<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-0208">
  <Title>ACL-COLING Workshop on Content Visualization and Intermedia Representations</Title>
  <Section position="1" start_page="0" end_page="0" type="metho">
    <SectionTitle>
ACL-COLING Workshop on Content Visualization and Intermedia Representations
Abstract
</SectionTitle>
    <Paragraph position="0"> This paper summarizes several initiatives at MITRE that are investigating the visualization of a range of content. We present results of our work in relevancy visualization, news visualization, world events visualization and sensor/battlefield visualization to enhance user interaction in information access and exploitation tasks. We summarize several initiatives we are currently pursuing and enumerate unsolved  problems.</Paragraph>
    <Paragraph position="1"> 1. Visualizing Semantic Content  Visualization can support effective and efficient interaction with a range of information for a variety of tasks. As Figure 1 illustrates, information (data elements, attributes, relations, events) can be encoded in (possibly interactive) visual displays which users can exploit for a variety of cognitive tasks such as retrieval, analysis (e.g., of trends, anomalies, relations), summarization, and inference. In this paper we consider a range of semantic content, visual mechanisms, and cognitive tasks to deepen our understanding of the role of interactive</Paragraph>
  </Section>
  <Section position="2" start_page="0" end_page="53" type="metho">
    <SectionTitle>
2. Document Relevancy Visualization
</SectionTitle>
    <Paragraph position="0"> Today's users are faced with a dizzying array of information sources. MITRE's Forager for Information on the SuperHighway (FISH) (Smotroff, Hirschrnan, and Bayer 1995) was developed to enable the rapid evaluation of information sources and servers. Figure 2a illustrates the application of FISH to three Wide Area Information Server (WAIS) TM databases containing information on joint ventures from the</Paragraph>
    <Section position="1" start_page="0" end_page="53" type="sub_section">
      <SectionTitle>
Message Understanding Conference (MUC).
</SectionTitle>
      <Paragraph position="0"> Figure 2b illustrates the application of FISH to visualize e-mail clustered by topic type for a moderator supporting a National Performance Review electronic town hall.</Paragraph>
      <Paragraph position="1">  Figure 2a. WAIS FISH Figure 2b. NPR FISH The traditional WAIS interface of a query box and a list of resulting hits is replaced by an interface which includes a query box, a historical list of queries, and a graphically encoded display of resulting hits (an example of which is shown in Figure 2a). In WAIS, the relevancy of a document to a given keyword query is measured on a scale from 1-1000 (where 1000 is the highest relevancy) by the frequency and location of (stems of) query keywords in documents.</Paragraph>
      <Paragraph position="2"> Motivated by the University of Maryland's TreeMap research for hierarchical information visualization, FISH encodes the relevance of each document to a given query (or set of compound queries) using both color saturation and size.</Paragraph>
      <Paragraph position="3"> In the example presented in Figure 2a, each database is allocated screen size in proportion to the number of and degree with which documents are relevant to the given query. For example, the MEAD database on the left of the output window is given more space than the PROMT database in the middle because it has many more relevant documents. Similarly, individual documents that have higher relevancy measures for a given query are given proportionally more space and a higher color saturation. In this manner, a user can rapidly scan several large lists of documents to find relevant ones by focusing on those with higher color saturation and more space.</Paragraph>
      <Paragraph position="4"> Compound queries can be formulated via the &amp;quot;Document Restrictions&amp;quot; menu by selecting the union or intersection of previous queries, in effect an AND or OR Boolean operator across queries.</Paragraph>
      <Paragraph position="5"> In Figure 2a, the user has selected the union of documents relevant to the query &amp;quot;japan&amp;quot; and the query &amp;quot;automobile&amp;quot;, which will return all documents which contain the keywords &amp;quot;japan&amp;quot; or &amp;quot;automobile&amp;quot;. Color coding can be varied on these documents, for example, to keep their color saturation distinct (e.g., blue vs. red) to enable rapid contrast of hits across queries within databases (e.g., hits on Japan vs. hits on automobile) or to mix their saturation so that intersecting keyword hits can be visualized (e.g., bright blue-reds could indicate highly relevant Japanese automobile documents, dark the opposite). In the example in Figure 2a, blue encodes Japan, red Automobile; the color coding is set for mixed saturation, the union of the relevant document sets for those two keywords is selected, and the order (from top to bottom in the display) is used to encode the WAIS relevancy ranking. One issue is just how effectively users can discriminate mixed colors.</Paragraph>
      <Paragraph position="6"> fGA.O study hnk$ (hemi&lt;al Oou~le clkk to edit (ol~s: ~rag (o remange</Paragraph>
