This course gives an introduction to machine translation. Since the audience for this class is both computational linguists and translators, we will cover both the theory of how it works as well as how to use it in everyday work.

Outline

1. Machine Translation Overview
• Überblick über maschinelle Übersetzung - Read all of section 1, sections 2.1-2.3.1, all of section 3. CoLi student are strongly encouraged to get this book
• Silly, artificial illustration of statistical machine translation
• Concepts: source language (s or f), target language (t or e), parallel text, Vauquois pyramid, fluency & fidelity
2. Noisy Channel Model
   
   • Concepts: encoding/decoding, probability, prior & posterior probabilities
   • Applications: machine translation, speech recognition, spellchecking, OCR, many others
   • 
3. Install and Prepare Text Corpora
   The following is adapted from the Moses documentation:
   1. Create a directory/folder to store your corpora (text files)

      mkdir ~/corpora
      					

   2. Go to that directory

      cd ~/corpora
      

   3. Download big parallel corpora

      wget -c http://www.statmt.org/wmt15/training-parallel-nc-v10.tgz
      wget -c http://www.statmt.org/wmt15/dev-v2.tgz
      

   4. Unzip the compressed corpora

      tar zxvf training-parallel-nc-v10.tgz
      tar zxvf dev-v2.tgz
      

   5. Tokenize data (separate punctuation from words)

      ~/moses/scripts/tokenizer/tokenizer.perl -l de \
        < news-commentary-v10.de-en.de \
        > news-commentary-v10.de-en.tok.de
      ~/moses/scripts/tokenizer/tokenizer.perl -l en \
        < news-commentary-v10.de-en.en \
        > news-commentary-v10.de-en.tok.en
      

   6. Train a Truecaser. This helps Moses know which words should have an uppercase first letter, like nouns in German

      ~/moses/scripts/recaser/train-truecaser.perl \
        --model truecase-model.de \
        --corpus news-commentary-v10.de-en.tok.de
      ~/moses/scripts/recaser/train-truecaser.perl \
        --model truecase-model.en \
        --corpus news-commentary-v10.de-en.tok.en
      

   7. Truecase the input data, so that words are in their normal case. For example "Das" -> "das"

      ~/moses/scripts/recaser/truecase.perl \
        --model truecase-model.de \
        < news-commentary-v10.de-en.tok.de \
        > news-commentary-v10.de-en.tok.truecase.de
      ~/moses/scripts/recaser/truecase.perl \
        --model truecase-model.en \
        < news-commentary-v10.de-en.tok.en \
        > news-commentary-v10.de-en.tok.truecase.en
      

4. Language Models
   • Concepts: bigram, trigram, & n-gram, history, perplexity, backoff, interpolation
5. Word Alignment
   • Overview, slides 17-30
   • Additional slides
   • Download the word alignment program MGIZA:

     cd ~/moses
     wget https://github.com/moses-smt/mgiza/archive/master.zip
     mv master.zip mgiza.zip
     					

   • Unzip the file

     unzip mgiza.zip
     mv mgiza-master mgiza
     cd mgiza/mgizapp
     					

   • Compile the program. You first need cmake and gcc (or clang)

     cmake .
     make -j 4
     cp scripts/*.{sh,py,pl} bin/
     					

   • After running EMS (below), here is a sample alignment from the dataset, in ~/corpora/training/giza.1/en-de.A3.final.gz:

     elected parliaments do not own our liberties .
     NULL ({ }) gewählte ({ 1 }) Parlamente ({ 2 }) besitzen ({ 5 }) unsere ({ 6 }) Freiheiten ({ 7 }) nicht ({ 3 4 }) . ({ 8 })
     					

     
     
     
6. Moses Experiment Management System (EMS)
   1. Download the example configuration file:
      

      cd ~/corpora
      wget http://jon.dehdari.org/teaching/uds/moses/moses_ems_nc10.conf
      

   2. Edit the configuration file:
      

      1. nano moses_ems_nc10.conf

      2. Change the first few lines according to your computer's setup
      3. Type Ctrl/Strg + o to save the file
      4. Type Ctrl/Strg + x to exit the Nano text editor
   3. Do a dry run (Probelauf):
      

      ~/moses/scripts/ems/experiment.perl -config moses_ems_nc10.conf

   4. Do a real Moses run:
      

      nice ~/moses/scripts/ems/experiment.perl -config moses_ems_nc10.conf -exec

   5. The last command can take several hours to run, so you can just leave it running overnight. The output of a test set will be in the file: ~/corpora/evaluation/newstest2009.truecased.number .
7. Phrase-based MT
   • Concepts: phrase extraction, log-linear model, weights
   • More info
8. Hierarchical MT
   • Concepts: synchronous grammar, hierarchical phrase extraction
   
   
   
   
   
   
   
   
   
   
   
   
   
   • More info
   
   
9. Computer-aided Translation
   • Translation Memory
   • Post-editing: MateCat, PET

Term Paper / Hausarbeit

Due September 30th, 2016. Feel free to email me if you have any questions.

External Links

• Online Demonstration of the Moses Machine Translation System
• Moses Main Page
• The Full Moses Manual
• Free Corpora:
• WMT. Previous years contain different languages, including Hindi, Spanish, and Hungarian
• ACL Wiki, "Resources by Language"
• OPUS - open parallel corpora
• Corpus Processing Tools:
• corpus.tools
• https://github.com/jonsafari/habeas-corpus
• https://github.com/kpu/preprocess

Die Umgangssprache ist ein Teil des menschlichen Organismus und nicht weniger kompliziert als dieser.

–Ludwig Wittgenstein      

(Machine Translation: "Maria did not slap the green witch." –Wittgenstein)