From Particular to General: A Preliminary Case Study of Transfer Learning in Reading Comprehension

In this paper we argue that transfer learning will be an important ingredient of general learning AI. We are especially interested in using data-rich domains to learn skills widely applicable in other domains. As a case study we explore transfer learning in reading comprehension. We train a neural-network-based model on two context-question-answer datasets, the Children’s Book Test and its larger extension, the BookTest, and we monitor transfer to a subset of bAbI tasks. Our initial experiments show only limited transfer between these domains. However, the transferred system is still significantly better than a random baseline.