# Stronger Transformers for Neural Multi-Hop Question Generation

arXiv: https://arxiv.org/abs/2010.11374 · HTML: https://mlsanity.com/paper/2010.11374/

## Key facts

- Cited by 6 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 22 Oct 2020.

## Built on by (most cited first, 6 of 6)

- [Automatic Generation of Socratic Subquestions for Teaching Math Word Problems](https://mlsanity.com/paper/2211.12835/) (2022)
- [Few-Shot Data Synthesis for Open Domain Multi-Hop Question Answering](https://mlsanity.com/paper/2305.13691/) (2023)
- [Improving Question Generation with Multi-level Content Planning](https://mlsanity.com/paper/2310.13512/) (2023)
- [QA4QG: Using Question Answering to Constrain Multi-Hop Question Generation](https://mlsanity.com/paper/2202.06538/) (2022)
- [Reasoning Circuits: Few-shot Multihop Question Generation with Structured Rationales](https://mlsanity.com/paper/2211.08466/) (2022)
- [Graph Neural Networks for Natural Language Processing: A Survey](https://mlsanity.com/paper/2106.06090/) (2021)

## Abstract (arXiv)

Prior work on automated question generation has almost exclusively focused on generating simple questions whose answers can be extracted from a single document. However, there is an increasing interest in developing systems that are capable of more complex multi-hop question generation, where answering the questions requires reasoning over multiple documents. In this work, we introduce a series of strong transformer models for multi-hop question generation, including a graph-augmented transformer that leverages relations between entities in the text. While prior work has emphasized the importance of graph-based models, we show that we can substantially outperform the state-of-the-art by 5 BLEU points using a standard transformer architecture. We further demonstrate that graph-based augmentations can provide complimentary improvements on top of this foundation. Interestingly, we find that several important factors--such as the inclusion of an auxiliary contrastive objective and data filtering could have larger impacts on performance. We hope that our stronger baselines and analysis provide a constructive foundation for future work in this area.