Dynamic Interactive Attention Network for Question Answering

摘要

The disclosed technology relates to an end-to-end neural network for query answering, referred to herein as "Dynamic Interactive Attention Network (DCN)". At a high level, the DCN includes an encoder neural network and a cross-interest encoder that capture the interaction between the query and the document in a so-called “cross-interest encoding”. The DCN further includes a decoder neural network and a highway max-out network that process the mutual attention encoding to estimate the start and end positions of the phrases in the document that answers the question.