<?xml version="1.0" standalone="yes"?>
<Paper uid="N06-4009">
  <Title>SconeEdit: A Text-guided Domain Knowledge Editor</Title>
  <Section position="3" start_page="284" end_page="284" type="metho">
    <SectionTitle>
2 The Knowledge Base
</SectionTitle>
    <Paragraph position="0"> SconeEdit is a software client to the Scone Knowledge Base System, or simply &amp;quot;Scone&amp;quot; (Fahlman, 2005). Scone is an efficient, open-source knowledge base (KB) system being developed in the Language Technologies Institute of Carnegie Mellon University. Scone is intended to be a practical KB system that can be used as a component in a wide range of AI and natural language software applications. One of the goals in developing Scone is to make it easy to use, especially when adding new knowledge.</Paragraph>
    <Paragraph position="1"> The SconeEdit interface makes Scone more usable in several ways: the Text View display gives the user a convenient and intuitive starting point for exploring the knowledge base. SconeEdit also provides an easy way of adding knowledge to the KB without learning the formal input language for Scone. This demonstration focuses on the effectiveness of SconeEdit and Scone together, but the design principles of SconeEdit are applicable to knowledge bases written in other formalisms.</Paragraph>
    <Paragraph position="2"> Figure 1 shows the SconeEdit window with a document and KB loaded. The left side of the interface contains the Text View, and the KB View is on the right. Each of these views is described in detail below.</Paragraph>
  </Section>
  <Section position="4" start_page="284" end_page="285" type="metho">
    <SectionTitle>
3 Architecture
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="284" end_page="284" type="sub_section">
      <SectionTitle>
3.1 Text View
</SectionTitle>
      <Paragraph position="0"> In a traditional ontology browser, the user starts looking for concepts of interest by typing words and phrases into a search field. This is the model for several existing tools, including the VisDic viewer for WordNet (Horak and Smrz, 2004), the INOH ontology viewer (INOH, 2004), and the Gene Ontology viewer presented by Koike and Takagi (2004), among others.</Paragraph>
      <Paragraph position="1"> SconeEdit improves on this browsing paradigm by giving a user who is unfamiliar with the knowledge base an easy way to start exploring. Rather than generating a series of guesses at what may be Figure 2. Excerpt from Text View, with Search and Text Tabs covered by the KB, the user can load natural language text into SconeEdit from a file or the system clipboard. We take an article from Xinhuanet News Service (Xinhuanet, 2006) as an example.</Paragraph>
      <Paragraph position="2"> Figure 2 shows an excerpt of this text after it has been loaded.</Paragraph>
      <Paragraph position="3"> When the text file is loaded, it appears in the Text Tab of the Text View pane. SconeEdit highlights all strings that it can identify as concepts from the knowledge base. In this example, &amp;quot;Washington&amp;quot; is correctly identified as the city, not the state. In many cases the concept may be ambiguous from the string alone. SconeEdit currently uses dynamic programming to highlight the longest-matching concept names it can find (see Section 5). More sophisticated disambiguation is a priority for our future work.</Paragraph>
      <Paragraph position="4"> The result of highlighting is a concise visual representation of what is &amp;quot;known&amp;quot; about that text. The Text View helps a user find relevant knowledge quickly, even in a large general-domain KB. Clicking on any highlighted term in the Text View brings up a hierarchical representation of that concept in the KB View.</Paragraph>
    </Section>
    <Section position="2" start_page="284" end_page="285" type="sub_section">
      <SectionTitle>
3.2 KB View
</SectionTitle>
      <Paragraph position="0"> The KB View contains two tabs: a Graph Tab and a List Tab. The Graph Tab displays an excerpt from the knowledge base as a network of linked concepts with one focus concept in the center.</Paragraph>
      <Paragraph position="1"> When the user clicks on a highlighted concept in the Text View, a graph focused on that concept appears in the Graph Tab. Continuing with our Xinhuanet example, Figure 3 shows the Graph Tab after a user has clicked on &amp;quot;Washington&amp;quot; in the text. The Graph View now displays concepts that are closely related to Washington-Dc in the knowledge base.</Paragraph>
      <Paragraph position="2">  Clicking on any of these related concepts in the Graph Tab moves the focus of the graph to that concept.</Paragraph>
