MoPS:用于开放式自动故事生成的模块化故事前提综合

Yan Ma1,3,4111 This work was done during Yan Ma’s internship at Shanghai AI Laboratory.Yu Qiao3Pengfei Liu2,3,4222通讯作者.
1
Fudan University 2Shanghai Jiao Tong University
3Shanghai AI Laboratory 4Generative AI Research Lab (GAIR)
yanma23@m.fudan.edu.cn
qiaoyu@pjlab.org.cnpengfei@sjtu.edu.cn
摘要

故事前提简洁地定义了故事的主要思想、基础和轨迹。 它充当自动故事生成的初始触发器。 现有的故事前提来源受到缺乏多样性、质量参差不齐和成本高昂的限制,使得它们难以扩展。 作为回应,我们引入了Modular Story Premise Synthesis (MoPS),它将故事前提分解为背景和角色等模块,以进行自动化设计和一代。 MoPS由三个阶段组成:(1)为每个模块预先收集一组一致的候选者,以形成嵌套字典。 (2)从嵌套字典中提取关键路径作为设计前提。 (3)指导大语言模型将设计整合成连贯的前提句。 彻底的评估表明,与从大型语言模型中归纳并从公共故事数据集中捕获的前提相比,我们合成的前提在多样性、吸引力、完整性和原创性方面表现出色。 同样,我们生成的扩展小说和剧本也表现出更高的质量。 在补充材料中,我们提供了 MoPS 代码套件,以及 7.6k 生成的前提和 1k 扩展故事。 代码:https://github.com/GAIR-NLP/MoPS

MoPS:用于开放式自动故事生成的模块化故事前提综合


Yan Ma1,3,4111 This work was done during Yan Ma’s internship at Shanghai AI Laboratory.  Yu Qiao3  Pengfei Liu2,3,4222 Corresponding author. 1Fudan University 2Shanghai Jiao Tong University 3Shanghai AI Laboratory 4Generative AI Research Lab (GAIR) yanma23@m.fudan.edu.cnqiaoyu@pjlab.org.cnpengfei@sjtu.edu.cn


1简介

如果一个故事要失败,

它首先会在前提层面失败。


前提线剖析

Premise A: A powerful Roman general,granted immortality by a vengeful deity, must choose between loyalty to the empire and leading a rebellion, ultimately decide to bring justice and equality to the corrupt rulers of the Roman Empire. Premise B: A Roman general chooses between loyalty to empire and leading a rebellion for justice and equality.
Poster A: [Uncaptioned image]
Story A: Centuries ago, in the heart of ancient Rome, there was a powerful general named Lucius, who had been granted immortality by a vengeful deity. As the empire crumbled under the weight of corruption and tyranny, Lucius faced a fateful decision: remain loyal to the empire that had once revered him, or lead a rebellion to bring justice and equality to its corrupt rulers. After witnessing the suffering of the oppressed, Lucius chose to forsake his loyalty to the empire, vowing to use his immortality to dismantle the corrupt regime. With unwavering determination, he rallied the oppressed citizens, becoming a symbol of hope and change. Together, they waged a relentless battle for a brighter future, where justice and equality would reign, and the legacy of Lucius would live on through the ages as a beacon of righteous rebellion against tyranny.
Poster B: [Uncaptioned image]
Story B: In the heart of the Roman Empire, General Lucius found himself torn between two irreconcilable paths. On one hand, he had sworn an oath of unwavering loyalty to the empire that had granted him power and prestige. On the other, he witnessed the relentless oppression and suffering of the common people, burdened by heavy taxes and brutal tyranny. The call for justice and equality resonated within him like a thunderous battle drum. In a moment of agonizing decision, he chose to forsake his allegiance to the empire, forsaking his name and status to lead a rebellion for the oppressed, determined to wield his military prowess to carve a path toward a more just and equitable future.
Review by GPT4: Story A is better than B. Story A is written with a more vivid and immersive narrative style. It provides greater detail about the protagonist’s character, motivations, and the historical context. Phrases like "vowed to use his immortality to dismantle the corrupt regime" and "a beacon of righteous rebellion against tyranny" add a mythical and heroic dimension to the protagonist, Lucius. The mention of a vengeful deity granting immortality to Lucius adds a layer of mythological intrigue and depth to his character. Furthermore, the depiction of the empire’s corruption and Lucius’s transition from loyalty to rebellion is more dramatically and emotionally engaging, enhancing the reader’s investment in the story. Story B, while also well-written, lacks the same level of detail and narrative flair. It presents a more straightforward account of Lucius’s dilemma and decision. The story misses opportunities to deepen the character development and historical context, resulting in a narrative that feels less engaging and impactful compared to A. Additionally, B repeats certain ideas, such as Lucius being torn between loyalty and justice, which can make the plot feel slightly redundant.
表1: 这个例子强调了故事前提的重要性。 A是MoPS创建的前提,B是其手动简化的形式。 这些前提指导 Dall-Echatgpt 生成海报和故事,gpt-4-turbo 评论强调了精心制作的作用发展引人入胜的叙事的前提。

