we use software agents, and more precisely social agents. We use the term “agents” for autonomous software pro- cesses that interact according to certain conventions. Agents may be solitary, as for instance personal information filtering agents, or they may be social and interact with each other on behalf of their clients. Agents for electronic markets, workflow agents, or information brokers a& examples of social agents. Agents have a number of properties which nicely supple- ment the qualities of humans in the process of