Siamese LSTMs for Translation Post-Edit Ranking

We consider the scenario of partially ranked crowd translations, where workers collaboratively post and peer edit prior translations. The iterative nature of the contributions, leads to interdependencies in the quality. We pose the problem as pair-wise comparisons of translations considering their history and model it using Siamese LSTM architecture. The LSTMs model translation dependencies and Siamese network model the preference function. We consider a supervised setting and predict the pair-wise comparisons for non-ranked translations. A sorting algorithm is used to get the complete set of rankings. The sequential nature of the problem is modeled well by LSTMs yielding 88.83% accuracy and 0.95 rank correlation, against a non-sequential Siamese DNN network providing an accuracy of 74.35% and 0.85 rank correlation, thus establishing the efficacy of the proposed approach.