Speed Vs. Accuracy in Moses

Written on September 4, 2016

I ran a few experiments for Speed Vs. Accuracy in Moses SMT. 3000 sentences were translated from de, fr, es to en, while varying the following params:

  1. Stack size
  2. Translation option size
  3. Cube Pruning limit

By changing the stack size and cube pruning limit, we are effectively controlling the speed of the translation. Smaller stack and cube pruning limit will make translation faster at the cost of model score. (TODO:use Bleu score instead) I then plot Model Score Vs. Time taken for various settings of stack size, translation option size and cube pruning limit.

For each language there are 4 plots, all, short, medium and long for different sentences lengths. Short sentences are \(\le 10\) words, medium are \(11\) to \(30\) and the rest are considered long.

French-English Experiments:

Varying stack size and translation options (in this experiment stack size was made equal to cube prune limit)

German-English Experiments:

Varying stack size and translation options (in this experiment stack size was made equal to cube prune limit)

Spanish-English Experiments:

Varying stack size and translation options (in this experiment stack size was made equal to cube prune limit)