geovedi / combrot

Expand, combine, evaluate multiple machine translation systems output

combrot

Expand, combine, evaluate multiple machine translation systems output

Neural

Build word representations

fasttext skipgram -dim 600 -minn 1 -maxn 6 -input corpus.tok.src -output model.src
fasttext skipgram -dim 600 -minn 1 -maxn 6 -input corpus.tok.tgt -output model.tgt

The models can be built from monolingual corpuses.

Train translation matrix

python train_tm.py model.src model.tgt parallel.tok.src parallel.tok.tgt tm_model.pickle

Preparing input data

paste source.txt output_system_1.txt output_system_2.txt ... | perl -pe 's/\t/ ||| /g' > combined.txt

Similarity scoring

python combrot_sim.py model.src model.tgt tm_model.pickle combined.txt > output.sim.txt

LM scoring

lamtram --operation nbest --src_in output.sim.txt ... > output.ll.txt
paste output.sim.txt output.ll.txt | perl -pe 's/\t/ /g' > output.score.txt

Statistical

Build LM

lmplz -o 5 < corpus.tok.tgt > klm_o5.arpa
build_binary klm_o5.arpa klm_o5.bin

LM scoring

python combrot_lm.py klm_o5.bin combined.txt > output.lm.score