<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-4008">
  <Title>K-QARD: A Practical Korean Question Answering Framework for Restricted Domain</Title>
  <Section position="4" start_page="0" end_page="29" type="metho">
    <SectionTitle>
2 Architecture of K-QARD
</SectionTitle>
    <Paragraph position="0"> As shown in Figure 1, K-QARD has four major components: Web information extraction, question analysis, answer finding, and answer generation. null The Web information extraction (IE) component extracts the domain-specific information for question answering from Web pages and stores the information into the relational database. For the domain portability, the Web IE component is based on the automatic wrapper induction approach which can be learned from small size of training examples.</Paragraph>
    <Paragraph position="1"> The question analysis component analyzes an  input question, extracts important information using the IE approach, and matches the question with pre-defined semantic frames. The component outputs the best-matched frame whose slots are filled with the information extracted from the question. In the answer finding component, K-QARD retrieves the answers from the database using the SQL generation script defined in each semantic frame. The SQL script dynamically generates SQL using the values of the frame slots.</Paragraph>
    <Paragraph position="2"> The answer generation component provides the answer to the user as a natural language sentence or a table by using the generation rules and the answer frames which consist of canned texts.</Paragraph>
  </Section>
  <Section position="5" start_page="29" end_page="31" type="metho">
    <SectionTitle>
3 Question Analysis
</SectionTitle>
    <Paragraph position="0"> The key component for ensuring the robustness and domain portability is the question analysis because it naturally requires many domain-dependent resources and has responsibility to solve the problem caused by various ways of expressing a question. In K-QARD, a question is analyzed using the methods devised by the information extraction approach. This IE-based question analysis method consists of several steps: 1. Natural language analysis: Analyzing the syntactic structure of the user's question and also identifiying named-entities and some important words, such as domain-specific predicate or terms.</Paragraph>
    <Paragraph position="1"> 2. Question focus recognition: Finding the intention of the user's question using the question focus classifier. It is learned from the training examples based on decision tree(C4.5)(Quinlan, 1993).</Paragraph>
    <Paragraph position="2">  3. Template Element(TE) recognition: Finding important concept for filling the slots of the semantic frame, namely template elements, using the rules, NE information, and ontology, etc.</Paragraph>
    <Paragraph position="3"> 4. Template Relation(TR) recognition: Find null ing the relation between TEs and a question focus based on TR rules, and syntactic information, etc.</Paragraph>
    <Paragraph position="4"> Finally, the question analysis component selects the proper frame for the question and fills proper values of each slot of the selected frame.</Paragraph>
    <Paragraph position="5"> Compared to other question analysis methods such as the complex semantic grammar(Martin et al., 1996), our approach has several advantages. First, it shows robust performance for the variation of a question because IE-based approach does not require the understanding of the entire sentence. It is sufficient to identify and process only the important concepts. Second, it also enhances the portability of the QA systems. This method is based on the divide-and-conquer strategy and uses only limited context information. By virture of these characteristics, the question analysis can be processed by using a small number of simple rules.</Paragraph>
    <Paragraph position="6"> In the following subsections, we will describe each component of our question analyzer in KQARD. null</Paragraph>
    <Section position="1" start_page="29" end_page="30" type="sub_section">
      <SectionTitle>
3.1 Natural language analysis
</SectionTitle>
      <Paragraph position="0"> The natural language analyzer in K-QARD identifies morphemes, tags part-of-speeches to them, and analyzes dependency relations between the morphemes. A stochastic part-of-speech tagger and dependency parser(Chung and Rim, 2004) for the Korean language are trained on a general domain corpus and are used for the analyzer. Then, several domain-specific named entities, such as a TV program name, and general named entities, such as a date, in the question are recognized using our dictionary and pattern-based named entity tagger(Lee et al., 2004). Finally some important words, such as domain-specific predicates, terminologies or interrogatives, are replaced by the proper concept names in the ontology. The manually constructed ontology includes two different types of information: domain-specific and general domain words.</Paragraph>
      <Paragraph position="1"> The role of this analyzer is to analyze user's question and transform it to the more generalized representation form. So, the task of the question focus recognition and the TE/TR recognition can be simplified because of the generalized linguistic information without decreasing the performance of the question analyzer.</Paragraph>
      <Paragraph position="2"> One of possible defects of using such linguistic information is the loss of the robustness caused by the error of the NLP components. However, our IE-based approach for question analysis uses the very restricted and essential contextual information in each step and can avoid such a risk successfully. null The example of the analysis process of this  'NE_*' denotes that the corresponding word is named entity of *. 'C_*' denotes that the corresponding word is belong to the concept C_* in the ontology. 'TE_*' denotes that the corresponding word is template element whose type is *. 'REL_OK' indicates that the corresponding TE and question focus are related.  component is shown in Figure 2-(1).</Paragraph>
