| Popis: |
present in the collection itself. In the following explanation, we use this last option. Using a standard statistical correlation technique, we begin by ordering the key concepts. Then for each document, we determine which concept from the ordered set is most strongly represented in that document, which is second most strongly represented, and so on, up to an arbitrary limit of 32 total concepts. In this case, this step is achieved by examining the n-dimensional document vectors. A three-dimensional visualization is created using the ordered set of concepts for each axis. The x-value for a document is the number of the concept that is most strongly represented in that document; the yand z-values are the numbers of the second and third most strongly represented concepts. Similar visualizations can be shown for any trio of concept strengths desired. Our system provides rotation to help the user examine the visualization; even in the static image, certain patterns are clear (see Figure 4 at the end of this document). Concentrations of dots show where particular closely related concepts show up strongly in multiple documents. Selections can be made within each “cell” in a particular hypercube, and the documents can be reprojected using any other topic strength combinations — such as the fourth, fifth, and sixth. This recursive interaction is suggestive of World Within Worlds [3]. 4 NEED FOR NEW METRICS With several alternate visualization approaches, the question of which one is better for what set of user needs and under what circumstances naturally arises. Each visualization method is designed to emphasize parts of the discovery process; however, we still need some way to measure the actual results. It is likely that interacting with more than one visual paradigm/method offers great promise. Metrics for query-based information retrieval are well established. The TIPSTER Text Program has ushered in truly major advances in document detection and information extraction through the Text Retrieval Conferences (TREC) and Message Understanding Conferences (MUC), respectively [URL: http://www.tipster.org/]. The querying and ranking methods of document detection can be characterized by the right answers to user-formulated retrievals. The metric of “recall” is used to measure what percentage of a predetermined set of “right answers” in the collection a system actually finds. The metric of “precision” is used to measure what percentage of answers returned by the system is actually correct. Since the start of the annual TREC competitions in November 1992, precision and recall performance has increased to focus the competitions on the fourth digit of accuracy. MUC has facilitated the growth in natural language processing technique to allow facile extraction of both proper and organization names and some scenario assessment in domainspecific areas. Our systems, though useful to our clients, would likely not shine in a TREC-type competition because the metrics do not actually measure what our system does best: make connections that are not known a priori. Note also that the measures of precision and recall do not require graphics. Our work depends heavily on graphic presentation for the communication of content. Bertin distinguished several distinct functions of graphics; among them are the capabilities to display information already understood and to facilitate the process of understanding of the information [2]. Guidelines for the presentation of static information have been used in the statistical community for years and have been incorporated into many statistical exploratory data analysis packages [6]. Perhaps best known are the principles of visual presentation as outlined by Tufte in his various books over the years [12 14]. Brath's recent work proposes four metrics to assess the efficacy of static 3-D presentations: number of data points and data density; number of dimensions and cognitive overhead; percentage of occlusion; and reference context and percentage of identifiable points [4]. The advantage of these measures is that they are objective and fairly easy to measure. The disadvantage is that they are for static pictures and thus have not been extended to interactive models. In the interactive use of graphics to gain insight into information, most researchers begin by a definition of data types. This area is more subjective and immature than the “static” metrics research but is very relevant to our work. Schneiderman has identified seven data types and seven tasks: overview, zoom, filter, details on demand, history, and extract [10]. These tasks fit very closely to our approach to information exploration. Other data types have been defined by others, such as Zhou [16] and Card [5]. Card proposes tables that can be used to track the processing for each variable used in the visualization and the interaction available for each. Data types and tables are important because they facilitate comparisons across very different visualization systems and can be used to succinctly identify similarities and differences. An interesting area of further research would be to investigate whether we can (1) predict successful visualizations based on objective quantities that can be easily measured, and then (2) design the experiments to take the measurements and verify the predictions. In our current experiment, the first informal effort was to compare our alternate visualizations for the 12,000-document corpus by importing the coordinates of each document into a common analysis package. Each document is represented in the figures below as an individual point. The points are colored according to an independent clustering so that the point representing a document is the same color in each visualization. The figures show that each display broadly classified documents similarly; the blues are separated from the reds and purples in all cases. We pose a question to the research community: How can we measure the “goodness” of a particular or combined visualization? The recently added (1995) Interactive Task in TREC somewhat more closely measures what it is we try to do in visual information analysis. The Interactive Task is used to study both user interaction with text retrieval systems and how quickly a variety of “aspects” can be discovered about specific topics. Using a similar approach, some ideas for making this assessment might include • contests to characterize main themes or topics addressed in a test corpus • contests to identify hidden connections that can only be made by integrating information across documents • variations on software usability tests. Figure 3: Cosmic Tumbleweed Projection Figure 4: 32D Hypercube Figure 5: Spire Galaxy Projection Figure 6: Fractal Projection |
