NeurIPS 2020

Compositional Generalization by Learning Analytical Expressions

Meta Review

This paper proposes a neuro-symbolic method (LAnE) for sequence-to-sequence tasks that exhibits systematicity and can solve both SCAN and MiniSCAN with perfect accuracy. LAnE works by identifying subexpressions in the input, translating them to the output domain (leveraging previously translated subexpressions stored in memory), and finally updating the outputs in memory. The reviewers found the paper to be of high quality, praising its novelty, the strength of the results, and the interest it is likely to generate amongst other researchers. In particular, R2 noted the “genuinely novel approach” and that the “results are striking”; similarly R4 stated that the paper presents “one of the most innovative approaches that has been proposed”. The biggest concern amongst the reviewers was clarity, with R2 and R3 mentioning they had to read the methods section several times and R1 also mentioning it was hard to follow. However, the reviewers were satisfied by the authors’ response in which they promise several concrete proposals to improve that section. I agree that the paper’s results are very strong and believe it will be of interest to the NeurIPS community. I therefore recommend acceptance.