Deep Neural Machine Translation with Linear Associative Unit
This addresses optimization difficulties in deep NMT systems, offering a simple yet effective solution for improving translation quality, though it appears incremental as it builds on existing RNN-based methods.
The paper tackled the problem of gradient diffusion in deep neural machine translation by proposing linear associative units (LAU), which improved Chinese-English translation by 11.7 BLEU over baseline models and achieved comparable results on English-German and English-French tasks.
Deep Neural Networks (DNNs) have provably enhanced the state-of-the-art Neural Machine Translation (NMT) with their capability in modeling complex functions and capturing complex linguistic structures. However NMT systems with deep architecture in their encoder or decoder RNNs often suffer from severe gradient diffusion due to the non-linear recurrent activations, which often make the optimization much more difficult. To address this problem we propose novel linear associative units (LAU) to reduce the gradient propagation length inside the recurrent unit. Different from conventional approaches (LSTM unit and GRU), LAUs utilizes linear associative connections between input and output of the recurrent unit, which allows unimpeded information flow through both space and time direction. The model is quite simple, but it is surprisingly effective. Our empirical study on Chinese-English translation shows that our model with proper configuration can improve by 11.7 BLEU upon Groundhog and the best reported results in the same setting. On WMT14 English-German task and a larger WMT14 English-French task, our model achieves comparable results with the state-of-the-art.