[Paper Review] Plan-And-Write: Towards Better Automatic Storytelling

Lili Yao, Nanyun Peng+4 more|arXiv (Cornell University)|Nov 14, 2018

Artificial Intelligence in Games34 references46 citations

TL;DR

The paper introduces a plan-and-write framework for open-domain story generation from titles, using dynamic and static storyline planning strategies to improve diversity, coherence, and topic relevance.

ABSTRACT

Automatic storytelling is challenging since it requires generating long, coherent natural language to describes a sensible sequence of events. Despite considerable efforts on automatic story generation in the past, prior work either is restricted in plot planning, or can only generate stories in a narrow domain. In this paper, we explore open-domain story generation that writes stories given a title (topic) as input. We propose a plan-and-write hierarchical generation framework that first plans a storyline, and then generates a story based on the storyline. We compare two planning strategies. The dynamic schema interweaves story planning and its surface realization in text, while the static schema plans out the entire storyline before generating stories. Experiments show that with explicit storyline planning, the generated stories are more diverse, coherent, and on topic than those generated without creating a full plan, according to both automatic and human evaluations.

Motivation & Objective

• Motivate automatic storytelling as a testbed for AI, aiming to generate long, coherent stories from a title.
• Propose a two-step framework: storyline planning and surface realization to improve quality.
• Explore two planning strategies (dynamic and static) and compare them.
• Leverage automatically extracted storylines to train planners without extra human annotation.
• Evaluate diversity, coherence, and topic relevance through automatic and human assessments.

Proposed method

• Represent storyline as an ordered sequence of words extracted from stories to train a plot-planning model.
• Implement two plan-and-write schemas: dynamic (incremental planning and writing) and static (plan full storyline first).
• Use content-introducing generation with context encoders (BiGRU) and attention-based decoders for both storyline and story generation.
• For static schema, use a Seq2Seq with attention to generate storyline from the title, then generate the story conditioned on title and planned storyline.
• Apply repetition suppression by forbidding repeated storyline words during generation to reduce redundancy.
• Evaluate with both automatic diversity metrics (inter- and intra-story repetition) and human judgments.

Experimental results

Research questions

• RQ1 Does explicit storyline planning improve diversity, coherence, and topic relevance of generated stories compared to baselines without planning?
• RQ2 Between dynamic and static planning, which yields more coherent and on-topic stories?
• RQ3 How do automatic diversity metrics correlate with human judgments in plan-and-write storytelling?

Key findings

• Plan-and-write methods produce more diverse and coherent stories than baselines without planning.
• Static schema generally outperforms dynamic schema in fidelity, coherence, interestingness, and overall preference.
• Dynamic schema reduces inter-story repetition but can increase final-sentence repetition compared to baselines.
• Static schema shows higher correlation between storyline quality (BLEU) and story quality (l–s score).
• Human evaluators preferred plan-and-write stories over baselines across fidelity, coherence, and interestingness, with coherence and interestingness being particularly influential.