<?xml version="1.0" standalone="yes"?>
<Paper uid="C02-1150">
  <Title>Learning Question Classifiers</Title>
  <Section position="3" start_page="0" end_page="1" type="intro">
    <SectionTitle>
2 Question Classification
</SectionTitle>
    <Paragraph position="0"> We define Question Classification(QC) here to be the task that, given a question, maps it to one of CZ classes, which provide a semantic constraint on the sought-after answer  . The intension is that this  We do not address questions like &amp;quot;Do you have a light?&amp;quot;, which calls for an action, but rather only factual Wh-questions. classification, potentially with other constraints on the answer, will be used by a downstream process which selects a correct answer from among several candidates.</Paragraph>
    <Paragraph position="1"> A question classification module in a question answering system has two main requirements. First, it provides constraints on the answer types that allow further processing to precisely locate and verify the answer. Second, it provides information that downstream processes may use in determining answer selection strategies that may be answer type specific, rather than uniform. For example, given the question &amp;quot;Who was the first woman killed in the Vietnam War?&amp;quot; we do not want to test every noun phrase in a document to see whether it provides an answer. At the very least, we would like to know that the target of this question is a person, thereby reducing the space of possible answers significantly. The following examples, taken from the TREC 10 question collection, exhibit several aspects of this point. Q: What is a prism? Identifying that the target of this question is a definition, strategies that are specific for definitions (e.g., using predefined templates) may be useful. Similarly, in: Q: Why is the sun yellow? Identifying that this question asks for a reason, may lead to using a specific strategy for reasons.</Paragraph>
    <Paragraph position="2"> The above examples indicate that, given that different answer types may be searched using different strategies, a good classification module may help the question answering task. Moreover, determining the specific semantic type of the answer could also be beneficial in locating the answer and verifying it. For example, in the next two questions, knowing that the targets are a city or country will be more useful than just knowing that they are locations. null Q: What Canadian city has the largest population? Q: Which country gave New York the Statue of Liberty? However, confined by the huge amount of manual work needed for constructing a classifier for a complicated taxonomy of questions, most question answering systems can only perform a coarse classification for no more than 20 classes. As a result, existing approaches, as in (Singhal et al., 2000), have adopted a small set of simple answer entity types, which consisted of the classes: Person, Location, Organization, Date, Quantity, Duration, Linear Measure. The rules used in the classification were of the following forms:  - If a query starts with Who or Whom: type Person.</Paragraph>
    <Paragraph position="3"> - If a query starts with Where: type Location.</Paragraph>
    <Paragraph position="4"> - If a query contains Which or What, the head noun  phrase determines the class, as for What X questions. While the rules used have large coverage and reasonable accuracy, they are not sufficient to support fine-grained classification. One difficulty in supporting fine-grained classification is the need to extract from the questions finer features that require syntactic and semantic analysis of questions, and possibly, many of them. The approach we adopted is a multi-level learning approach: some of our features rely on finer analysis of the questions that are outcomes of learned classifiers; the QC module then applies learning with these as input features.</Paragraph>
    <Section position="1" start_page="1" end_page="1" type="sub_section">
      <SectionTitle>
2.1 Classification Standard
</SectionTitle>
      <Paragraph position="0"> Earlier works have suggested various standards of classifying questions. Wendy Lehnert's conceptual taxonomy (Lehnert, 1986), for example, proposes about 13 conceptual classes including causal antecedent, goal orientation, enablement, causal consequent, verification, disjunctive, and so on. However, in the context of factual questions that are of interest to us here, conceptual categories do not seem to be helpful; instead, our goal is to semantically classify questions, as in earlier work on TREC (Singhal et al., 2000; Hovy et al., 2001; Harabagiu et al., 2001; Ittycheriah et al., 2001).</Paragraph>
      <Paragraph position="1"> The key difference, though, is that we attempt to do that with a significantly finer taxonomy of answer types; the hope is that with the semantic answer types as input, one can easily locate answer candidates, given a reasonably accurate named entity recognizer for documents.</Paragraph>
    </Section>
    <Section position="2" start_page="1" end_page="1" type="sub_section">
      <SectionTitle>
2.2 Question Hierarchy
</SectionTitle>
      <Paragraph position="0"> We define a two-layered taxonomy, which represents a natural semantic classification for typical answers in the TREC task. The hierarchy contains 6 coarse classes (ABBREVIATION, ENTITY, DESCRIPTION, HUMAN, LOCATION and NU-MERIC VALUE) and 50 fine classes, Table 1 shows the distribution of these classes in the 500 questions of TREC 10. Each coarse class contains a non-overlapping set of fine classes. The motivation behind adding a level of coarse classes is that of compatibility with previous work's definitions, and comprehensibility. We also hoped that a hierarchical classifier would have a performance advantage over a multi-class classifier; this point, however is not fully supported by our experiments.</Paragraph>
      <Paragraph position="1">  Class # Class # ABBREV. 9 description 7 abb 1 manner 2 exp 8 reason 6 ENTITY 94 HUMAN 65 animal 16 group 6 body 2 individual 55 color 10 title 1 creative 0 description 3 currency 6 LOCATION 81 dis.med. 2 city 18 event 2 country 3 food 4 mountain 3 instrument 1 other 50 lang 2 state 7 letter 0 NUMERIC 113 other 12 code 0 plant 5 count 9 product 4 date 47 religion 0 distance 16 sport 1 money 3 substance 15 order 0 symbol 0 other 12 technique 1 period 8 term 7 percent 3 vehicle 4 speed 6 word 0 temp 5 DESCRIPTION 138 size 0 definition 123 weight 4  over the question hierarchy. Coarse classes (in bold) are followed by their fine class refinements.</Paragraph>
    </Section>
    <Section position="3" start_page="1" end_page="1" type="sub_section">
      <SectionTitle>
2.3 The Ambiguity Problem
</SectionTitle>
      <Paragraph position="0"> One difficulty in the question classification task is that there is no completely clear boundary between classes. Therefore, the classification of a specific question can be quite ambiguous. Consider  1. What is bipolar disorder? 2. What do bats eat? 3. What is the PH scale?  Question 1 could belong to definition or disease medicine; Question 2 could belong to food, plant or animal; And Question 3 could be a numeric value or a definition. It is hard to categorize those questions into one single class and it is likely that mistakes will be introduced in the downstream process if we do so. To avoid this problem, we allow our classifiers to assign multiple class labels for a single question. This method is better than only allowing one label because we can apply all the classes in the later precessing steps without any loss.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>