Emotional tension score in sentences

Consider multi-label classification, instead of a binary sentiment your dataset would have a level/degree/probability target for each emotional label of whatever dimension size you want.

In the context of neural networks, it's just a matter of replacing softmax with sigmoid loss layer.

If getting such large dataset with desired supervised labels is hard or not possible, one idea may be to use models trained on any sentiment datasets out there (there are even some multi-dimensional too) and apply additional un(semi)-supervised learning techniques to extract what you want.