So in my vision, this depends on the depth of the network. If you're doing simple word recognition then yes, you're going to end up with the most midwit of the midwit.

But, and this is a simple implementation: Suppose you use a text classifier on the individual sub-phrases, then for each one of those, you output neural layer snapshot represented as an image, then you take the images making up a sentence, and you feed them through a net to pattern recognize similar sentences and again you're outputting a neural snapshot as an image.

At each level of this, you can train using 2 similar phrases and one different. The reward function is based on the neural image of the similar phrases being more similar (XOR of the pixels is less) than the different one.

Feed those images back in, this time per-paragraph, and you should have a form of paragraph level classification. Then you feed that output into a network which which classifies text into a score and you train on things you find worthwhile.