This paper presents a hierarchical neural story generation approach that improves the coherence and fluency of generated stories. The authors collect a large dataset of 300,000 human-written stories paired with writing prompts from an online forum. They propose a hierarchical model that first generates a prompt, then conditions on this prompt when generating the story. This approach improves the relevance of the story to the prompt and allows for more structured and coherent text generation. The model uses a convolutional sequence-to-sequence architecture to efficiently model long documents, and introduces a novel gated multi-scale self-attention mechanism to capture long-range context. To further improve performance, the authors introduce a fusion mechanism that combines a pre-trained seq2seq model with a second model that focuses on the link between the prompt and the story. This fusion mechanism allows the model to better capture the dependencies between the input and output. Experiments show that the hierarchical model significantly outperforms strong baselines on both automated and human evaluations. Human judges prefer stories generated by the hierarchical model over those from a non-hierarchical model by a factor of two to one. The model also demonstrates improved performance in terms of fluency, topicality, and overall quality of generated stories. The authors also introduce new evaluation metrics to assess the performance of story generation models. These metrics isolate different aspects of story generation, such as coherence, fluency, and adherence to the prompt. The results show that the proposed hierarchical model and self-attention mechanisms significantly improve the quality of generated stories. The paper also discusses related work in story generation and hierarchical text generation, and highlights the importance of modeling long-range dependencies and conditioning on abstract prompts. The authors conclude that their approach represents a significant advancement in the field of creative text generation, enabling the generation of longer, more consistent, and more fluent stories.This paper presents a hierarchical neural story generation approach that improves the coherence and fluency of generated stories. The authors collect a large dataset of 300,000 human-written stories paired with writing prompts from an online forum. They propose a hierarchical model that first generates a prompt, then conditions on this prompt when generating the story. This approach improves the relevance of the story to the prompt and allows for more structured and coherent text generation. The model uses a convolutional sequence-to-sequence architecture to efficiently model long documents, and introduces a novel gated multi-scale self-attention mechanism to capture long-range context. To further improve performance, the authors introduce a fusion mechanism that combines a pre-trained seq2seq model with a second model that focuses on the link between the prompt and the story. This fusion mechanism allows the model to better capture the dependencies between the input and output. Experiments show that the hierarchical model significantly outperforms strong baselines on both automated and human evaluations. Human judges prefer stories generated by the hierarchical model over those from a non-hierarchical model by a factor of two to one. The model also demonstrates improved performance in terms of fluency, topicality, and overall quality of generated stories. The authors also introduce new evaluation metrics to assess the performance of story generation models. These metrics isolate different aspects of story generation, such as coherence, fluency, and adherence to the prompt. The results show that the proposed hierarchical model and self-attention mechanisms significantly improve the quality of generated stories. The paper also discusses related work in story generation and hierarchical text generation, and highlights the importance of modeling long-range dependencies and conditioning on abstract prompts. The authors conclude that their approach represents a significant advancement in the field of creative text generation, enabling the generation of longer, more consistent, and more fluent stories.