Multi-hop Question Generation without Supporting Fact Information

Question generation is the parallel task of question answering, where given an input context and optionally, answer, goal to generate a relevant fluent natural language question. Although recent works on have experienced success by utilizing sequence-to-sequence models, there need for models handle increasingly complex contexts with producing elaborate questions. Multi-hop more challenging that aims questions connecting multiple facts from contexts. In this work we apply transformer model multi-hop generation, without any sentence-level supporting fact information. We utilize concepts proven effective in single-hop including copy mechanism placeholder tokens. evaluate our model's performance HotpotQA dataset using automated evaluation metrics human evaluation, show improvement over previous works.