<?xml version="1.0" standalone="yes"?>
<Paper uid="P03-2028">
  <Title>Spoken Interactive ODQA System: SPIQA</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Spoken Interactive QA system: SPIQA
</SectionTitle>
    <Paragraph position="0"> Figure 1 shows the components of our system, and the data that flows through it. This system comprises an ASR system (SOLON), a screening filter that uses a summarization method, and ODQA engine (SAIQA) for a Japanese newspaper text corpus, a Deriving Disambiguating Queries (DDQ) module, and a Text-to-Speech Synthesis (TTS) engine (FinalFluet). null</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
ASR system
</SectionTitle>
    <Paragraph position="0"> Our ASR system is based on the Weighted Finite-State Transducers (WFST) approach that is becoming a promising alternative formulation for the traditional decoding approach. The WFST approach offers a unified framework representing various knowledge sources in addition to producing an optimized search network of HMM states. We combined cross-word triphones and trigrams into a single WFST and applied a one-pass search algorithm to it.</Paragraph>
    <Paragraph position="1"> Screening filter To alleviate degradation of the QA's performance by recognition errors, fillers, word fragments, and other distractors in the transcribed question, a screening filter that removes these redundant and irrelevant information and extracts meaningful information is required. The speech summarization approach (C. Hori et. al., 2003) is applied to the screening process, wherein a set of words maximizing a summarization score that indicates the appropriateness of summarization is extracted automatically from a transcribed question, and these words are then concatenated together. The extraction process is performed using a Dynamic Programming (DP) technique.</Paragraph>
    <Paragraph position="2"> ODQA engine The ODQA engine, SAIQA, has four components: question analysis, text retrieval, answer hypothesis extraction, and answer selection.</Paragraph>
    <Paragraph position="3"> DDQ module When the ODQA engine cannot extract an appropriate answer to a user's question, the question is considered to be &amp;quot;ambiguous.&amp;quot; To disambiguate the initial questions, the DDQ module automatically derives disambiguating queries (DQs) that require information indispensable for answer extraction. The situations in which a question is considered ambiguous are those when users' questions exclude indispensable information or indispensable information is lost through ASR errors. These instances of missing information should be compensated for by the users.</Paragraph>
    <Paragraph position="4"> To disambiguate a question, ambiguous phrases within it should be identified. The ambiguity of each phrase can be measured by using the structural ambiguity and generality score for the phrase. The structural ambiguity is based on the dependency structure of the sentence; phrase that is not modified by other phrases is considered to be highly ambiguous. Figure 2 has an example of a dependency structure, where the question is separated into phrases. Each arrow represents the dependency between two phrases. In this example, &amp;quot;the World Cup&amp;quot; has no Which country won the world cupin Southeast Asia ?  modifiers and needs more information to be identified. &amp;quot;Southeast Asia&amp;quot; also has no modifiers. However, since &amp;quot;the World Cup&amp;quot;appears more frequently than &amp;quot;Southeast Asia&amp;quot; in the retrieved corpus, &amp;quot;the World Cup&amp;quot; is more difficult to identify. In other words, words that frequently occur in a corpus rarely help to extract answers in ODQA systems. Therefore, it is adequate for the DDQ module to generate questions relating to &amp;quot;World Cup&amp;quot; in this example, such as &amp;quot;What kind of World Cup?&amp;quot;,&amp;quot;What year was the World Cup held?&amp;quot;.</Paragraph>
    <Paragraph position="5"> The structural ambiguity of the n-th phrase is de-</Paragraph>
    <Paragraph position="7"> where the complete question is separated into N phrases, and D(P</Paragraph>
    <Paragraph position="9"> , which can be calculated using Stochastic Dependency Context-Free Grammar (SDCFG) (C. Hori et. al., 2003). Using this SDCFG, only the number of non-terminal symbols is determined and all combinations of rules are applied recursively. The non-terminal symbol has no specific function, such as a noun phrase. All the probabilities of rules are stochastically estimated based on data. Probabilities for frequently used rules become greater, and those for rarely used rules become smaller. Even though transcription results given by a speech recognizer are ill-formed, the dependency structure can be robustly estimated by our SDCFG.</Paragraph>
    <Paragraph position="10"> The generality score is defined as A</Paragraph>
    <Paragraph position="12"> where P(w) is the unigram probability of w based on the corpus to be retrieved. Thus, &amp;quot;w = cont&amp;quot; means that w is a content word such as a noun, verb or adjective.</Paragraph>
    <Paragraph position="13"> We generate the DQs using templates of interrogative sentences. These templates contain an interrogative and a phrase taken from the user's question, i.e., &amp;quot;What kind of * ?&amp;quot;, &amp;quot;What year was * held?&amp;quot; and &amp;quot;Where is * ?&amp;quot;.</Paragraph>
    <Paragraph position="14"> The DDQ module selects the best DQ based on its linguistic appropriateness and the ambiguity of the phrase. The linguistic appropriateness of DQs can be measured by using a language model, N-gram.</Paragraph>
    <Paragraph position="15">  are weighting factors to balance the scores.</Paragraph>
    <Paragraph position="16"> Hence, the module can generate a sentence that is linguistically appropriate and asks the user to disambiguate the most ambiguous phrase in his or her question.</Paragraph>
  </Section>
class="xml-element"></Paper>