Classic SMT and feed-forward neural networks
From Asad Sayeed's statistical NLP course at the University of Gothenburg.

Azfar Imtiaz
Elin Hagman
Sandra Derbring

The models are trained by running the script train_model.py.
Targetfile: File for target language. (Default: UN-english.txt)
Sourcefile: File for source language. (Default: UN-french.txt)
Modelfile: Pre-trained vector file. (Default: GoogleNews-vectors-negative300.bin)
Trainedmodelfile: File used as output for the trained model. (Default: trainedmodel)
-M: Model type. 0 for trigram model or 1 for the translation model. (Default 0)
-B: Batch size for loading data into the model. A batch size of 1 is one sentence. (Default: 100)
-E: Epochs used to train the model for every batch. (Default: 20)
-L: Layer size for the neural network. (Default: 600)
-P: Using CPU or GPU for training the model. Select GPU with e.g. "cuda:0". (Default CPU)
-R: Learning rate for the Adam optimizer (Default: 0.01)
-S: Size in percentage of test set (Default: 0.2)\

The test is done by running test_models.py.
Targetfile: File for target language. (Default: UN-english.txt)
Sourcefile: File for source language. (Default: UN-french.txt)
Modelfile: Pre-trained vector file. (Default: GoogleNews-vectors-negative300.bin)
Languagemodel: The trained language model (Default: trained_language_model)
Translationmodel: The trained translation model (Default: trained_translation_mode)
-T: Select number of top n predicted words.(Default: 50)
-P: Using the CPU or GPU for training the model. Select GPU with e.g. "cuda:0". (Default CPU) -S: Size in percentage of test set (Default: 0.2)\