Teaching Machines to Read and Comprehend

The paper addresses the challenge of teaching machines to read natural language documents, which can be tested by having systems answer questions about documents they have seen. A key obstacle had been the absence of large-scale training and test datasets for this kind of evaluation. To resolve this bottleneck, the authors define a new methodology that provides large-scale supervised reading comprehension data.

With this data available, they develop a class of attention-based deep neural networks that learn to read real documents and answer complex questions using minimal prior knowledge of language structure. Both the dataset-creation methodology and the attention-based reading models were influential, helping establish machine reading comprehension as a benchmark-driven research area.