[Paper Review] A Skeleton-Based Model for Promoting Coherence Among Sentences in Narrative Story Generation

Jingjing Xu, Xuancheng Ren+4 more|arXiv (Cornell University)|Aug 21, 2018

Topic Modeling22 references4 citations

TL;DR

This paper proposes a skeleton-based generative model for narrative story generation that improves coherence by first generating key semantic phrases (skeletons) and then expanding them into full sentences. Using reinforcement learning to automatically learn skeletons without manual annotation, the model achieves a 20.1% improvement in G-score over state-of-the-art models in human evaluation.

ABSTRACT

Narrative story generation is a challenging problem because it demands the generated sentences with tight semantic connections, which has not been well studied by most existing generative models. To address this problem, we propose a skeleton-based model to promote the coherence of generated stories. Different from traditional models that generate a complete sentence at a stroke, the proposed model first generates the most critical phrases, called skeleton, and then expands the skeleton to a complete and fluent sentence. The skeleton is not manually defined, but learned by a reinforcement learning method. Compared to the state-of-the-art models, our skeleton-based model can generate significantly more coherent text according to human evaluation and automatic evaluation. The G-score is improved by 20.1% in the human evaluation. The code is available at https://github.com/lancopku/Skeleton-Based-Generation-Model

Motivation & Objective

• To address the challenge of low coherence in narrative story generation, where generated sentences often lack tight semantic connections.
• To model inter-sentence dependencies more effectively by focusing on key semantic phrases rather than generating full sentences in one step.
• To automatically learn sentence skeletons without relying on manually annotated or rule-based extraction methods.
• To improve story fluency and coherence through a two-stage generation process: skeleton generation followed by sentence expansion.
• To evaluate the model’s effectiveness using both automatic metrics and human evaluation, particularly focusing on coherence and semantic relevance.

Proposed method

• The model uses a two-part architecture: a skeleton-based generative module and a skeleton extraction module.
• The input-to-skeleton component maps input descriptions to semantic skeletons using a sequence-to-sequence model with attention.
• The skeleton-to-sentence component expands the generated skeleton into a fluent, complete sentence using another sequence-to-sequence model.
• A reinforcement learning framework connects the skeleton extraction module with the generative module to optimize skeleton quality end-to-end.
• The skeleton extraction module uses a differentiable policy network to select key phrases (e.g., predicates, subjects, objects) from full sentences, treating word selection as a discrete decision.
• The reward signal for reinforcement learning is derived from human evaluation scores (G-score), encouraging the model to generate skeletons that lead to more coherent stories.

Experimental results

Research questions

• RQ1 Can a two-stage generation approach—first generating skeletons and then expanding them—improve the coherence of narrative stories compared to end-to-end sequence-to-sequence models?
• RQ2 How effective is reinforcement learning in automatically learning sentence skeletons without human-annotated supervision?
• RQ3 To what extent does skeleton-based generation reduce semantic sparsity and improve inter-sentence dependency modeling?
• RQ4 How does the model perform under varying input lengths and unseen input phrases compared to existing models?
• RQ5 What types of coherence failures (e.g., irrelevant scenes, chaotic timelines) are most common in generated stories, and how does the model mitigate them?

Key findings

• The proposed skeleton-based model improves the G-score by 20.1% compared to state-of-the-art models in human evaluation, demonstrating significant gains in coherence.
• The model achieves a human evaluation coherence score of 5.62, which is substantially higher than the 4.98 score of the baseline Seq2Seq model.
• The skeleton extraction module successfully identifies key phrases, removing redundant modifiers while preserving essential semantic content, as shown in qualitative analysis of extracted skeletons.
• The model maintains relatively stable performance across varying input lengths and unseen phrase ratios, with a smaller drop in coherence compared to baselines.
• Error analysis reveals that irrelevant scenes (62.8%) are the primary cause of low coherence, followed by chaotic syntax (25.8%) and timeline inconsistencies (5.7%).
• The ablation study confirms that both the skeleton extraction module and reinforcement learning are essential, with the full model outperforming variants lacking either component.