The Goldilocks Principle: Reading Children's Books with Explicit Memory Representations

We introduce a new test of how well language models capture meaning in children's books. Unlike standard language modelling benchmarks, it distinguishes the task of predicting syntactic function words