<?xml version="1.0" standalone="yes"?>
<Paper uid="E06-1050">
  <Title>A Probabilistic Answer Type Model</Title>
  <Section position="2" start_page="0" end_page="393" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Given a question, people are usually able to form an expectation about the type of the answer, even if they do not know the actual answer. An accurate expectation of the answer type makes it much easier to select the answer from a sentence that contains the query words. Consider the question &amp;quot;What is the capital of Norway?&amp;quot; We would expect the answer to be a city and could filter out most of the words in the following sentence: The landed aristocracy was virtually crushed by Hakon V, who reigned from 1299 to 1319, and Oslo became the capital of Norway, replacing Bergen as the principal city of the kingdom.</Paragraph>
    <Paragraph position="1"> The goal of answer typing is to determine whether a word's semantic type is appropriate as an answer for a question. Many previous approaches to answer typing, e.g., (Ittycheriah et al., 2001; Li and Roth, 2002; Krishnan et al., 2005), employ a predefined set of answer types and use supervised learning or manually constructed rules to classify a question according to expected answer type. A disadvantage of this approach is that there will always be questions whose answers do not belong to any of the predefined types.</Paragraph>
    <Paragraph position="2"> Consider the question: &amp;quot;What are tourist attractions in Reims?&amp;quot; The answer may be many things: a church, a historic residence, a park, a famous intersection, a statue, etc. A common method to dealwiththisproblemistodefineacatch-allclass.</Paragraph>
    <Paragraph position="3"> This class, however, tends not to be as effective as other answer types.</Paragraph>
    <Paragraph position="4"> Another disadvantage of predefined answer types is with regard to granularity. If the types are too specific, they are more difficult to tag. If they are too general, too many candidates may be identified as having the appropriate type.</Paragraph>
    <Paragraph position="5"> In contrast to previous approaches that use a supervised classifier to categorize questions into a predefined set of types, we propose an unsupervised method to dynamically construct a probabilistic answer type model for each question. Such a model can be used to evaluate whether or not a word fits into the question context. For example, given the question &amp;quot;What are tourist attractions in Reims?&amp;quot;, we would expect the appropriate answers to fit into the context &amp;quot;X is a tourist attraction.&amp;quot; From a corpus, we can find the words that appeared in this context, such as: A-Ama Temple, Aborigine, addition, Anak Krakatau, archipelago, area, baseball, Bletchley Park, brewery, cabaret, Cairo, Cape Town, capital, center, ...</Paragraph>
    <Paragraph position="6"> Using the frequency counts of these words in the context, we construct a probabilistic model to compute P(in(w,G)|w), the probability for a word w to occur in a set of contexts G, given an occurrence of w. The parameters in this model are obtained from a large, automatically parsed, unlabeled corpus. By asking whether a word would  occurinaparticularcontextextractedfromaques- null tion, we avoid explicitly specifying a list of possible answer types. This has the added benefit of being easily adapted to different domains and corpora in which a list of explicit possible answer types may be difficult to enumerate and/or identify within the text.</Paragraph>
    <Paragraph position="7"> The remainder of this paper is organized as follows. Section 2 discusses the work related to answer typing. Section 3 discusses some of the key concepts employed by our probabilistic model, including word clusters and the contexts of a question and a word. Section 4 presents our probabilistic model for answer typing. Section 5 compares the performance of our model with that of an oracle and a semi-automatic system performing the same task. Finally, the concluding remarks in are made in Section 6.</Paragraph>
  </Section>
class="xml-element"></Paper>