<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-2010">
  <Title>Generating and Visualizing a Soccer Knowledge Base</Title>
  <Section position="3" start_page="0" end_page="124" type="metho">
    <SectionTitle>
2 System Overview
</SectionTitle>
    <Paragraph position="0"> The SOBA system consists of a web crawler, linguistic annotation components and a component for the transformation of linguistic annotations into an ontology-based representation.</Paragraph>
    <Paragraph position="1"> The web crawler acts as a monitor on relevant web domains (i.e. the FIFA2 and UEFA3 web sites), automatically downloads relevant documents from them and sends them to a linguistic annotation web service.</Paragraph>
    <Paragraph position="2"> Linguistic annotation and information extraction is based on the Heart-of-Gold (HoG) architecture (Callmeier et al. 2004), which provides a uniform and flexible infrastructure for building multilingual applications that use semantics- and XML-based natural language processing components.</Paragraph>
    <Paragraph position="3"> The linguistically annotated documents are further processed by the transformation component, which generates a knowledge base of soccer-related entities (players, teams, etc.) and events (matches, goals, etc.) by mapping annotated entities or events to ontology classes and their properties.</Paragraph>
    <Paragraph position="4"> Finally, an automatic hyperlinking component is used for the visualization of extracted entities and events. This component is based on the VieWs system, which was developed independently of SmartWeb (Buitelaar et al., 2005). In what follows we describe the different components of the system in detail.</Paragraph>
    <Section position="1" start_page="0" end_page="123" type="sub_section">
      <SectionTitle>
2.1 Web Crawler
</SectionTitle>
      <Paragraph position="0"> The crawler enables the automatic creation of a football corpus, which is kept up-to-date on a daily basis. The crawler data is compiled from texts, semi-structured data and copies of original  HTML documents. For each football match, the data source contains a sheet of semi-structured data with tables of players, goals, referees, etc. Textual data comprise of match reports as well as news articles.</Paragraph>
      <Paragraph position="1"> The crawler is able to extract data from two different sources: FIFA and UEFA. Semi-structured data, news articles and match reports covering the WorldCup2006 are identified and collected from the FIFA website. Match reports and news articles are extracted from the UEFA website. The extracted data are labeled by IDs that match the filename. The IDs are derived from the corresponding URL and are thus unique.</Paragraph>
      <Paragraph position="2"> The crawler is invoked continuously each day with the same configuration, extracting only data which is not yet contained in the corpus. In order to distinguish between available new data and data already present in the corpus, the URLs of all available data from the website are matched against the IDs of the already extracted data.</Paragraph>
    </Section>
    <Section position="2" start_page="123" end_page="123" type="sub_section">
      <SectionTitle>
2.2 Linguistic Annotation and Information
Extraction
</SectionTitle>
      <Paragraph position="0"> As mentioned before, linguistic annotation in the system is based on the HoG architecture, which provides a uniform and flexible infrastructure for building multilingual applications that use semantics- and XML-based natural language processing components.</Paragraph>
      <Paragraph position="1"> For the annotation of soccer game reports, we extended the rule set of the SProUT (Drozdzynski et al. 2004) named-entity recognition component in HoG with gazetteers, part-of-speech and morphological information. SProUT combines finite-state techniques and unification-based algorithms. Structures to be extracted are ordered in a type hierarchy, which we extended with soccer-specific rules and output types.</Paragraph>
      <Paragraph position="2"> SProUT has basic grammars for the annotation of persons, locations, numerals and date and time expressions. On top of this, we implemented rules for soccer-specific entities, such as actors in soccer (trainer, player, referee ...), teams, games and tournaments. Using these, we further implemented rules for soccer-specific events, such as player activities (shots, headers ...), game events (goal, card ...) and game results. A soccer-specific gazetteer contains soccer-specific entities and names and is supplemented to the general named-entity gazetteer.</Paragraph>
      <Paragraph position="3"> As an example, consider the linguistic annotation for the following German sentence from one of the soccer game reports: Guido Buchwald wurde 1990 in Italien Weltmeister (Guido Buchwald became world champion in 1990 in Italy)</Paragraph>
      <Paragraph position="5"/>
    </Section>
    <Section position="3" start_page="123" end_page="124" type="sub_section">
      <SectionTitle>
2.3 Knowledge Base Generation
</SectionTitle>
      <Paragraph position="0"> The SmartWeb SportEventOntology (a subset of SWIntO) contains about 400 direct classes onto which named-entities and other, more complex structures are mapped. The mapping is represented in a declarative fashion specifying how the feature-based structures produced by SProUT are mapped into structures which are compatible with the underlying ontology. Further, the newly extracted information is also interpreted in the context of additional information about the match in question.</Paragraph>
      <Paragraph position="1"> This additional information is obtained by wrapping the semi-structured data on relevant soccer matches, which is also mapped to the ontology. The information obtained in this way about the match in question can then be used as contextual background with respect to which the newly extracted information is interpreted.</Paragraph>
      <Paragraph position="2"> The feature structure for player as displayed above will be translated into the following F-Logic (Kifer et al. 1995) statements, which are then automatically translated to RDF and fed to the visualization component:</Paragraph>
      <Paragraph position="4"/>
    </Section>
    <Section position="4" start_page="124" end_page="124" type="sub_section">
      <SectionTitle>
2.4 Knowledge Base Visualization
</SectionTitle>
      <Paragraph position="0"> The generated knowledge base is visualized by way of automatically inserted hyperlink menus for soccer-related named-entities such as players and teams. The visualization component is based on the VIeWs4 system. VIeWs allows the user to simply browse a web site as usual, but is additionally supported by the automatic hyperlinking system that adds additional information from a (generated) knowledge base.</Paragraph>
      <Paragraph position="1"> For some examples of this see the included figures below, which show extracted information for the Panama team (i.e. all of the football players in this team in Figure 1) and for the player Roberto Brown (i.e. his team and events in which he participated in Figure 2).</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="124" end_page="124" type="metho">
    <SectionTitle>
3 Implementation
</SectionTitle>
    <Paragraph position="0"> All components are implemented in Java 1.5 and are installed as web applications on a Tomcat web server. SOAP web services are used for communication between components so that the system can be installed in a centralized as well as decentralized manner. Data communication is handled by XML-based exchange formats. Due to a high degree of flexibility of components, only a simple configuration over environment variables is needed.</Paragraph>
  </Section>
  <Section position="5" start_page="124" end_page="124" type="metho">
    <SectionTitle>
4 Conclusions and Future Work
</SectionTitle>
    <Paragraph position="0"> We presented an ontology-based approach to information extraction in the soccer domain that aims at the automatic generation of a knowledge base from match reports and the subsequent visualization of the extracted information through automatic hyperlinking. We argue that such an approach is innovative and enhances the user experience.</Paragraph>
    <Paragraph position="1"> Future work includes the extraction of more complex events, for which deep linguistic analysis and/or semantic inference over the ontology and knowledge base is required. For this purpose we will use an HPSG-based parser that is available within the HoG architecture (Callmeier, 2000) and combine this with a semantic inference approach based on discourse analysis (Cimiano et al., 2005).</Paragraph>
  </Section>
class="xml-element"></Paper>