Numerical question answering with Neural Module Networks

Primary supervisor

Yuan-Fang Li

Neural module networks (NMNs) [1] support explainable question answering over text [2] by parsing a natural-language question into a program. Such a program consists of a number of differentiable neural modules that can be executed on text in a soft way, operating over attention scores. As a result, NMNs learn to jointly program and execute these programs in an end-to-end way.

Designed mainly for compositional (multi-hop) reasoning, NMNs provide limited support for numerical reasoning. It only supports number comparison, finding min/max in a paragraph and date differences.

Student cohort

Single Semester
Double Semester

Aim/outline

This project will design new modules that enhance the numerical reasoning capabilities of NMNs so that it can handle more complex questions.

URLs/references

  1. Andreas, J., Rohrbach, M., Darrell, T., & Klein, D. (2016). Neural module networks. In IEEE conference on computer vision and pattern recognition (CVPR) (pp. 39-48). https://arxiv.org/pdf/1912.04971

  2. Gupta, N., Lin, K., Roth, D., Singh, S., & Gardner, M. (2019). Neural module networks for reasoning over text. In International Conference on Learning Representations (ICLR). https://openreview.net/pdf?id=SygWvAVFPr

Required knowledge

Natural language processing

Deep learning