      <Paragraph position="8"> More recently, we have explored multiple server evaluation on popular World Wide Web search engines. For example, Figure 2c illustrates a query across multiple servers. Research issues include differences in relevancy ranking algorithms, encoding of multiple attributes beyond relevancy using color or size (e.g., length,  quality, cost, source), and document collections which are heterogeneous in size, content, and format.</Paragraph>
    </Section>
  </Section>
  <Section position="3" start_page="53" end_page="55" type="metho">
    <SectionTitle>
3. Document Structure/Content Visualization
</SectionTitle>
    <Paragraph position="0"> Figure 3a (Gershon et al. 1995; Gershon 1996) illustrates another navigation mechanism in which the user is able to view a hierarchy of the browse space. The left: hand of Figure 3a displays the traditional HTML layout of a web page whereas the right hand side illustrates a hierarchical, navigable view automatically generated from the underlying structure of the browsing space. The user can create a personal space by interactively and visually modify the structure of hyperspace or extracting segments of the documents.</Paragraph>
    <Paragraph position="1">  For discovery and analysis of new information and relationships in retrieved documents, we have developed a method for aggregating relevant information and representing it visually (Gershon, et al, 1995). The method is based on representing correlations of words within a document in a table. These tables could be very large depending on the size of the document thus making it difficult for the user to perceive and make sense of all the highly relevant correlations. Since the order of the words is not usually based on contents, the order of the words is permuted until the highly relevant correlations are concentrated in one comer.</Paragraph>
    <Paragraph position="2">  Other research at MITRE has focused on automatic discovery and visualization of semantic relations among individual and groups of documents (Mani and Bloedom 1997). Figure 3c illustrates the results of visualization of a set of documents using the NetMap visualization software after clustering these into related groups which appear around a circle. Outside of each cluster on the circle are displayed intracluster relations; in the center of the circle are intercluster relations (e.g., a shared named entity such as a person, place, or thing which appears in multiple documents). The user can zoom in any part of the graph. This is shown in Figure 3d, which shows individual people (green) and organizations (aquamarine).</Paragraph>
    <Paragraph position="3"> Selecting an individual entity from a document returns a display such as that in Figure 3e. Figure 3e illustrates individual entities encoded with color and shapes (e.g., people in green stick figures, organizations in aquamarine diamonds, locations in purple jagged rectangles, documents in yellow circles, person-organization relations in white squares). Lines and their properties (e.g., color, dashed) can encode relations among these entities (e.g., co-occurrence in documents). This provides a richer mechanism for discovering  interdocument and interentity relationships during analysis. Current research is investigating the role of automated text summarization, document retrieval and navigation and visualization.</Paragraph>
  </Section>
  <Section position="4" start_page="55" end_page="56" type="metho">
    <SectionTitle>
4. Named Entity/News Visualization
</SectionTitle>
    <Paragraph position="0"> MITRE's Broadcast News Navigator (BNN) is a system that is investigating analysis of trends in news reporting. BNN performs multistream (audio, video, text) analysis to eliminate commercials, segment stories, extract named entities (i.e., people, organization, location) and keyframes, and classify and summarize stories (Merlino, Morey, and Maybury 1997). BNN's intuitive web-based interface gives the user the ability to browse, query, extract from and customize digitized broadcasts. Figure 4 illustrates a trend analysis display from BNN that shows the most frequently mentioned named entities reported on CNN Prime News TM from October to November of 1997. &amp;quot;China&amp;quot; spikes in the center of the graph, associated with a state visit to Washington. Later &amp;quot;Iraq&amp;quot; spikes which is correlated with news regarding UN site inspections. The user can click on any point on the line graphs and be brought to a list of stories that mention that named entity.</Paragraph>
    <Paragraph position="1">  In contrast, the user can formulate a query specifying keywords, named entities or subjects. Figure 5a shows the results of executing the  query: Find me stories which have a topic of 1 Note in the display the occurrence of the terms &amp;quot;U.S.&amp;quot; and &amp;quot;United States&amp;quot;. BNN performs no co-reference resolution, a topic of current research at MITRE.</Paragraph>
    <Paragraph position="2"> &amp;quot;chemicals&amp;quot;, the keywords &amp;quot;chemical weapons&amp;quot;, person &amp;quot;Sadam Hussein&amp;quot;, organization &amp;quot;Pentagon&amp;quot;, and location &amp;quot;Iraq&amp;quot;. Each story in this &amp;quot;Story Skim&amp;quot; view is represented by a keyframe and the three most frequent named entities. Selecting one of these stories yields a &amp;quot;Story Detail&amp;quot; display, which as shown in Figure 5b including a keyframe, named entities, subject classification and pointers to the closed caption and video source.</Paragraph>
