Using Verb Paraphrases for Arabic-to-English Example-Based Translation
Kfir Bar

Abstract:

We have developed an experimental Arabic-to-English example-based
machine translation (EBMT) system, which exploits a bilingual corpus
to find examples that match fragments of the input source-language
text--Modern Standard Arabic (MSA), in our case--and imitates its
translations. Translation examples were extracted from a collection of
parallel, sentence-aligned, unvocalized Arabic-English documents,
taken from several corpora published by the Linguistic Data
Consortium. The system is non-structural: translation examples are
stored as textual strings, with some additional inferred linguistic
features. In working with a highly inflected language, finding an
exact match for an input phrase with reasonable precision presumably
requires a very large parallel corpus. Since we are interesting in
studying the use of relatively small corpora for translation, matching
phrases to the corpus is done on a spectrum of linguistic levels, so
that not only exact phrases are discovered, but also related ones. In
this work, we investigate particularly the effect of matching
synonymous verbs.