Multilingual Neural Translation

Machine translation (MT) refers to the technology that can automatically translate contents in one language into other languages. Being an important research area in the field of natural language processing, machine translation has typically been considered one of most challenging yet exciting problems. Thanks to research progress in the data-driven statistical machine translation (SMT), MT is recently capable of providing adequate translation services in many language directions and it has been widely deployed in various practical applications and scenarios. Nevertheless, there exist several drawbacks in the SMT framework. The major drawbacks of SMT lie in its dependency in separate components, its simple modeling approach, and the ignorance of global context in the translation process. Those inherent drawbacks prevent the over-tuned SMT models to gain any noticeable improvements over its horizon. Furthermore, SMT is unable to formulate a multilingual approach in which more than two languages are involved. The typical workaround is to develop multiple pair-wise SMT systems and connect them in a complex bundle to perform multilingual translation. Those limitations have called out for innovative approaches to address them effectively. On the other hand, it is noticeable how research on artificial neural networks has progressed rapidly since the beginning of the last decade, thanks to the improvement in computation, i.e faster hardware. Among other machine learning approaches, neural networks are known to be able to capture complex dependencies and learn latent representations. Naturally, it is tempting to apply neural networks in machine translation. First attempts revolve around replacing SMT sub-components by the neural counterparts. Later attempts are more revolutionary by fundamentally changing the whole core of SMT with neural networks, which is now popularly known as neural machine translation (NMT). NMT is an end-to-end system which directly estimate the translation model between the source and target sentences. Furthermore, it is later discovered to capture the inherent hierarchical structure of natural language. This is the key property of NMT that enables a new training paradigm and a less complex approach for multilingual machine translation using neural models. This thesis plays an important role in the evolutional course of machine translation by contributing to the transition of using neural components in SMT to the completely end-to-end NMT and most importantly being the first of the pioneers in building a neural multilingual translation system. First, we proposed an advanced neural-based component: the neural network discriminative word lexicon, which provides a global coverage for the source sentence during the translation process. We aim to alleviate the problems of phrase-based SMT models that are caused by the way how phrase-pair likelihoods are estimated. Such models are unable to gather information from beyond the phrase boundaries. In contrast, our discriminative word lexicon facilitates both the local and global contexts of the source sentences and models the translation using deep neural architectures. Our model has improved the translation quality greatly when being applied in different translation tasks. Moreover, our proposed model has motivated the development of end-to-end NMT architectures later, where both of the source and target sentences are represented with deep neural networks. The second and also the most significant contribution of this thesis is the idea of extending an NMT system to a multilingual neural translation framework without modifying its architecture. Based on the ability of deep neural networks to modeling complex relationships and structures, we utilize NMT to learn and share the cross-lingual information to benefit all translation directions. In order to achieve that purpose, we present two steps: first in incorporating language information into training corpora so that the NMT learns a common semantic space across languages and then force the NMT to translate into the desired target languages. The compelling aspect of the approach compared to other multilingual methods, however, lies in the fact that our multilingual extension is conducted in the preprocessing phase, thus, no change needs to be done inside the NMT architecture. Our proposed method, a universal approach for multilingual MT, enables a seamless coupling with any NMT architecture, thus makes the multilingual expansion to the NMT systems effortlessly. Our experiments and the studies from others have successfully employed our approach with numerous different NMT architectures and show the universality of the approach. Our multilingual neural machine translation accommodates cross-lingual information in a learned common semantic space to improve altogether every translation direction. It is then effectively applied and evaluated in various scenarios. We develop a multilingual translation system that relies on both source and target data to boost up the quality of a single translation direction. Another system could be deployed as a multilingual translation system that only requires being trained once using a multilingual corpus but is able to translate between many languages simultaneously and the delivered quality is more favorable than many translation systems trained separately. Such a system able to learn from large corpora of well-resourced languages, such as English → German or English → French, has proved to enhance other translation direction of low-resourced language pairs like English → Lithuania or German → Romanian. Even more, we show that kind of approach can be applied to the extreme case of zero-resourced translation where no parallel data is available for training without the need of pivot techniques. The research topics of this thesis are not limited to broadening application scopes of our multilingual approach but we also focus on improving its efficiency in practice. Our multilingual models have been further improved to adequately address the multilingual systems whose number of languages is large. The proposed strategies demonstrate that they are effective at achieving better performance in multi-way translation scenarios with greatly reduced training time. Beyond academic evaluations, we could deploy the multilingual ideas in the lecture-themed spontaneous speech translation service (Lecture Translator) at KIT. Interestingly, a derivative product of our systems, the multilingual word embedding corpus available in a dozen of languages, can serve as a useful resource for cross-lingual applications such as cross-lingual document classification, information retrieval, textual entailment or question answering. Detailed analysis shows excellent performance with regard to semantic similarity metrics when using the embeddings on standard cross-lingual classification tasks