    <Paragraph position="3">  intemet stones, over time, and spoken language stories. Other investigations are focusing on which presentation mixes (e.g., keyffames, named entities, one line summary, full video source) are most effective for story retrieval and fact extraction from news (Merlino and Maybury 1998).</Paragraph>
  </Section>
  <Section position="5" start_page="56" end_page="57" type="metho">
    <SectionTitle>
5. Geographic Event Visualization
</SectionTitle>
    <Paragraph position="0"> The Geospatial News on Demand Environment (GeoNODE) initiative at MITRE is a new project investigating visualizing geographic aspects of news events. This program builds on MITRE's BNN, described in the previous section, and MSIIA, addressed in the subsequent section.</Paragraph>
    <Paragraph position="1"> GeoNODE is based on the research area of Geographic Visualization which investigates methods and tools that impact the way scientists and others conceptualize and explore georeferenced data, make decisions critical to society, and learn about the world (MacEachren and Ganter 1990, Taylor 1991). Since news reports are about events in the world, the reported events and trends can be assessed, queried, and reviewed effectively by leveraging a person's preexisting knowledge of the world's geography.</Paragraph>
    <Paragraph position="2"> The objective of GeoNODE is to understand the information integration of geospatial/temporal visualizations, information retrieval, multimedia, and other technologies to support browsing, analysis, and rapid inference from broadcast news.</Paragraph>
    <Paragraph position="3"> As shown in Figure 6, GoeNODE will analyze global and local cooperation and conflict found in broadcast news, internet, newswire and radio sources as well as broadcast news. Processing will include the identification, extraction, and summarization of events from national and international sources. GeoNODE will consider event types (e.g., terrorist acts, narcotrafficking, peace accords), frequency, and severity in an interactive geo-spatial/temporal context that supports browsing, retrieval, analysis and inference.</Paragraph>
    <Paragraph position="4">  Although a geographical context can enhance a person's understanding of reported events and therefore facilitate news retrieval and further queries, the same familiar visualization concerns apply to geographic presentation that are salient in visualizing any data rich multivariate information space. The GeoNODE user experience is derived from research, experience and standard practice in the visual search and retrieval domains: Overview first, zoom and filter, then details-on-demand (Shneiderman 1994). During each stage of the visualization process, cartographic methods and spatial analysis techniques are applied. These can be considered as a kind of grammar that allows for the optimal design, production and use of maps, depending on the application (Kraat 1997). Select cartographic generalization operators are applied to address key multi-scale and information overload problems (Buttenfield 1991).</Paragraph>
    <Paragraph position="5"> GeoNODE addresses Knowledge Representation (KR) and information fusion issues that are important to the news event presentation. The KR activities specific to GeoNODE are concerned with discovering and manipulating geospatial and temporal information, specifically investigating the following: improved natural language processing of place names that are central to understanding a news report  Spatial information management is currently growing in its utility to commercial applications, and several industries have already begun to explicitly rely on GIS systems, although most (53%) companies are evaluating while an average of only 7% are implementing or using a GIS (IDC 1997). Accompanying the growing interest in spatial information is a technology trend influencing the architecture of GeoNODE, mainly, a shift from single-purpose/standalone GIS applications to geospatial extensions and services for databases, component frameworks, data warehouses and data analysis applications. By supporting a component-based architecture, GeoNODE can more readily take advantage of future geospatial services and an expanding number of news sources (internet, newswire, radio, and other broadcast sources).</Paragraph>
    <Paragraph position="6"> Further research will investigate incorporation of summarization, geospatial/temporal KR, and other traditional visualization techniques. For example, Figure 7 illustrates some of the kinds of visualizations that are being explored by other researchers, such as the use of color and geolocation to encode relations among geographic entries. Figure 7 is a geographic visualization of early WWW usage available at http ://www.cybergeography.org/atlas/atlas.html. These and other research threads will shape GeoNODE into a visualization component for reasoning about news events in geographic space. As a long term objective, the system architecture should allow for navigation and retrieval from topic, conceptual, and web spaces where a user can access, update and annotate existing data with spatial information.</Paragraph>
  </Section>
class="xml-element"></Paper>