TL;DR: ICLR 2017 workshop submission
Abstract: In this paper, we propose to combine neural execution and symbolic execution to query a table with natural languages. Our approach makes use the differentiability of neural networks and transfers (imperfect) knowledge to the symbolic executor before reinforcement learning. Experiments show our approach achieves high learning efficiency, high execution efficiency, high interpretability, as well as high performance.
Conflicts: pku.edu.cn, huawei.com, DeeplyCurious.ai