Abstract

We have seen significant improvements in machine translation due to the usage of
deep learning. While the improvements in translation quality are impressive, the
encoder-decoder architecture enables many more possibilities. In this talk, I 
will present methods to integrate additional constraints in the model: 
length-constraints to display the output in a given format as well as 
time-constraints in low-latency speech translation.

We will focus on methods that learn these additional constraints in an 
unsupervised way. So the model is able to make use of the standard parallel 
training data for machine translation. In order to achieve this, multiple 
components of the development of a machine translation system have to be 
addressed. In this talk, we will cover the adaptation of the learning, the 
architecture and the inference. 

By applying these methods, we are able to generate shorter translations that 
directly summarize the source sentence. In a study, we will compare these to 
human subtitles used in TV. By combining the methods with multi-lingual machine
translation, we are also able to perform monolingual shortening. For low-latency
 translation, we show that the latency can be significantly reduced with only 
 minimal losses in translation quality.