[Paper Review] PlotMachines: Outline-Conditioned Generation with Dynamic Plot State Tracking

Hannah Rashkin, Aslı Çelikyılmaz+2 more|arXiv (Cornell University)|Apr 30, 2020

Topic Modeling37 references44 citations

TL;DR

PlotMachines introduces outline-conditioned story generation with dynamic plot state memory and discourse structure, showing improved coherence and outline adherence over strong baselines like GPT-2 and Grover across multiple datasets.

ABSTRACT

We propose the task of outline-conditioned story generation: given an outline as a set of phrases that describe key characters and events to appear in a story, the task is to generate a coherent narrative that is consistent with the provided outline. This task is challenging as the input only provides a rough sketch of the plot, and thus, models need to generate a story by interweaving the key points provided in the outline. This requires the model to keep track of the dynamic states of the latent plot, conditioning on the input outline while generating the full story. We present PlotMachines, a neural narrative model that learns to transform an outline into a coherent story by tracking the dynamic plot states. In addition, we enrich PlotMachines with high-level discourse structure so that the model can learn different writing styles corresponding to different parts of the narrative. Comprehensive experiments over three fiction and non-fiction datasets demonstrate that large-scale language models, such as GPT-2 and Grover, despite their impressive generation performance, are not sufficient in generating coherent narratives for the given outline, and dynamic plot state tracking is important for composing narratives with tighter, more consistent plots.

Motivation & Objective

• Define the task of outline-conditioned story generation and motivate the need for dynamic plot state tracking.
• Develop PlotMachines, a memory-augmented transformer that generates multi-paragraph stories from outlines.
• Incorporate high-level discourse structure to learn writing styles for different narrative parts (beginning, middle, end).
• Create and release three datasets pairing multi-paragraph narratives with automatically constructed outlines.
• Demonstrate limitations of large pre-trained models and show advantages of dynamic plot state tracking against baselines.

Proposed method

• Represent the outline as a token sequence with dedicated point delimiters and an end token, and condition generation on this outline for each paragraph.
• Maintain a memory matrix with two components K (outline point tracking) and D (latent document state) to track plot state across paragraphs.
• Update memory via gates after each paragraph using previous paragraph representation h^{i-1} to refine M^{i}.
• Modify Transformer blocks to include an additional memory-attention pathway that attends over memory while also performing standard self-attention; outputs are averaged.
• Incorporate a discourse tag (beginning, body, end) into the input to learn stylistic differences across narrative sections.
• Train end-to-end on cross-entropy loss predicting each paragraph, using gold previous paragraphs to update memory during training, and use a preset five-paragraph structure during decoding.

Experimental results

Research questions

• RQ1 Can an outline-conditioned generation model produce coherent long-form stories that adhere to a given outline?
• RQ2 Does dynamic plot state tracking improve coherence and outline adherence compared to non-memory baselines?
• RQ3 What is the impact of discourse-level structure on learning writing styles for different narrative parts?
• RQ4 Are memory-augmented models more effective than large pre-trained models (e.g., GPT, GPT-2, Grover) for outline-conditioned generation?

Key findings

• PlotMachines achieves comparable or higher ROUGE scores than baselines across Wikiplots, WritingPrompts, and NYTimes datasets.
• PlotMachines with GPT-2 base yields higher ROUGE than Grover on all three datasets in several metrics.
• Memory and discourse components are beneficial; ablations removing memory or discourse degrade performance.
• Human evaluations favor PlotMachines over GPT and Fusion in outline utilization, narrative flow, and overall ordering.
• PlotMachines demonstrates better diversity (lower self-BLEU) while maintaining coherence and outline adherence compared to baselines.