<?xml version="1.0" standalone="yes"?>
<Paper uid="W05-0830">
  <Title>Deploying Part-of-Speech Patterns to Enhance Statistical Phrase-Based Machine Translation Resources</Title>
  <Section position="4" start_page="164" end_page="165" type="metho">
    <SectionTitle>
3 Discussion
</SectionTitle>
    <Paragraph position="0"> The motivation behind this investigation has been to test whether syntactic or structural language aspects can be reflected or represented in the resources used in statistical phrase-based machine translation.</Paragraph>
    <Paragraph position="1"> We adopted a line of investigation that concentrates on the correspondence of part-of-speech patterns between French and English. We measured the usability of syntactic structures for statistical phrase-based machine translation by comparing translation performance when a standard phrase table was used, and when a syntactically enhanced phrase table was used. Both approaches scored very similarly. This similarity in the performance is justified by the following three factors.</Paragraph>
    <Paragraph position="2"> Firstly, the difference between the two translation resources, namely the lexical phrase table and the enhanced phrase table, does not relate to their entries, and thus their coverage, but to a simple alteration of the translation probability values of some of their entries. The coverage of these resources is exactly identical.</Paragraph>
    <Paragraph position="3"> Secondly, a closer examination of the translation probability value alterations that took place in order to reflect part-of-speech correspondences reveals that the proportion of the entries of the phrase table that were matched syntactically to phrases from the parallel corpus, and thus underwent a modification in their translation probability score, was very low (less than 1%). The reason behind this is the fact that the part-of-speech patterns produced by the parallel corpus were long strings in their vast majority, while the part-of-speech patterns found in the phrase table were significantly shorter strings. The inclusion of phrases longer than three words in translation resources has been avoided, as it has been shown not to have a strong impact on translation performance [Koehn et al., 2003].</Paragraph>
    <Paragraph position="4"> Thirdly, the above described translation probability value modifications were not parameterized, but consisted in a straightforward increase of the translation probability to its maximum value. It remains to be seen how these probability value alterations can be expanded to a type of probability value 'reweighing', in line with specific parameters, such as the size of the resources involved, the frequency of part-of-speech patterns in the resources, the length of part-of-speech patterns, as well as the syntactic classification of the members of part-of-speech patterns. If one is to compare the impact that such parameters have had upon the performance of automatic information summarisation [Mani, 2001] and retrieval technology [Belew, 2000], it may be worth experimenting with such parameter tuning when refining machine translation resources.</Paragraph>
    <Paragraph position="5"> A note should be made to the choice of tagger for our experiments. A possible risk when attempting any syntactic examination of a large set of data may stem from the overriding role that syntax often assumes over semantics. Statistical phrase-based machine translation has been faced with instances of this phenomenon, often disguised as linguistic idiosyncrasies. This phenomenon accounts for such instances as when nouns appear in pronominal positions, or as adverbial modifiers.</Paragraph>
    <Paragraph position="6"> On these occasions, and in order for the syntactic examination to be precise, words would have to be defined on the basis of their syntactic distribution rather than their semantic function. The TreeTagger abides by this convention, which is one of the main reasons why we chose it over a plethora of other freely available taggers, the remaining reasons being its high speed and low error rate. In addition, it should be clarified that there is no statistical, linguistic, or other reason why we chose to adopt the English version of the Penn TreeBank tagset over the French, as they are both equally conclusive and transparent.</Paragraph>
    <Paragraph position="7"> The overall driving force behind our investigation has been to test whether part-of-speech structures can be of assistance to the enhancement of translation resources for statistical phrase-based machine translation. We view our use of part-of-speech patterns as a natural extension to the introduction of structural elements to statistical machine translation by Wang [1998] and Och et al. [1999].</Paragraph>
    <Paragraph position="8">  Our empirical results suggest that the use of part-of-speech pattern correspondences to enhance existing translation resources does not damage machine translation performance. What remains to be investigated is how this approach can be optimized, and how it would respond to known statistical machine translation issues, such as mapping nested structures, or the handling of 'unorthodox' language pairs, i.e. agglutinative-fusion languages.</Paragraph>
  </Section>
class="xml-element"></Paper>