Transformer-based NMT : modeling, training and implementation

Abstract

International trade and industrial collaborations enable countries and regions to concentrate their developments on specific industries while making the most of other countries' specializations, which significantly accelerates global development. However, globalization also increases the demand for cross-region communication. Language barriers between many languages worldwide create a challenge for achieving deep collaboration between groups speaking different languages, increasing the need for translation. Language technology, specifically, Machine Translation (MT) holds the promise to enable communication between languages efficiently in real-time with minimal costs. Even though nowadays computers can perform computation in parallel very fast, which provides machine translation users with translations with very low latency, and although the evolution from Statistical Machine Translation (SMT) to Neural Machine Translation (NMT) with the utilization of advanced deep learning algorithms has significantly boosted translation quality, current machine translation algorithms are still far from accurately translating all input. Thus, how to further improve the performance of state-of-the-art NMT algorithm remains a valuable open research question which has received a wide range of attention. In the research presented in this thesis, we first investigate the long-distance relation modeling ability of the state-of-the-art NMT model, the Transformer. We propose to learn source phrase representations and incorporate them into the Transformer translation model, aiming to enhance its ability to capture long-distance dependencies well. Second, though previous work (Bapna et al., 2018) suggests that deep Transformers have difficulty in converging, we empirically find that the convergence of deep Transformers depends on the interaction between the layer normalization and residual connections employed to stabilize its training. We conduct a theoretical study about how to ensure the convergence of Transformers, especially for deep Transformers, and propose to ensure the convergence of deep Transformers by putting the Lipschitz constraint on its parameter initialization. Finally, we investigate how to dynamically determine proper and efficient batch sizes during the training of the Transformer model. We find that the gradient direction gets stabilized with increasing batch size during gradient accumulation. Thus we propose to dynamically adjust batch sizes during training by monitoring the gradient direction change within gradient accumulation, and to achieve a proper and efficient batch size by stopping the gradient accumulation when the gradient direction starts to fluctuate. For our research in this thesis, we also implement our own NMT toolkit, the Neutron implementation of the Transformer and its variants. In addition to providing fundamental features as the basis of our implementations for the approaches presented in this thesis, we support many advanced features from recent cutting-edge research work. Implementations of all our approaches in this thesis are also included and open-sourced in the toolkit. To compare with previous approaches, we mainly conducted our experiments on the data from the WMT 14 English to German (En-De) and English to French (En-Fr) news translation tasks, except when studying the convergence of deep Transformers, where we alternated the WMT 14 En-Fr task with the WMT 15 Czech to English (Cs-En) news translation task to compare with Bapna et al. (2018). The sizes of these datasets vary from medium (the WMT 14 En-De, ~ 4.5M sentence pairs) to very large (the WMT 14 En-Fr, ~ 36M sentence pairs), thus we suggest our approaches help improve the translation quality between popular language pairs which are widely used and have sufficient data.