In this course we did two projects to create machine translation models.
This study applies two well-known translation models, namely IBM1 and IBM2, to train bilingual corpora and assign alignments to possible translations. Apart from the classical Expectation-Maximization training, a Variational Bayes model was also implemented for the first model. The performance of both models were indicated through both the perplexity of the train data and the alignment error-rate of the validation data.
Abstract: This project applies Neural Machine Translation in a parallel corpora. The model makes use of a sequence-to-sequence positional embedding with different encoders and decoders, namely: linear, Long Short-Term Memory and Gated Recurrent Unit. An attention mechanism was implemented in the decoder through dot and bilinear product in order to focus on specific areas of the sentences. Finally, the output sentences from the model were evaluated through four different scores.