It's interesting that the first two - listen even without comprehension, and deprioritize grammar - correspond significantly to advances in NLP on the last few

It's interesting that the first two - listen even without comprehension, and deprioritize grammar - correspond significantly to advances in NLP on the last few years. Unsupervised language models (that learn to predict text without any labeled task) turn out to be valuable through transfer learning in specific comprehension tasks. And old school NLP methods that focused on grammar tree construction have fallen behind attention and transformer based methods that learn softer, more abstract representations of structure.