# Natural Question Generation with Reinforcement Learning Based Graph-to-Sequence Model

**Authors:** Yu Chen, Lingfei Wu, Mohammed J. Zaki  
**arXiv:** [1910.08832](https://arxiv.org/abs/1910.08832)  
**Date:** 2019-10-19  
**Summary page:** https://www.summarizepaper.com/en/paper/1910.08832v1/

## Key points

- The paper proposes a novel approach to generate questions from a given passage and answer.
- The proposed model is based on reinforcement learning (RL) and graph-to-sequence (Graph2Seq) architecture.
- The model comprises a Graph2Seq generator that uses a Bidirectional Gated Graph Neural Network to embed the passage, and a hybrid evaluator that combines cross-entropy and RL losses to ensure the generation of syntactically and semantically valid text.
- The proposed model outperforms previous state-of-the-art methods by a significant margin on the SQuAD dataset.
- Natural question generation (QG) is an essential task in natural language processing with numerous applications such as conversational agents, educational systems, and search engines.
- To address the challenge of generating high-quality questions, the authors propose using graph-based representations of passages to capture their structural information better.
- The authors evaluate their proposed model on the SQuAD dataset and demonstrate that their approach outperforms several state-of-the-art models significantly in terms of various evaluation metrics such as BLEU-4, ROUGE-L, METEOR scores.
- Overall, this paper presents an innovative approach to natural question generation using reinforcement learning based graph-to-sequence models with promising results on benchmark datasets like SQuAD.

## Summary

The paper titled "Natural Question Generation with Reinforcement Learning Based Graph-to-Sequence Model" proposes a novel approach to generate questions from a given passage and answer. The proposed model is based on reinforcement learning (RL) and graph-to-sequence (Graph2Seq) architecture. The model comprises a Graph2Seq generator that uses a Bidirectional Gated Graph Neural Network to embed the passage, and a hybrid evaluator that combines cross-entropy and RL losses to ensure the generation of syntactically and semantically valid text. The proposed model outperforms previous state-of-the-art methods by a significant margin on the SQuAD dataset. The authors highlight that natural question generation (QG) is an essential task in natural language processing, which has numerous applications such as conversational agents, educational systems, and search engines. However, generating high-quality questions requires understanding the context of the passage and its relation to the answer. To address this challenge, the authors propose using graph-based representations of passages to capture their structural information better. The authors evaluate their proposed model on the SQuAD dataset, which is widely used for QG tasks. They compare their results with several state-of-the-art models and demonstrate that their approach outperforms them significantly in terms of various evaluation metrics such as BLEU-4, ROUGE-L, METEOR scores. Overall, this paper presents an innovative approach to natural question generation using reinforcement learning based graph-to-sequence models. The proposed method shows promising results on benchmark datasets like SQuAD and can have significant implications for various NLP applications such as chatbots or educational systems.