Quality Estimation from Scratch

This thesis presents a deep neural network for word-level machine translation quality estimation. The model extends the feedforward multi-layer architecture by [Collobert et al., 2011] to learning continuous space representations for bilingual contexts from scratch. By means of stochastic gradient descent and backpropagation of errors, the model is trained for binary classification of translated words, given only the source sentence and the machine translation. We enhance this model with alignments, and unsupervised pre-training of word representations allows for leveraging large monolingual corpora for supervised quality estimation training. Evaluating it on the data provided by the Workshop on Statistical Machine Translation 2014 and 2015, the model yields competitive results across languages and datasets. A linear combination of the deep model and a shallow linear model trained on baseline features further improves over both individual models. Furthermore, the bilingual word representations learnt during supervised training for quality estimation prove useful for other cross-lingual tasks.