Plan for setup of machine translation

** NOTE: This documentation is old. It is kept since it may contain methodological points still valid.**

We plan to look at at least [Apertium

http://www.apertium.org/] (a rule-based system, cf. its wiki) and Moses (a statistically-based system). This document discusses the setup of Moses.

Overview

The programs should be installed on the Xserve machine, in order to facilitate long runs (may last for days).

Files

We need 5 different programs, cf. the download information on each page:

• alignment.jar, our Bergen-Tromsø sentence aligner
• Mosesdecoder (the mt program itself)
• giza++ (word alignment)
• srilm (language model)
• mkcls (word class/POS? training)

They shall be installed on the Xserve, and installed to standard paths.

The process

• Input is a set of parallel sentences

Setup

• Files where they belong
• Paths and access
• Modify makefiles

Make catalogues in gtsvn/mt

Today we have the catalogues:

• courses
• dev
• doc
• giza
• grantapplications
• script

Needed:

• change giza to wordalign, make one for sentencealign.
• have catalogues for the language pairs, and for the machine runs

MT systems, usage

smenob

• A gist system, i.e. in order to get an idea of what is written

nobsme

engsme

• Only KDE input