<?xml version="1.0" standalone="yes"?>
<Paper uid="W04-2207">
  <Title>Identifying correspondences between words: an approach based on a bilingual syntactic analysis of French/English parallel corpora</Title>
  <Section position="17" start_page="4" end_page="4" type="concl">
    <SectionTitle>
7 Discussion
</SectionTitle>
    <Paragraph position="0"> The results achieved by the syntactic propagation method are quite encouraging. They show a high global precision rate - 94.3% for the INRA corpus and 93.1% for the JOC - assessed respectively against a reference list of approximately 8000 and 4600 alignments.</Paragraph>
    <Paragraph position="1"> Various reasons make it difficult to compare the results of this experiment with those reported in the literature and presented in section 2. Indeed, each approach has been tested on a different corpus and the results achieved could depend on the type of texts used for evaluation purposes. Moreover, the reference alignment lists, i.e. the gold standards, have probably been established according to different annotation criteria, which could also influence the quality of the results. Finally, each system has been designed, or at least used, to perform a specific task and evaluated in this respect. Daille, Gaussier and Lange (1994), as well as Gaussier (1998) and Hull (2001), were interested in bilingual terminology extraction so that word alignment could not be considered as an end in itself but rather as a basis for term alignment. The system proposed by Wu (2000) aims at bilingual language modelling, word and phrase alignment is incorporated as a subtask.</Paragraph>
    <Paragraph position="2"> Finally, Arhenberg, Andersson and Merkel (2000) as well as Lin and Cherry (2003) addressed the problem of full word alignment without restricting themselves to content words. Both noticed that the integration of linguistic knowledge, morphological and lexical for the former, syntactic for the latter, improves the alignment quality. However, concerning the approach proposed by Lin and Cherry (2003), it should be pointed out that linguistic knowledge is considered secondary to statistical information. As regards the alignment by syntactic propagation, linguistic knowledge is the kernel of the approach rather than an additional information.</Paragraph>
    <Paragraph position="3"> The propagation of alignments links using syntactic relations has proved very efficient when the same propagation pattern is used in both languages, i.e. when the syntactic structures are identical. A high level of precision is also achieved in the case of noun/adjective transpositions, even if the category of the words to be aligned varies. We are actually pursuing the study of non-correspondence between syntactic structures in English and French outlined in (Ozdowska and Bourigault, 2004). The aim is to determine whether there are some regularities in rendering certain English structures into certain French ones or not.</Paragraph>
    <Paragraph position="4"> If variation across languages is subjected to such regularities, the syntactic propagation could then be extended to the cases of non correspondence.</Paragraph>
  </Section>
class="xml-element"></Paper>