CLIN 2005 Abstracts
  • Extracting, Representing, and Grounding Events for Temporal Question Answering
    David Ahn (ISLA, University of Amsterdam)
    Steven Schockaert (Dept. Applied Mathematics and Computer Science, Ghent University),
    Maarten de Rijke (ISLA, University of Amsterdam)
    Current question answering systems are tailored mainly to answer simple factoid questions, but research is beginning to focus on more complex classes of questions, including the class which we are investigating--temporal questions. There are two particularly challenging reasoning tasks involved in answering temporal questions. The system must be able to determine whether two descriptions refer to the same event (or to compatible events), and it must also be able to deduce which temporal relation holds between two events, a reasoning task that is complicated by the fact that historical events are often characterized by a gradual beginning and a gradual ending.

    In order to support temporal question answering, we have constructed a knowledge base of historical events off-line. The knowledge base is initially populated with tens of thousands of temporally grounded event descriptions extracted from the year and decade entries in Wikipedia, a freely available online encyclopedia. The hypertext structure of Wikipedia entries makes it straightforward not only to temporally ground events but also to identify important participants in events. The knowledge base is then further expanded using patterns to identify temporally related events in documents found on the Web. Finally, it is refined by constructing fuzzy sets for temporally vague events by polling the web for beginnings and endings.