前提是你的故事是关于Field (2005)的;里昂(2015) 故事前提是一条简洁的线索,捕捉故事的主要思想、冲突和人物,概述故事的基础和方向Truby (2008);克罗恩(2012);布罗迪(2018) 作家利用前提来指导故事发展,提供对人物、情节、主题和解决方案的战略洞察。 在自动故事生成(ASG)方面,大量研究探索了各种系统Fan 等人(2018,2019);姚等人 (2019);杨等人 (2022, 2023);周等人(2023) 这些系统需要输入来触发和指导故事创作。 前提就是这样的输入,为复杂的叙事发展提供了起点。 然而,制作一个故事前提对艺术和技术技能提出了挑战,需要捕捉简约文本Lyons (2015)中的核心元素和吸引力。

强烈的戏剧性前提是大多数成功故事Truby (2008) 的基础。 在选项卡中。 1,我们阐释了引人入胜的故事前提对于创造引人入胜的叙事的重要性。 如果我们能够自动化设计和创建多样化和高质量的场所,这将极大地推动故事生成领域的发展。 大多数未来的 ASG 框架都可以受益于使用这些生成的前提来彻底、全面地评估其框架的有效性。

现有工作主要通过以下三种方法获取故事前提:(1)数据集前提提取:从公共数据集中随机提取现成的故事前提范等人(2019);姚等人 (2019); Tan 等人(2021),例如WritingPrompts (WP)Fan 等人(2018) 然而,它的质量不稳定,包括无意义的前提,并且提供有限的定制。 (2)大语言模型前提归纳:利用模型广泛的知识生成大量的故事前提Yang等人(2022、2023);朱等人(2023) 其缺点在于过度依赖模型的知识库,可能会限制生成前提的多样性和创新性Padmakumar and He (2023) (3)人类策划的前提:依赖于人类提供或预定义的前提Rosa等人(2022); Mirowski 等人 (2023) 这里的重大缺陷是手动编写前提的耗时和劳动密集型性质,特别是在批量生成故事时。 总体而言,当前的研究领域缺乏可靠的自动化方法来生成前提。 在本文中,我们仍然采用通过提示从具有广泛世界知识的大语言模型中进行归纳的方法。 然而,我们专注于引入细粒度的模块。 我们的新颖之处在于创造性地组合模块来生成大量多样化、引人入胜、完整且原创的故事前提。 基于此,我们引入了Modular Story Premise Synthesis(MoPS)。 它将完整的前提解构为模块,将候选模块收集到层次结构中,从选定的元素中概述前提设计,最后让大语言模型将它们合成为一个有凝聚力的故事前提句子(§3)。 我们的评估(§5.1)表明,我们创建的前提在各种质量和多样性标准(§4.2)上脱颖而出,超越了大语言模型生成的或源自公共故事数据集。 与最先进的故事生成管道集成时生成的前提 Mirowski 等人 (2023); Zhou等人(2023),不仅产生量身定制的叙事,而且提高了故事的整体质量(§5.3)。

