#Context-Sensitive Translation with Non-Linear Rerankers

Installing

Make sure you have cnn installed. Also ensure that you have the development version of the Eigen library available.

Locate the necessary libraries for us:

ln -s [PATH_TO_CNN] ./.cnn
ln -s [PATH_TO_EIGEN] ./.eigen

Then to compile, run

make -j 2

Training Models and decoding

To train a non-linear reranker on the dense features, run:

./bin/train [k-best-file] > model_file

To train a non-linear reranker on the dense features with the context sensitive features, run:

./bin/train [k-best-file] --context [SOURCE_EMBEDDINGS] [TARGET_EMBEDDINGS] [SOURCE_SENTENCES] > model_file

To decode (rerank) with a pre-trained model, run

./bin/rerank [MODEL] [k-best-file] [SOURCE_SENTENCES]

Many further options are available are training. To find out more, run

./bin/train --help