MSc project at the University of Edinburgh under the supervision of Dr. Kenneth Heafield. Implemented the Reward Augmented Maximum Likelihood method and improved the quality of translation of the University of Edinburgh’s neural machine translation system by augmenting the training objective with reinforcement learning-inspired task rewards.
Obtained an improvement of 1.07 BLEU points over the baseline cross-entropy trained system.
Open-source implementation based on the Python Theano-based Nematus framework.
If you want details about methods and results, see the final thesis.