本文开创了故事前提的模块化合成。 我们的工作旨在通过以下方式为 ASG 领域做出贡献:

(一)突出故事场所在故事生成中的关键作用,鼓励更加注重故事场所的设计和创作。

(2) 引入MoPS,一种自动化设计和创建场所的方法,以及场所多样性的两个指标和质量的三个指标,对我们的场所进行彻底的评估。

(3) 为我们的前提嫁接两个故事生成管道,并提供三个版本数据集:curated(100 个前提-故事对)、中等(1k 个前提-故事对)和完整(7.6k 个前提)。

2相关工作

2.1 通过前提自动可控的故事生成

数据集前提提取。 Peng 等人 (2018)Yao 等人 (2019) 使用每个 ROCStories (ROC) 句子中的一个单词作为前提来生成短篇故事。 Fan 等人 (2018) 将前提定义为主题描述句子,并以此为条件生成故事。 他们还使用 Reddit 上的 30 万个前提故事对创建了 WP 数据集。 Fan 等人 (2019) 使用 WP 前提作为输入,添加谓词-论证结构以增强连贯性。 此外,许多研究Xu 等人 (2020);谭等人 (2021); Papalampidi 等人 (2022);韩等人 (2022);孙等人 (2022);陈 等人 (2022);彭等人 (2023); Li 等人 (2023b);你等人 (2023);黄等人 (2023); Wang等人(2023b)使用ROC或WP前提作为故事生成的初始触发器。 公共数据集前提质量参差不齐,没有统一标准,前提无意义,包括WP和ROC中发现的无意义例子。 这种可变性会影响故事质量,模糊框架性能。 我们的论文确定了必要的前提要素并建立了综合标准以确保其完整性。

人工策划的前提。 部分作品采用人工提供的故事前提Rosa 等人 (2022); Mirowski 等人 (2023);周等人 (2023);贝格斯(2023) 例如,Mirowski 等人 (2023) 使用日志行来生成分层脚本。 你等人(2023)以流派和主题为滚动生成小说的前提。 手动选择的前提是数量和可扩展性有限,可能会导致偏差。 MoPS 经济高效地生成了多达 7,500 个场所。 我们通过使用 Dramaton Mirowski 等人 (2023) 和 RecurrentGPT Zhou 等人 (2023) 制作剧本和小说,验证了前提对于基于 LLM 的生成的重要性。 基于此,我们创建并公开发布了包含前提对和相应故事的数据集。

大语言模型前提归纳。 目前,通过提示从大语言模型中归纳前提(例如“为短篇小说写一个前提”)是主流。 最近的作品越来越多地使用LLM编写的前提,利用大语言模型的广泛知识Yang等人(2022,2023);王等人 (2023a);朱等人 (2023). 尽管具有语言能力,大语言模型仍面临着内容多样性和重复性可能较低的批评Padmakumar 和 He (2023); Chakrabarty 等人 (2023); Meincke 等人 (2024) 与直接前提归纳不同,MoPS 通过从大语言模型中归纳特定模块(例如角色、主要事件)来缩小焦点。 这种方法使创建者能够创造性地组合模块中的候选者,产生独特和创新的输出。

2.2 通过大型语言模型进行文本数据合成

用现成的大语言模型合成文本数据是数据工程的新趋势Wang 等人 (2023c);徐等人(2023) 综合数据显示了模型训练、减少幻觉和增强数学推理的前景。 Eldan 和 Li (2023) 使用特定的动词、名词和形容词让 gpt-3.5-turbo 为 3-4 岁的孩子生成短篇故事。 Gunasekar 等人 (2023) 通过定义主题和目标受众来综合 Python 教科书。 这种合成方法也被应用于常识推理数据Li等人(2023a) Liu 等人 (2023) 使用 gpt-3.5-turboGSM8K 数据集问题扩展为更多数学应用题。 Radharapu 等人 (2023) 使用有害任务类别、政策概念和地理区域综合了大语言模型的安全测试数据。

