Abstract

We present a formal language with expressions denoting general symbolstructures and queries which access information in those structures. Asequence-to-sequence network processing this language learns to encode symbolstructures and query them. The learned representation (approximately) shares asimple linearity property with theoretical techniques for performing this task.

 

Quick Read (beta)