Neural Sequence-to-sequence Models with Constraints for Language Translation

Neural network architectures have been developed in order to address two frequently occurring tasks in language translation:

a) Given an original sentence and a translation of this sentence, a sentence similar to the original sentence has to be translated.

b) Given a sentence and a translation of a part of this sentence, a translation of the complete sentence has to be generated.

The above systems were developed particularly in view of patent translation and especially of patent translation from English to German. As part of this framework, a state-of-the-art compound word splitter for German was designed.