Reinforcement Learning with TensorFlow

Mixed objective and deep residual coattention for Question Answering

The framework proposed in this research is based on the DCN model (see the preceding diagram), which consists of a coattention encoder and dynamic decoder pointer. The encoder encodes the question and document context separately and then forms a collaborative representation of the both through coattention followed by the decoder outputting the start and end position estimate as per the coattention.

In the new framework of DCN+, two new changes are introduced to the original DCN framework. They are as follows:

• Adding a deep residual coattention encoder
• Mixed training objective function which is the combination of the maximum likelihood cross-entropy loss function and reward ...