Neural machine translation with limited resources

With the advent of deep neural networks in recent years, Neural Machine Translation (NMT) systems have achieved state-of-the-art performance on standard translation benchmarks. NMT is a way to translate from one language to another with a single neural network in an end-to-end manner. The NMT models have emerged quickly, and within a few years of research, they have outperformed the traditional statistical systems with impressive performance. Despite the success of NMT models in standard benchmarks, there are some notable limitations. One of them is that NMT models are known to be data-hungry, i.e., they tend to work very well only when a massive amount of parallel training data (a.k.a. bitext) is available, but perform poorly when the data is limited. Except for some mainstream languages, e.g., English, French, or Chinese, most natural languages are low-resourced and lack large parallel data. Moreover, acquiring large bitext corpora is not viable in most scenarios, especially with resource-constrained conditions like low-resource languages. Researchers have made numerous endeavors to expand the success of NMT from high-resource to low-resource languages like transfer learning, data augmentation, and pivoting. However, they still require strong cross-lingual signals, i.e., lots of parallel data. One solution to this problem might be transferring cross-lingual signals through cross-lingual word embeddings (CLWEs), which can be learned from monolingual data in an unsupervised way or with the help of a small seed dictionary. CLWEs seem to be very promising in resource-constrained machine translation (MT). Most of the successful and predominant CLWE methods (a.k.a. word translation methods) learn a linear mapping function based on the isomorphic assumption, which is problematic. We hypothesize to learn the cross-lingual mapping in a projected latent space which would give the model enough flexibility to induce the required geometric structures such that it would be easier to align the embeddings. Based on this hypothesis, we propose two novel models for learning CLWEs. We empirically show that our methods are particularly very effective for low-resource languages. We then turn our attention from word- to sentence-level translation with limited resources. Specifically, we focus on data augmentation strategies widely used in NLP and Computer Vision to increase the robustness of the models in resource-constrained scenarios. We investigate the domain-mismatch issue thoroughly that hinders the all-embracing success of the existing techniques in NMT. Eventually, we introduce a novel data augmentation framework for low-resource NMT that leverages the neighboring samples of the original parallel data without explicitly using additional monolingual data. Our framework can diversify the in-domain parallel data in a controlled way. We perform extensive experiments on four low-resource language pairs comprising data from different domains. We have shown that our method is comparable to the traditional back-translation that uses extra in-domain monolingual data. Typically, NMT systems are trained on heterogeneous data from different domains, sources, topics, styles, and modalities. The quality of the data also varies a lot. Usually, during training, all the data are concatenated and randomly shuffled. However, not all of them may be useful, some data may be redundant, and some might even be noisy and detrimental to the final NMT system performance. These problems are more acute in low-resource languages compared to the high-resource ones. Consequently, we explore the possibilities of curriculum training for NMT systems, i.e., presenting the data to the NMT systems in a systematic order during training. We introduce a two-stage curriculum training framework for NMT where we fine-tune a base NMT model on subsets of data. To select the data subsets, we propose two scoring approaches --- deterministic scoring using pre-trained methods and online scoring that considers prediction scores of the emerging NMT model. Our curriculum strategies consistently demonstrate better translation quality and faster convergence (approximately 50% fewer updates) on both high- and low-resource languages.Doctor of Philosoph