Change truecaser

[dmlls]

The truecasing process (which is carried out in the post-processing stage) is of great importance since it's the last step performed before returning the summary to the user. Even if the generated summary could be considered to be "good", the overall user's perception would be greatly affected if the truecasing was applied with low accuracy,

The current truecaser works just fine, but its accuracy could be improved. However, its "not-too-efficient" (to put it lightly) implementation, makes this task quite tedious both to train and later on, to use it, since the whole vocabulary is stored in memory, which means that dozens of GB of RAM are needed to make it work with a big vocabulary. A big vocabulary that, on the other hand, is needed for good accuracy.

We are currently considering the Stanford NLP Group's TrueCaseAnnotator as a suitable candidate to substitute the current truecaser.