philip30 / emnlp2016 Goto Github PK

## Lexicon - Experiment ##
#### Preparation
1. Extract the code and the data:
    $ tar -xvzf data.tar.gz   
    $ tar -xvzf chainn.tar.gz

2. Install all third party software:
    - Moses    : http://www.statmt.org/moses/                   # PBMT
    - GIZA++   : https://github.com/moses-smt/giza-pp           # Word Alignment
    - Travatar : http://www.phontron.com/travatar/              # Hiero
    - Chainer  : http://chainer.org/                            # Neural Network Toolkit
    - MTEval   : https://github.com/odashi/mteval               # Evaluator to calculate NIST and BLEU

3. Copy the script/bash/config-run.sh script/bash/config.sh
    $ cp script/bash/config-run.sh script/bash/config.sh

4. Open config.sh and resolve all the dependencies there!
    $ vi script/bash/config.sh

5. Run prepare.sh, this will create the filtered data and will extract lexical probability from eijiro
    $ ./prepare.sh

##### Execution
6. For convenience, you need to run the baseline experiment first because we will use the lexical translation probabilities
   produced from them:
    $ ./hiero.sh
    $ ./pbmt.sh

7. Let's make the hybrid lexicon first!
    $ ./script/hybrid.sh    

8. Once we have lexicons, you can reproduce the experiments by running btec.sh and kftt.sh:
    $ ./btec.sh
    $ ./kftt.sh

That's it. You can find the result in test/$experiment/$method/test-13.{out,align,result,nist.result}
Enjoy :)