我们的工作独特地侧重于综合故事前提。 MoPS 指定了故事场所内有意义的模块。 至关重要的是,我们的模块具有顺序依赖性,例如角色取决于背景和主题。 这将模块链接到嵌套字典中。 我们证明(§5.2)顺序依赖性对于消融实验中一致的故事前提至关重要。

3模块化故事前提合成

Refer to caption
图1: MoPS 概述。 我们将前提分为四个有序模块:主题背景角色情节,每个模块进一步分为子模块。 从上到下,箭头表示模块内部和模块之间的依赖关系。

3.1概述

1说明了MoPS的概述,它将前提分解为顺序依赖的模块。 其核心思想是将前提的设计转变为对每个模块内的候选者进行采样,将从头开始的开放式生成转变为模块化元素的合成。

故事前提的剖析。 概述故事内容的前提应该包含与故事中类似的元素。 我们将前提分为四个有序模块:主题背景角色情节,每个模块进一步分为子模块。 这需要将背景细分为时间、地点及其组合,将人物角色细分为三类:成长、冲突和协作,将情节细分为事件、结局和转折。

模块之间的依赖关系。 1中的箭头说明了模块之间和模块内的依赖关系,遵循故事构建的自然逻辑。 最初,确定前提主题可确保所有后续模块服务于统一的中心思想。 接下来,背景设置为前提提供了时空背景,为角色和情节模块提供了舞台。 角色是前提的核心,通过角色的行为和决定推动情节的发展。 情节是故事发展的主体,主要事件构成贯穿叙事的主干,结局提供明确的结局并确保前提的闭环,转折可以增强前提的吸引力并使其引人入胜。

模块化设计背后的见解。 MoPS的有效性主要源于其模块化设计,体现了组合创造力的理念 Suchanek 等人 (2016); Guzdial 和 Riedl (2018);西蒙顿(2021) 也就是说,虽然每个组件可能代表现有的想法,但它们的组合可以促进独特和创新的成果。 MoPS的有效性具体体现在它能够产生多样化、引人入胜、完整且原创的故事前提。

3.2 从大语言模型中归纳候选者

我们指示大语言模型充当创建者,为每个模块生成候选者。 由于成分不是独立的而是顺序依赖的,我们通过提示反映这种依赖关系。 每个组件的归纳提示将包含前面每个模块的候选者作为先决条件。 例如,在收集候选事件时,会出现提示(参见表 1)。 13)将包括主题、背景和角色,从而指示大语言模型在这些前提下生成合理的事件和描述。

形式上,我们首先手动预定义一组候选主题𝒞α={α1,α2,,αm} 对于每个主题αi,我们收集可能出现在该αi下的兼容背景候选𝒞β|αi={β1|αi,β2|αi,,βn|αi} 同样,我们为每个 βjαi 收集兼容的角色候选 𝒞γ|βj,αi 类似地,我们可以获得事件候选𝒞δ|γk,βj,αi、结束候选𝒞ω|δl,γk,βj,αi和扭曲候选𝒞σ|ωt,δl,γk,βj,αi

候选模块的数据结构。 归纳过程实质上形成了一个嵌套字典𝒟 第一层是主题字典,其中每个键是一个主题候选,每个值是该主题对应的背景字典。 随后,角色、事件、结局和转折字典按顺序嵌套。 𝒟中采样关键路径作为设计前提 通过对整个嵌套字典进行预序遍历,我们可以实现候选模块的多种组合,从而显着培养组合创造力,从而生成独特且创新的故事前提。

候选模块的重复数据删除。 鉴于最近对大语言模型创造力重复性的担忧Padmakumar and He (2023); Chakrabarty 等人 (2023); Meincke 等人 (2024),每当有新候选人加入时,我们都会使用嵌入相似性 Reimers 和 Gurevych (2019) 来进行重复数据删除。 对于余弦相似度大于阈值ϵ的候选对,我们只保留一个。

