<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-0503">
  <Title>The Problem of Precision in Restricted-Domain Question-Answering. Some Proposed Methods of Improvement</Title>
  <Section position="8" start_page="4" end_page="4" type="concl">
    <SectionTitle>
8 Discussions and Conclusions
</SectionTitle>
    <Paragraph position="0"> RDQA, working on small document collections and restricted subjects, seems to be a task no less difficult than open-domain QA. Due to candidate scarcity, the precision performance of a RDQA system, and in particular that of its IR module, becomes a problematic issue. It affects seriously the entire success of the system, because if most of the retrieved candidates are incorrect, it is meaningless to apply further techniques of QA to refine the answers.</Paragraph>
    <Paragraph position="1"> In this paper, we have discussed several methods to improve the precision performance of the IR module. They include the use of domain-specific terminology to rearrange the candidate list and to better characterize the question-document relevance relationship. Once this relationship has been well established, one can expect to obtain a small set of (almost) all relevant documents for a given question, and use this to guide the IR engine in a two-level search strategy.</Paragraph>
    <Paragraph position="2"> Also, long and complex answers may be a common characteristic of RDQA systems. Being aware of this, one can design appropriate systems which are more tolerant on answer size to achieve a higher precision, and to avoid the need of expanding a short but insufficient answer into a complete one. However, what a good answer should be is still an open question, which would need a lot more study to clarify.</Paragraph>
    <Paragraph position="3"> We have also presented applications of these methods in the real QA system for Bell Canada.</Paragraph>
    <Paragraph position="4"> Good improvements achieved compared to results of the original IR module show that these methods are applicable and effective.</Paragraph>
    <Paragraph position="5"> Many other problems on the precision performance of a RDQA system have not been tackled in this paper. Some of them relate to the free form of the questions: how to identify the category of the question (e.g. the mapping 'Who' -Person, 'When' - Time, 'How many' - Quantity, etc.), how to analyze the question into pragmatic parts (pre-suppositions, problem context, question focus), etc. Certainly, they are also problems of open-domain QA if one wants to go further than pre-defined question pattern tasks.</Paragraph>
  </Section>
class="xml-element"></Paper>