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

• ./align aligns words using Dice's coefficient.
• ./check checks for out-of-bounds alignment points.
• ./grade computes alignment error rate.

The commands are designed to work in a pipeline. For instance, this is a valid invocation:

./align -t 0.9 -n 1000 | ./check | ./grade -n 5

The data/ directory contains a fragment of the German/English Europarl corpus.

• data/dev-test-train.de-en is the German/English parallel data to be aligned. The first 150 sentences are for development; the next 150 is a blind set you will be evaluated on; and the remainder of the file is unannotated parallel data.

• data/dev.align contains 150 manual alignments corresponding to the first 150 sentences of the parallel corpus. When you run ./check these are used to compute the alignment error rate. You may use these in any way you choose. The notation i-j means the word at position i (0-indexed) in the German sentence is aligned to the word at position j in the English sentence; the notation i?j means they are "probably" aligned.