Chain-of-Questions Training with Latent Answers for Robust Multistep Question Answering

Wang Zhu, Jesse Thomason, Robin Jia

Submission Type: Regular Long Paper

Submission Track: Question Answering

Keywords: multistep reasoning, question answering, latent variable learning

TL;DR: We propose Chain-of-Questions, a framework that generates sub-questions and sub-answers one at a time, trained by latent answer learning.

Abstract:

We propose Chain-of-Questions, a framework that trains a model to robustly answer multistep questions by generating and answering sub-questions. We obtain supervision for sub-questions from human-annotated question decomposition meaning representation (QDMR),
but QDMR does not include annotated answers to sub-questions. To overcome this technical challenge, we treat sub-answers as latent variables and infer them with a novel dynamic mixture of Hard-EM and MAPO. Chain-of-Questions is effective and robust, greatly outperforming strong neuro-symbolic methods by 9.0 F1 on a DROP contrast set and GPT-3.5 by 24.3 F1 on a HotpotQA adversarial set.

Submission Number: 5134