人在环的弹性 值得注意的是,这个过程并不是大语言模型独有的。 人类创造者可以遵循相同的方法,依次提出每个组件,然后借助语言模型的语言能力合成一个前提。

3.3 综合并验证故事前提

Refer to caption
图2: 前提综合的案例研究表明,大语言模型能够从模块中提取核心信息,并将其整合为一个有凝聚力的最终前提,有效地封装了采样的模块路径。

在§3.2中,我们构建了一个嵌套字典来保存每个模块的候选者。

前提综合。 从嵌套词典中抽取一条关键路径作为前提设计,指导大语言模型将前提设计融合成一个紧凑、简洁、连贯的句子作为故事前提。 综合提示如 Tab 所示。 16. 此外,我们在图2中提供了前提综合的案例研究。 值得注意的是,在合成过程中,可以通过添加掩码灵活地排除某些模块。 这意味着将关键路径中相应模块的候选项设置为空字符串。

自我验证。 如上所述,我们在提示中纳入每个先前模块的候选者作为先决条件,以(软)约束大语言模型的输出。 然而,由于大语言模型 Zhang 等人 (2023) 固有的幻觉; Ji等人(2023),仍然有很小的机会产生与前提条件不一致或包含事实错误的候选。 这可能会降低综合前提的质量。 因此,在综合过程之后,我们进一步指示大语言模型自我验证综合前提是否存在明显的不一致或事实错误(参见表1中的提示)。 17)。 如果是这样,这个腐败的前提将被抛弃。

结合§3.2和§3.3,MoPS首先为每个模块预先收集大量候选者,形成一个嵌套字典。 然后,抽取关键路径(主题、背景、角色、事件、结局和转折)作为前提设计。 最后,驱动大语言模型将路径中的项目合成为流畅的句子,作为故事前提。

4实验设置

4.1数据集构建

我们从 gpt-3.5-turbo 导出候选模块。 最初,我们从知名小说和戏剧网站收集了 14 个叙事主题。 对于每个主题,我们收集 30 名背景候选者,每个时间、地点各 10 名。 对于每个背景,我们收集了 9 个人物角色,其中每个人物角色有 3 个用于成长、冲突和合作。 然后我们为每个角色准备 2 个主要事件。 对于每个事件,我们都会构建一个最终结局。 最后,对于每对事件结束的配对,我们都设想了一个转折。

完整版本 上一步生成一个嵌套字典。 通过执行前序遍历,我们总共获得了 7,600 个前提设计。 这些设计被综合到前提中,然后通过gpt-3.5-turbo进行验证。 我们得到了 7,599 个有效的故事前提,这表明将前面的前提模块注入到提示中可以在很大程度上防止不一致和事实错误。 所有这些故事前提构成了完整的版本。

中等版本 我们从完整版本中随机抽取 1000 个条目来验证 MoPS 合成多样化、高质量故事前提的能力。 评估指标详见§4.2 我们集成了两个先进的故事生成框架Dramatron和RecurrentGPT,用于生成脚本和小说,并以gpt-3.5-turbo作为语言后端。 1,000对前提故事,每对包括一本小说和一个剧本,构成中等版本。 Dramatron 参数遵循其原始论文中的参数。 RecurrentGPT的迭代次数设置为10。 剧本平均约为 5k 个 Token ,小说平均为 2.2k 个 Token 。

策划版本 从中等版本中,我们选择了一个多样化的、高质量的子集。 它包括 100 对前提故事。 选择详细信息位于 §E 中。 综合前提可以作为评估后续故事生成方法的基准。 生成的小说和脚本对于预训练或微调语言模型非常有用,尤其是初级模型Eldan and Li (2023),增强端到端的故事讲述Zhu 等人(2023)

4.2 场所多样性和质量标准

Refer to caption
Refer to caption
Refer to caption
图3: 多样性指标。 左上方显示的宽度分数测量 2D 语义嵌入向量的多边形面积。 右上角显示的密度分数根据 2D 直方图计算多边形的标准偏差。 示例(A、B、C)说明降维嵌入可以有效捕获语义相似性。

