Answer-based Adversarial Training for Generating Clarification Questions

Sudha Rao, Hal Daumé III

Abstract: We propose a generative adversarial training approach for the problem of clarification question generation. Our approach generates clarification questions with the goal of eliciting new information that would make the given context more complete. We develop a Generative Adversarial Network (GAN) where the generator is a sequence-to-sequence model and the discriminator is a utility function that models the value of updating the context with the answer to the clarification question. We evaluate on two datasets, using both automatic metrics and human judgments of usefulness, specificity and relevance, showing that our approach outperforms both a retrieval-based model and ablations that exclude the utility model and the adversarial training.

Keywords: natural language processing, text generation, generative adversarial network

TL;DR: We propose an adversarial training approach to the problem of clarification question generation which uses the answer to the question to model the reward.

Community Implementations: [ 2 code implementations](https://www.catalyzex.com/paper/arxiv:1904.02281/code)