DOC: Improving Long Story Coherence With Detailed Outline Control
Kevin Yang11{}^{1}start_FLOATSUPERSCRIPT 1 end_FLOATSUPERSCRIPT Dan Klein11{}^{1}start_FLOATSUPERSCRIPT 1 end_FLOATSUPERSCRIPT Nanyun Peng22{}^{2}start_FLOATSUPERSCRIPT 2 end_FLOATSUPERSCRIPT Yuandong Tian33{}^{3}start_FLOATSUPERSCRIPT 3 end_FLOATSUPERSCRIPT 11{}^{1}start_FLOATSUPERSCRIPT 1 end_FLOATSUPERSCRIPTUC Berkeley, 22{}^{2}start_FLOATSUPERSCRIPT 2 end_FLOATSUPERSCRIPTUCLA, 33{}^{3}start_FLOATSUPERSCRIPT 3 end_FLOATSUPERSCRIPTMeta AI
Abstract
We propose the Detailed Outline Control (DOC) framework for improving long-range plot coherence when automatically generating several-thousand-word-long stories. DOC consists of two complementary components: a detailed outliner and a detailed controller. The detailed outliner creates a more detailed, hierarchically structured outline, shifting creative burden from the main drafting procedure to the planning stage. The detailed controller ensures the more detailed outline is still respected during generation by controlling story passages to align with outline details. In human evaluations of automatically generated stories, DOC substantially outperforms a strong Re ${}^{3}$ baseline Yang et al. (2022) on plot coherence (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.
中文速览
为了让语言模型写出情节连贯、长达数千词的故事，研究者提出了"详细大纲控制"（Detailed Outline Control，DOC）框架，针对现有系统在长程情节一致性上的两大痛点——大纲太粗略、生成时又容易偏离大纲——同时发力。DOC 包含两个核心模块：一个"详细大纲生成器"，通过层次化、递归式展开把简短的三点大纲细化成带场景和角色信息的树状详细大纲，把创意决策从临场发挥提前到规划阶段；一个"详细控制器"，基于 FUDGE 方法在生成每个词时持续施加控制力，让正文始终贴合大纲细节，并通过对比训练引入流畅的难负例，防止长段落后半段跑题。在人工评测中，DOC 相比此前最强基线 Re³ 在情节连贯性上绝对提升 22.5%，大纲相关性提升 28.2%，故事趣味性提升 20.7%；在人机协作写作场景下，DOC 还支持用户在高层规划阶段介入，而非逐段干预，可控性获得压倒性好评。这项工作为长文本自动生成中"如何让模型既有详细计划又能忠实执行"这一核心难题提供了系统性解决思路，也为未来人机协同创作打下了基础。
原文 arXiv:2212.10077;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2212.10077v3