Teaching Machines to Read and Comprehend

This paper tackles the challenge of teaching machines to read natural language documents, where machine reading systems can be tested on their ability to answer questions posed about the contents of documents they have seen. The authors observe that, until their work, large-scale training and test datasets for this kind of evaluation were missing, and they define a new methodology that resolves this bottleneck by providing large-scale supervised reading comprehension data.

Using this data, the authors develop a class of attention-based deep neural networks that learn to read real documents and answer complex questions with minimal prior knowledge of language structure. The work is significant for both establishing a methodology to create large-scale reading comprehension datasets and demonstrating attention-based models that learn to comprehend text.