“Ten years ago, deep learning for natural language processing was largely ineffective, with models failing on simple negation tasks like "This movie is not bad."”