X-SRL Dataset and mBERT Word Aligner

This code contains a method to automatically align words from parallel sentences by using multilingual BERT pre-trained embeddings. This can be used to transfer source annotations (for example labeled English
sentences) into the target side (for example a German translation of the
sentence) by transferring the label into the best-aligned target word.
This newly labeled data can be used to train different multilingual SOTA models to improve performance, especially for the lower-resource languages.