DOC: Improving Long Story Coherence With Detailed Outline Control

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 3378–3465 July 9-14, 2023 ©2023 Association for Computational Linguistics DOC: Improving Long Story Coherence With Detailed Outline Control Kevin Yang1 Dan Klein1 Nanyun Peng2 Yuandong Tian3 1UC Berkeley, 2UCLA, 3Meta AI {yangk,klein}@berkeley.edu,violetpeng@cs.ucla.edu,yuandong@meta.com Abstract We propose the Detailed Outline Control (DOC) framework for improving long-range plot coherence when automatically generating several-thousand-word-long stories. DOC con- sists of two complementary components: a de- tailed outliner and a detailed controller. The detailed outliner creates a more detailed, hier- archically structured outline, shifting creative burden from the main drafting procedure to the planning stage. The detailed controller ensures the more detailed outline is still re- spected during generation by controlling story passages to align with outline details. In hu- man evaluations of automatically generated sto- ries, DOC substantially outperforms a strong Re3 baseline (Yang et al., 2022) on plot coher- ence (22.5% absolute gain), outline relevance (28.2%), and interestingness (20.7%). Humans also judged DOC to be much more controllable in an interactive generation setting. 1 Introduction Recent advancements in natural language gener- ation systems have fueled increased interest in long-form text generation, in which texts may span thousands of words or more. Compared to tasks with shorter outputs, long-form generation involves meaningfully different challenges. It is nontrivial to maintain overarching coherence, or even basic relevance to an initial premise or plan. Even the most advanced language models to date, such as GPT4 (OpenAI, 2023), still cite long context as a major direction for further improvement, and re- quire structured planning to generate text longer than a few hundred words. In this work, we focus on long-form story gen- eration, which is representative of the major diffi- culties in long text generation. Only recently have prior efforts even attempted to generate stories of comparable length to human-authored “short sto- ries” (Re3, Yang et al. (2022)). Compared to hu- mans, state-of-the-art story generation systems like 1. Sue flies to Italy. a. … i. … ii. Sue packs up. b. … 2. … … … … … The day before her flight, Sue packed her bags, making sure to bring sunscreen … Detailed Outline Generated Story 1. Sue … 2. … Outline … Detailed Controller Detailed Outliner Figure 1: High-level overview of DOC. Our detailed outliner expands a brief initial outline into a more detailed outline. Our detailed controller then maintains faithfulness to the more detailed outline when drafting the main story. Re3 still fall short in numerous areas: common fail- ure modes include insufficient high-level planning resulting in local fluency amid global incoherence, or deviating from said planning even when it exists. To bridge some of this gap, we propose the De- tailed Outline Control (DOC) framework. While reusing the high-level planning-drafting-revision structure of Re3, DOC improves long-range plot coherence via two complementary approaches. First, our detailed outliner refines a brief ini- tial outline into a more detailed, hierarchical one (Figure 1 left). As motivation, a human author might also …

Content truncated to 500 words