<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1119">
  <Title>Leveraging Reusability: Cost-effective Lexical Acquisition for Large-scale Ontology Translation</Title>
  <Section position="3" start_page="0" end_page="945" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> Multilingual access to digital collections is an important problem in today's increasingly interconnected world. Although technologies such as cross-language information retrieval and machine translation help humans access information they could not otherwise find or understand, they are often inadequate for highly specific domains.</Paragraph>
    <Paragraph position="1"> Most digital collections of any significant size use a system of organization that facilitates easy access to collection contents. Generally, the organizing principles are captured in the form of a controlled vocabulary of keyword phrases (descriptors) representing specific concepts. These descriptors are usually arranged in a hierarchic thesaurus or ontology, and are assigned to collection items as a means of providing access (either via searching for keyword phases, browsing the hierarchy, or a combination both). MeSH (Medical Subject Headings) serves as a good example of such an ontology; it is a hierarchicallyarranged collection of controlled vocabulary terms manually assigned to medical abstracts in a number of databases. It provides multilingual access to the contents of these databases, but maintaining translations of such a complex structure is challenging (Nelson, et al, 2004).</Paragraph>
    <Paragraph position="2"> For the most part, research in multilingual information access focuses on the content of digital repositories themselves, often neglecting significant knowledge that is explicitly encoded in the associated ontologies. However, information systems cannot utilize such ontologies by simply applying off-the-shelf machine translation. General-purpose translation resources provide insufficient coverage of the vocabulary contained within these domain-specific ontologies.</Paragraph>
    <Paragraph position="3"> This paper tackles the question of how one might efficiently translate a large-scale ontology to facilitate multilingual information access. If we need humans to assist in the translation process, how can we maximize access while minimizing cost? Because human translation is associated with a certain cost, it is preferable not to incur costs of retranslation whenever components of translated text are reused. Moreover, when exhaustive human translation is not practical, the most &amp;quot;useful&amp;quot; components should be translated first. Identifying reusable elements and prioritizing their translation based on utility is essential to maximizing effectiveness and reducing cost.</Paragraph>
    <Paragraph position="4">  We present a process of prioritized translation that balances the issues discussed above. Our work is situated in the context of the MALACH project, an NSF-funded effort to improve multi-lingual information access to large archives of spoken language (Gustman, et al., 2002). Our process leverages a small set of manuallyacquired English-Czech translations to translate a large ontology of keyword phrases, thereby providing Czech speakers access to 116,000 hours of video testimonies in 32 languages. Starting from an initial out-of-vocabulary (OOV) rate of 85%, we show that a small set of prioritized translations can be elicited from human informants, aligned, decomposed and then recombined to cover 90% of the access value in a complex ontology. Moreover, we demonstrate that prioritization based on hierarchical position and frequency of use facilitates extremely efficient reuse of human input. Evaluations show that our technique is able to boost performance of a simple translation system by 65%.</Paragraph>
  </Section>
class="xml-element"></Paper>