Dialog system with adaptive recurrent hopping and dual context encoding

An apparatus comprising:

• an encoder configured to generate an original context representation and a canonical context representation for a natural language query, wherein the original context representation comprises an entity and the canonical context representation comprises an entity type in place of the entity;
• a memory controller configured to obtain memory readouts from a knowledge base by performing an adaptive number of memory hops based on the natural language query;
• and
• a decoder configured to generate a response based on the canonical context representation and the response.

A method of training a memory network, comprising:

• identifying a training set, wherein each item in the training set comprises a natural language query and a ground truth response;
• predicting a response to the natural language query for each item of the training set using a memory network, wherein the memory network performs an adaptive number of memory hops;
• computing a loss function for each of the memory hops;
• and
• training the memory network is to determine the adaptive number of memory hops based on the computed loss functions for each of the adaptive number of memory hops.