      <Paragraph position="3"> The List Tab shows an alternative view of the same focus concept. It displays KB information as a set of property lists. As in the Graph Tab, the user can double-click on any concept in the List Tab to bring that concept into focus. When the focus concept is densely connected to other concepts in the KB, the List Tab can be easier to interpret than the Graph Tab. In general, research has shown that preference for the list style or graph style is personal and varies from user to user (Tribble and Rose, 2006). Figure 4 shows the List Tab, focused on the concept City.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="285" end_page="285" type="metho">
    <SectionTitle>
4 Adding Knowledge
</SectionTitle>
    <Paragraph position="0"> Browsing the knowledge base in this way gives the user a detailed, domain-targeted view of its contents. A natural extension of this paradigm is to allow the user to edit the KB while browsing. For example, a user may encounter a concept in the Figure 5. Adding a concept synonym text that is not present in the knowledge base.</Paragraph>
    <Paragraph position="1"> SconeEdit allows the user to simply click on a word in the text to create a new concept in the KB (see Figure 5). To specify where the new concept belongs, the user navigates to the appropriate loca-tion in the KB View (List Tab or Graph Tab).</Paragraph>
    <Paragraph position="2"> The user can also modify an existing KB concept by adding English synonyms. For example, the word &amp;quot;United States&amp;quot; may be highlighted in a text example, while &amp;quot;U.S.&amp;quot; is not. To add a synonym for the &amp;quot;United States&amp;quot; concept, the user navigates to this concept in the KB View, and then clicks on the text &amp;quot;U.S.&amp;quot;. A menu offers the choice of adding a synonym to the existing focus concept.</Paragraph>
    <Paragraph position="3"> Figure 5 illustrates this process.</Paragraph>
  </Section>
  <Section position="6" start_page="285" end_page="286" type="metho">
    <SectionTitle>
5 Identifying KB Concepts in Text
</SectionTitle>
    <Paragraph position="0"> Elements in a Scone knowledge base represent specific concepts, rather than words or word senses. Each concept is linked with a list of English names (words or phrases). This association between Scone elements and English names is many-to-many.</Paragraph>
    <Paragraph position="1"> To map a sentence to the set of concepts that appear there, a dynamic-programming alignment is performed using the English names in the KB as a dictionary. SconeEdit searches for an alignment that covers as much of the input text as possible. The result of aligning an input string with concepts is a set of triples, each consisting of a concept, an offset, and a length. These triples are used directly by the Text Tab to highlight substrings and associate them with KB concepts.</Paragraph>
    <Paragraph position="2"> Consider the sentence &amp;quot;Washington, D.C. is a city.&amp;quot; Table 1 shows some example Scone concepts and their English names. Given a knowledge</Paragraph>
    <Section position="1" start_page="286" end_page="286" type="sub_section">
      <SectionTitle>
Name lists
</SectionTitle>
      <Paragraph position="0"> base with these concepts, SconeEdit returns the alignment: (concept: Washington-DC, offset: 1, length: 16) (concept: City, offset: 23, length: 4).</Paragraph>
    </Section>
  </Section>
  <Section position="7" start_page="286" end_page="286" type="metho">
    <SectionTitle>
6 Planned Features
</SectionTitle>
    <Paragraph position="0"> A single node in the KB could have hundreds or thousands of outgoing links. For readability, the browser must select a subset of these links to display to the user. We plan to leverage Scone's reasoning ability, along with SconeEdit's document-driven design, to select which nodes are likely to be relevant to the user in the context of the loaded document(s). For example, a user who views sub-classes of disease in a medical ontology may be presented with thousands of disease types. If the current document loaded into SconeEdit is a document about food, Scone may be able to prune the subclasses it lists to only food-borne illnesses.</Paragraph>
    <Paragraph position="1"> Another feature we hope to add is better integration with an entire corpus. The current system allows the user to work with individual documents.</Paragraph>
    <Paragraph position="2"> This could be extended to allow a user to navigate to a particular concept in the knowledge base and retrieve all documents in a corpus containing that concept (in its various forms). These documents could then be used to generate more KB concepts of interest.</Paragraph>
  </Section>
class="xml-element"></Paper>