There are three Python programs here (-h for usage):

• ./rerank chooses the best candidate translations from a k-best list using a linear model.
• ./oracle computes a lower bound of BLEU on the development data.
• ./compute-bleu computes the BLEU score of a set of translations.

The commands are designed to work in a pipeline. For instance, these are valid invocations:

python rerank | python compute-bleu

python oracle | python compute-bleu

The data/ directory contains training, devlopment, and test data.

• train.src: Russian source sentences.
• train.ref: English reference sentences.
• train.100best: Candidate translations of train.src from a machine translation system.
• dev+test.src: Russian source sentences.
• dev.ref: English references sentences for the first half of dev+test.src.
• dev+test.100best: Candidate translations of dev+test.src.