This README contains the experimental settings used to create
translation models in our experiments on predicting learning curves.

A detailed step-by-step guide to this procedure can be found at 
the following url: http://www.statmt.org/wmt11/baseline.html

1. Pre-processing: Tokenization, Filtering, Lowercasing ( using scripts available as part of Moses )
2. Language Model: Order 5, Kneser-Ney smoothing on monolingual data of target language (using SRILM)
3. Training: Alignment heuristic: grow-diag-final-and, Re-ordering: msd-bidirectional-fe. 
4. MERT optimization procedure on held-out dataset.
5. Translation/Decoding using the optimized parameters from MERT.
6. Evaluation using BLEU.

For configurations where monolingual data in the target language is available (Europarl, News), 
we used such data to train the language model. 
However, in the case of EMEA and KFTT corpora, we used the target side of the parallel corpus
due to the unavailability of monolingual data. 
