Pytorch 1.4+ and no NLTK

• Works with Pytorch versions after Pytorch 1.3
• Reduces memory footprint by converting LongTensors to IntTensors
• Removes NLTK dependency (self-contained tokenization)