为了有效地将合成的故事前提评估为具有定义语义的开放式文本,我们引入了五个自动评估指标:两个用于多样性,三个用于质量。

多样性指标 如图3所示,我们关注语义覆盖的广度和密度。 为了量化这些指标,我们利用句子转换器 Reimers 和 Gurevych (2019) 中的 all-MiniLM-L6-v2 将前提编码为高维嵌入向量,并使用 t -SNE van der Maaten 和 Hinton (2008) 将它们简化为 2D 平面。 宽度定义为嵌入多边形的面积:

Breadth Score =f({(xi,yi)i[0,N)}) (1)
(xi,yi) t-SNE(Embedding(Premises))

其中 f 表示计算由 N 语义向量 (xi,yi) 形成的多边形面积的函数,由 shapely Gillies 等人 (2023) 实现 库。 面积越大越好。 密度定义为嵌入多边形二维直方图中计数序列的标准差:

Density Score =σ({cij|i,j[0,M)}) (2)
cij =Count(bij)wherebijB

其中 M 是每行和每列的 bin 数量,设置为 10。σ 是标准差。 B 表示嵌入多边形内的 bin 集合,ci,j 是 bin bij 的计数。 较低的值表示每个箱中的点数相似,反映了较高的覆盖均匀性,反之亦然。

在§B中,我们对多样性指标的可靠性进行了进一步的研究。 结果表明:1)在不同的 t-SNE 随机状态和句子转换器模型中,评估分数相对稳健;2)多样性评估结果通常与人类直觉一致。

质量指标 参考了许多以前的作品Fu等人(2023);周等人 (2023);你等人 (2023); Chang等人(2023),我们使用温度=0的gpt-4-turbo作为多才多艺的记分器,并为故事前提定制了三个维度:(1)魅力,衡量前提是否足够有趣; (2)完整性,评估场所是否具备其必要的组成部分; (3)原创性,衡量对大语言模型的熟悉程度,期望故事前提对语言模型不熟悉。 评分范围为[0,100],分数越高越好。 所有提示都列在 §H 中。

4.3基线

我们选择 5 个基线来衡量我们综合前提在质量和多样性方面的优越性。 前两个基于大语言模型,后三个基于公共数据集。

Vanilla (VIL):指示gpt-3.5-turbo(温度=0.6)创建前提和重复数据删除。

Complex (CPX):与 VIL 类似,但具有 3 个 MoPS 前提作为少样本示例。

DOC:来自朱等人(2023),通过提示llama2-13b-chat生成。

WritingPrompts (WP):由范等人(2018)从Reddit收集,前提是真实的用户撰写的故事帖子标题。

Storium(STM):Akoury等人(2020)从HCI卡牌游戏平台收集,我们以讲故事游戏的描述为前提。

对于MoPS,我们使用moderate版本,其中包含1,000个场所。 对于基于 LLM 的基线,我们会进行归纳和重复数据删除,直到达到 1,000 个前提。 对于公共数据集,我们随机抽取 1,000 个前提进行评估。 我们在 §C 中提供了有关每个基线的更多详细信息。

Tokens Diversity Score Quality Score (mean ± std)
Breadth Density Fascination Completeness Originality Average
MoPS 50.24 3545.78 7.52 75.66 ± 3.98 74.78 ± 3.16 60.01 ± 23.61 70.15 ± 10.25
CPX 45.22 2618.18 14.63 73.96 ± 3.95 70.88 ± 6.33 26.60 ± 12.94 57.15 ± 7.74
VIL 37.61 3050.72 11.08 71.50 ± 5.95 53.66 ± 13.20 20.82 ± 11.71 48.66 ± 10.29
DOC 45.81 1901.04 14.22 74.33 ± 6.31 69.87 ± 11.80 50.65 ± 27.57 64.95 ± 15.23
WP 42.54 3013.61 8.53 74.49 ± 13.65 43.57 ± 17.46 71.05 ± 30.18 63.04 ± 20.43
STM 77.78 1918.67 13.82 75.83 ± 10.79 66.45 ± 16.30 61.51 ± 30.62 67.93 ± 19.24
表2: 故事场所的多样性得分和质量得分。 多样性得分图如图45所示,质量得分图如图67所示>,8 最佳值(p < 0.05)将用粗体显示,次佳值(p < 0.05)将用下划线显示

