<?xml version="1.0" standalone="yes"?>
<Paper uid="P05-3025">
  <Title>Interactively Exploring a Machine Translation Model</Title>
  <Section position="3" start_page="0" end_page="97" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> There are many new approaches to statistical machine translation, and more ideas are being suggested all the time. However, it is difficult to determine how well a model will actually perform. Experienced researchers have been surprised by the capability of unintuitive word-for-word models; at the same time, seemingly capable models often have serious hidden problems -- intuition is no substitute for experimentation. With translation ideas growing more complex, capturing aspects of linguistic structure in different ways, it becomes difficult to try out a new idea without a large-scale software development effort.</Paragraph>
    <Paragraph position="1"> Anyone who builds a full-scale, trainable translation system using syntactic information faces this problem. We know that syntactic models often do not fit the data. For example, the syntactic system described in Yamada and Knight (2001) cannot translate n-to-m-word phrases and does not allow for multi-level syntactic transformations; both phenomena are frequently observed in real data. In building a new syntax-based MT system which addresses these flaws, we wanted to find problems in our framework as early as possible. So we decided to create a tool that could help us answer questions like: 1. Does our framework allow good translations for real data, and if not, where does it get stuck? 2. How does our framework compare to existing state-of-the-art phrase-based statistical MT systems such as Och and Ney (2004)? The result is DerivTool, an interactive translation visualization tool. It allows a user to build up a translation from one language to another, step by step, presenting the user with the myriad of choices available to the decoder at each point in the process. DerivTool simplifies the user's experience of exploring these choices by presenting only the decisions relevant to the context in which the user is working, and allowing the user to search for choices that fit a particular set of conditions. Some previous tools have allowed the user to visualize word alignment information (Callison-Burch et al., 2004; Smith and Jahr, 2000), but there has been no corresponding deep effort into visualizing the decoding experience itself. Other tools use visualization to aid the user in manually developing a grammar (Copestake and Flickinger, 2000), while our tool visualizes  Starting with: 0 and applying the rule: NPB(DT(the) NNS(police)) - 0 we get: NPB(DT(the) NNS(police)) If we then apply the rule: VBN(killed) we get: NPB(DT(the) NNS(police)) VBN(killed) Applying the next rule: NP-C(x0:NPB) - x0 results in: NP-C(NPB(DT(the) NNS(police))) VBN(killed) Finally, applying the rule: VP(VBD(was) VP-C(x0:VBN PP(IN(by) x1:NP-C))) - x1 x0 results in the final phrase: VP(VBD(was) VP-C(VBN(killed) PP(IN(by) NP-C(NPB(DT(the) NNS(police))))))  the translation process itself, using rules from very large, automatically learned rule sets. DerivTool can be adapted to visualize other syntax-based MT models, other tree-to-tree or tree-to-string MT models, or models for paraphrasing.</Paragraph>
  </Section>
class="xml-element"></Paper>