<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2211">
  <Title>Speech/Language Technology Research, ETRI</Title>
  <Section position="2" start_page="0" end_page="0" type="intro">
    <SectionTitle>
1 Introduction
PBMT (Pattern-based Machine Translation)
</SectionTitle>
    <Paragraph position="0"> approach has been adopted by many MT researchers, mainly due to the portability, customizability and the scalability of the approach. cf. Hong et al. (2003a), Takeda (1996), Watanabe &amp; Takeda (1998). However, major drawback of the approach is that it is often very costly and time-consuming to construct a large amount of data enough to assure the performance of the PBMT system. From this reason many studies from PBMT research circles have been focused on the data acquisition issue. Most of the data acquisition studies were about automatic acquisition of lexical resources from bilingual corpus.</Paragraph>
    <Paragraph position="1"> Since 2001, ETRI has developed a Korean-Chinese MT system, TELLUS K-C, under the auspices of the MIC (Ministry of Information and Communication) of Korean government.</Paragraph>
    <Paragraph position="2"> We have adopted verb pattern based approach for Korean-Chinese MT. The verb patterns play the most crucial role not only in the transfer but also in the source language analysis. In the beginning phase of the development, most of the verb patterns were constructed manually by experienced Korean-Chinese lexicographers with some help of editing tools and electronic dictionaries. In the setup stage of a system, the electronic dictionary is very useful for building a verb pattern DB. It provides with a comprehensive list of entries along with some basic examples to be added to the DB. In most cases, however, the examples in the dictionary with which the lexicographers write a verb pattern are basic usages of the verb in question, and other various usages of the verb are often neglected. Bilingual corpus can be useful resources to extract verb patterns. However, as for language pairs like Korean-Chinese for which there are not so much bilingual corpus available in electronic form, the approach does not seem to be suitable. Another serious problem with the bilingual corpus-based approach is that the patterns extracted from the corpus can be domain-dependent.</Paragraph>
    <Paragraph position="3"> The verb pattern generation based on translation equivalency is another good alternative to data acquisition from bilingual corpus. The idea was originally introduced by Fujita &amp; Bond (2002) for Japanese to English MT.</Paragraph>
    <Paragraph position="4"> In this paper, we present a method to construct Korean-Chinese verb patterns from existing Korean-Chinese verb patterns that are manually written by lexicographers. The clue for the semi-automatic generation is provided by the idea that verbs of similar meanings often share the argument structure as already shown in Levin (1993). The synonymy among Korean verbs can be indirectly inferred from the fact that they have the same Chinese translation.</Paragraph>
    <Paragraph position="5"> We have already applied the approach to TELLUS K-C and increased the number of verb patterns from about 110,000 to 350,000. Though 350,000 patterns still contain many erroneous patterns, the evaluations in section 5 will show that the accuracy of the semi-automatically generated patterns is noteworthy and the pattern matching ratio improves significantly with</Paragraph>
  </Section>
class="xml-element"></Paper>