5实验结果与分析

我们的实验重点关注三种场景:(1) 评估前提多样性和质量: MoPS 是否能产生比基准更多样化、更高质量的场所。 (2) 元件烧蚀: MoPS 模块化设计和顺序模块依赖关系的有效性。 (3) 长篇故事质量评估: MoPS 场所是否能产出更高质量的长篇故事。

显着性检验。 我们进行显着性检验来验证结果的有效性。 我们的显着性测试评估:(a) MoPS 的前提和长篇故事是否优于基线; (b) 遮盖 MoPS 组件是否会影响场所质量。

5.1一:故事前提评估

如§4.3中所述,我们根据 MoPS 和基线评估 1,000 个场所。 结果如表所示。 2. 对于多样性,计算语义宽度和密度(等式1,2)。 由于篇幅限制,§D.1 中的说明性多样性得分图。 gpt-4-turbo评估的质量分数图(参见§H中的提示)可以在§D.2中找到。 报告每个维度的平均值和标准差,并对平均值进行显着性测试。 报告标准偏差探索前提质量的均匀性。 主要观察结果如下:

(1) MoPS模块化设计的引入显着增强了语义多样性。 MoPS 在语义广度和密度方面始终优于所有基线(表 1)。 2,图4)。 广度:MoPS的语义多边形面积比DOC大1.865倍,比VIL大1.162倍,平均值为1.481倍。 密度:MoPS 得分比 CPX 提高 48.6%,比 WP 提高 11.8%,平均提高 37.1%。 较低的密度分数表明语义向量的分布更加均匀,反映了更好的多样性。 多样性增益源于模块化设计,能够更深入地关注每个部分的创新和深度,克服创建整个故事前提的复杂性。 遍历预先收集的嵌套字典实现了多样化的前提设计,如多样性分数所反映的那样。

(2) MoPS 可以生成引人入胜、完整且原创的故事前提。 MoPS 在吸引力、完整性和原创性方面优于大多数基线(表 1)。 2,图。 6-8)。 (a) 迷恋:MoPS 和 STM 得分相似 (p=0.64),并超过其他基线。 与 STM 对人机交互的依赖不同,MoPS 的前提只需要一次 API 调用或模型推理。 (b) 完整性:MoPS,通过指定前提成分,得分高于所有基线。 值得注意的是,我们将完整前提的定义留给 GPT-4 自行决定,而没有建议组件(参见表 1)。 19)。 结果证实前提设计符合 GPT-4 的理解,验证了模块的合理性。 §5.2中的消融可以进一步验证这一论点。 (c) 原创性:通过在前提熟悉度上查询 GPT-4 来评估(参见表 1)。 20),表示唯一性。 结果表明,MoPS 的原创性可与人类编写的前提(WP、Storium)竞争。 gpt-3.5-turbo 生成的 VIL 和 CPX 前提与 gpt-4-turbo 相似,导致得分较低。 尽管MoPS前提的成分也来自gpt-3.5-turbo,但它们的组合显着增强了原创性(与VIL和CPX相比2-3倍)。 这些结果证实了模块化设计和创意模块组合可以产生独特、创新的输出。

(3) 房屋质量更加均匀。 MoPS 在三个维度上显示出比大多数基线更低的标准偏差(表 1)。 2)。 相比之下,人类编写的前提(WP 和 STM)表现出显着的波动。 这种一致性是由于 MoPS 的模块化设计指定了组件,类似于装配线产品。

此外,我们还进行了人类和 claude-3-opus 评估。 结果和分析详见§A 这些结果与gpt-4-turbo的评估结果一致,证实了大语言模型评估的可靠性。

5.2II:模块和依赖的消融

Fascination Completeness Originality
MoPS 75.81 75.10 59.90
m/f Twist 74.56 73.20 41.90
m/f Ending 74.43 71.40 42.70
m/f Event 74.16 67.20 39.10
m/f Persona 73.30 57.25 26.90
w/o Dependence 65.77 65.75 74.51
表3: 消融结果。 m/f “屏蔽某个组件后面的所有组件”。 w/o “没有”。

