Machine translation (MT) technologies have been improved significantly in the last two decades, with the developments on phrased-based statistical MT (SMT) and recently the neural MT (NMT). However, most of these methods rely on the availability of large parallel data (millions to tens of millions sentence pairs) in the training, which are resources that do not exist in many language pairs. The development of monolingual MT is a recent approach that enables building MT systems without parallel data. However, a large amount of monolingual corpus is still required to train this kind of MT systems.
展开▼