Nominated for Best Paper Award (top 6.97% of accepted regular papers).

Ranked 1st for Czech→Polish and 2nd for Spanish→Portuguese translation systems.

Lifelong learning aims to enable information systems to learn from a continuous stream of data across time. However, this scenario is very challenging as the general limitations of machine learning methods apply to neural network-based models. Contemporary neural networks learn in isolation, and are not able to effectively learn new information without forgetting previously acquired knowledge.

[Read the full post]

The majority of current NMT systems is still trained using large bilingual corpora, which are available only for a handful of domains and high-resource language pairs. This is mainly caused by the fact that creating parallel corpora requires a great amount of resources (e.g. data, knowledge, time, money). Therefore, research on unsupervised MT focuses on eliminating the dependency on labeled data, which is especially beneficial for low-resource languages.

[Read the full post]