Neural Module Networks for Reasoning over Text

TL;DR

This paper extends neural module networks to handle reasoning over natural language text, enabling better performance on complex question answering tasks involving symbolic reasoning and argument extraction.

Contribution

It introduces modules for reasoning over text with symbolic operations, an unsupervised auxiliary loss for argument extraction, and demonstrates improved results on the DROP dataset.

Findings

Outperforms state-of-the-art models on DROP subset

Modules effectively handle symbolic reasoning tasks

Auxiliary loss improves argument extraction accuracy

Abstract

Answering compositional questions that require multiple steps of reasoning against text is challenging, especially when they involve discrete, symbolic operations. Neural module networks (NMNs) learn to parse such questions as executable programs composed of learnable modules, performing well on synthetic visual QA domains. However, we find that it is challenging to learn these models for non-synthetic questions on open-domain text, where a model needs to deal with the diversity of natural language and perform a broader range of reasoning. We extend NMNs by: (a) introducing modules that reason over a paragraph of text, performing symbolic reasoning (such as arithmetic, sorting, counting) over numbers and dates in a probabilistic and differentiable manner; and (b) proposing an unsupervised auxiliary loss to help extract arguments associated with the events in text. Additionally, we show that a limited amount of heuristically-obtained question program and intermediate module output supervision provides sufficient inductive bias for accurate learning. Our proposed model significantly outperforms state-of-the-art models on a subset of the DROP dataset that poses a variety of reasoning challenges that are covered by our modules.