Revealing Collection Structure through Information Access Interfaces

In fo rmat ion Access research at Xerox PARC focuses on ampl i f y ing the users' cognitive abil i t ies, rather than t r y i ng to completely automate them. This framework emphasizes the par t ic ipat ion of the user in a cycle of query fo rmu la t ion , presentation of results, followed by query re formulat ion, and so on. Th is framework is intended to help the user i teratively refine a vaguely understood in format ion need. Since the focus is on query repair, the in fo rmat ion presented is typical ly not document descriptions, but rather intermediate in format ion that indicates relationships between the query and the retrieved documents. We have developed in format ion access tools intended to supply some of this funct ional i ty , and describe two of these here. As an i l lus t ra t ion , suppose a user is interested in medical diagnosis software. Assume that in i t ia l ly the user has available a large, unfamil iar in format ion source. In our example, this source is the 2.2 Gigabyte T I P S T E R text col lection [Harman, 1993]. Because the collection is unfami l ia r , the user wi l l be unsure whether it contains relevant in fo rmat ion , and if so, how to access i t . To address this s i tuat ion, we have developed a browsing me thod , called Scatter/Gather [Cut t ing et a/., 1992; 1993], tha t allows a user to rapidly assess the general contents of a very large collection by scanning through a dynamic , hierarchical representation that is mot ivated by a table-of-contents metaphor. In i t ia l ly the system automat ica l ly scatters, or clusters, the collection into a smal l number of document groups, and presents short summaries of the groups to the user. These summaries consist of two types of in format ion: topical t i t les (t i t les of documents close to the cluster centroid) and typical terms ( terms of importance in the cluster). Based on these summaries, the user selects one or more of the groups for fur ther study. The selected groups are gathered) or unioned, together to form a subcollection. The system then applies clustering again to scatter the new subcol lect ion in to a smal l number of document groups, which are again presented to the user. W i t h each successive i te ra t ion the groups become smaller, and therefore more detai led. The user may, at any t ime, switch to a more focused search method. Figure 1 shows a port ion Figure 1: A port ion of a top-level view of the Scatter /Gather algori thm over the T I P S T E R corpus.