    </Section>
    <Section position="2" start_page="30" end_page="30" type="sub_section">
      <SectionTitle>
3.2 Question focus recognition
</SectionTitle>
      <Paragraph position="0"> We define a question focus as a type of information that a user wants to know. For example, in the question GOWhat movies will be shown on TV tonight?GP, the question focus is a program title, or titles. For another example, the question focus is a current rainfall in a question GOSan Francisco is raining now, is it raining in Los Angeles too?GP.</Paragraph>
      <Paragraph position="1"> To find the question focus, we define question focus region, a part of a question that may contain clues for deciding the question focus. The question focus region is identified with a set of simple rules which consider the characteristic of the Korean interrogatives. Generally, the question focus region has a fixed pattern that is typically used in interrogative questions(Akiba et al., 2002). Thus a small number of simple rules is enough to cover the most of question focus region pattern. Figure 2-(2) shows the part recognized as a question focus region in the sample question.</Paragraph>
      <Paragraph position="2"> After recognizing the region, the actual focus of the question is determined with features extracted from the question focus region. For the detection, we build the question focus classifier using decision tree (C4.5) and several linguistic or domain-specific features such as the kind of the interrogative and the concept name of the predicate.</Paragraph>
      <Paragraph position="3"> Dividing the focus recognition process into two parts helps to increase domain portability. While the second part of deciding the actual question focus is domain-dependent because every domainapplication has its own set of question foci, the first part that recognizes the question focus region is domain-independent.</Paragraph>
    </Section>
    <Section position="3" start_page="30" end_page="31" type="sub_section">
      <SectionTitle>
3.3 TE recognition
</SectionTitle>
      <Paragraph position="0"> In the TE identification phase, pre-defined words, phrases, and named entities are identified as slot-filler candidates for appropriate slots, according to TE tagging rules. For instance, movie and NBC are tagged as Genre and Channel in the sample question GOTell me the movie on NBC tonight.GP (i.e.</Paragraph>
      <Paragraph position="1"> movie will be used to fill Genre slot and NBC will be used to fill Channel slot in a semantic frame). The hand-crafted TE tagging rules basically consider the surface form and the concept name (derived from domain ontologies) of a target word. The context surrounding the target word or word dependency information is also considered in some cases. In the example question of Figure 2, the date expression 'EAF4CIH2D8(today)', time expression 'B0DJELFIA3(night)' and the channel name 'MBC' are selected as TE candidates.</Paragraph>
      <Paragraph position="2"> In K-QARD, such identification is accomplished by a set of simple rules, which only examines the semantic type of each constituent word in the question, except the words in the question region. It is mainly because of our divide-and-conquer strategy motivated by IE. The result of this component may include some wrong template elements, which do not have any relation to the user's intention or the question focus. However, they are expected to be removed in the next component, the TR recognizer which examines the relation between the recognized TE and the question focus.</Paragraph>
    </Section>
    <Section position="4" start_page="31" end_page="31" type="sub_section">
      <SectionTitle>
3.4 TR recognition
</SectionTitle>
      <Paragraph position="0"> In the TR recognition phase, all entities identified in the TE recognition phase are examined whether they have any relationships with the question focus region of the question. For example, in the question GOIs it raining in Los Angeles like in San Francisco?GP, both Los Angeles and San Francisco are identified as a TE. However, by the TR recognition, only Los Angeles is identified as a related entity with the question focus region.</Paragraph>
      <Paragraph position="1"> Selectional restriction and dependency relations between TEs are mainly considered in TR tagging rules. Thus, the TR rules can be quite simplified.</Paragraph>
      <Paragraph position="2"> For example, many relations between the TEs and the question region can be simply identified by examining whether there is a syntactic dependency between them as shown in Figure 2-(4). Moreover, to make up for the errors in dependency parsing, lexico-semantic patterns are also encoded in the TR tagging rules.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="31" end_page="31" type="metho">
    <SectionTitle>
4 Application of K-QARD
</SectionTitle>
    <Paragraph position="0"> To evaluate the K-QARD framework, we built restricted domain question answering systems for the several domains: weather, broadcast, and traffic. For the adaptation of QA system to each domain, we rewrote the domain ontology consisting of about 150 concepts, about 30 TE/TR rules, and 7-23 semantic frames and answer templates. In addition, we learned the question focus classifier from training examples of about 100 questions for the each domain. All information for the question answering was automatically extracted using the Web IE module of K-QARD, which was also learned from training examples consisting of several annotated Web pages of the target Web site. It took about a half of week for two graduate students who clearly understood the framework to build each QA system. Figure 3 shows an example of QA system applied to the broadcast domain.</Paragraph>
  </Section>
class="xml-element"></Paper>