In this thesis, we present the LoRe-MT system (Low Resources Machine
Translation) to translate Dutch to English. It is a demonstration of
techniques for constructing MT systems for languages lacking extensive
resources like bilingual corpora or complex parsers (so called low resource
languages). We develop and test methods for relatively cheaply and easily
building MT systems for new language pairs where extensive pre-existing
resources may not be available.
This thesis describes the system in detail, the resources used in its
development, and extensively evaluates it and its components.
The LoRe-MT system is a response to several important MT implementation
How far can we go in implementing quality MT systems using only basic
tools and rudimentary resources like shallow source language analysis,
bilingual dictionaries and processed monolingual target language corpora?
Can we resolves word and phrase order issues using exclusively target
Is it possible to perform adequate lexical selection (choosing the correct
translation from a set of candidates) using only target language corpora?
This thesis includes a general introduction to the problems posed by
machine translation, and proceeds to describe in detail the processed used
*Minimal resource shallow source analysis
*Source to target language transfer using a bilingual dictionary and a very
small rule set
*Matching outputs to the contents of target language corpora in order to
extract the most likely correct target language sentence structure
This system and the implementation choices we made are evaluated in depth,
and we discuss the potential for each component to be reused in other
natural language processing tasks. We also provide an outline for future
work and development of the LoRe-MT paradigm.
Keywords: hybrid machine translation, low resource languages, Dutch to