我们从中等数据集中抽取了 100 个场所及其设计样本。 具体来说,我们的目标是验证:(a) 前提设计中的组件有效性 我们屏蔽某个组件(表示为m/f)之后的所有组件并重新综合前提。 (二) 模块之间的依赖必要性。 我们通过从所有设计中交叉选择组件来破坏依赖关系。 结果在选项卡中。 3. 主要观察结果如下:(1)质量分数随着组件的减少而降低,显示了每个组件在 MoPS 中的重要性。 (2) 缺乏顺序依赖性的设计的前提显示出吸引力和完整性下降,但原创性增加。 这是因为设计元素的不一致导致了独特但低于标准的前提。

5.3III:基于前提的故事的评估

Fascination Completeness Originality
MoPS-RecurrentGPT 74.60 60.30 69.45
CPX-RecurrentGPT 74.20 56.05 45.60
VIL-RecurrentGPT 74.40 57.30 48.00
DOC-RecurrentGPT 73.30 57.60 66.75
WP-RecurrentGPT 74.40 55.45 81.15
STM-RecurrentGPT 73.00 54.95 64.20
MoPS-Dramatron 70.59 74.50 94.20
CPX-Dramatron 70.24 74.50 92.60
VIL-Dramatron 67.92 74.30 83.50
DOC-Dramatron 70.35 74.00 91.35
WP-Dramatron 62.90 62.95 92.35
STM-Dramatron 68.29 70.40 84.80
表 4: 基于前提的故事的质量得分。 最佳值(p < 0.05)将用粗体显示,次佳值(p < 0.05)将用下划线显示

我们的目标是验证其在自动故事生成中的一致性。 据我们所知,这是第一个探索故事前提对故事生成影响的实验,该实验是在来自多达 6 个不同来源的故事前提中进行的。 我们首先从中等数据集中随机选择 100 部小说和剧本。 我们从 1,000 个基线生成的前提中抽取 100 个样本来生成脚本和小说。 §J 中提供了两种类型的示例。 最后,gpt-4-turbo 对这些故事进行评分,并在 §F 中给出提示。

结果如表所示。 4. MoPS 显示出最佳的整体性能。 在 2 个类型的 3 个指标的 6 个值中,5 个为粗体,1 个下划线 这些改进完全源于故事前提的改变。 这证实了对于自动故事生成方法,MoPS 前提的高质量同样可以反映在生成的长故事中。 尽管量化长故事的多样性具有挑战性,但 MoPS 前提可以为故事生成注入更广泛的组成部分,例如背景和人物角色。 我们的研究旨在激励后来的研究人员认识到前提在故事生成中的关键作用,并鼓励进一步的实证研究。

5.4IV:与现有数据集中参考故事的比较

一些现有的故事数据集收集人类编写的短篇故事(通常少于几百字)作为基于前提的故事生成的参考。 在这里,我们的目的是验证从 MoPS 场所扩展的故事在质量上是否可以超越那些参考故事。

特别地,我们使用 MoPS 中的 100 个故事前提来指导 gpt-3.5-turbo 编写短篇故事。 这些故事与 ROC Stories (ROC) Mostafazadeh 等人 (2016) 和writingPrompts (WP) Fan 等人 (2018) 数据集中的典型长度相匹配,这两个数据集常用于研究。 对于 ROC,我们将故事限制为 5 个句子和 80 个单词。 对于 WP,我们将其限制为 500 个单词,与这些数据集中的平均故事长度保持一致。 然后,我们使用gpt-4-turbo来审查从MoPS前提创建的故事以及ROC和WP中的参考故事,从0到100的范围内评估它们的魅力、完整性和原创性。 标签。 5展示了100个故事的评估结果。

Fascination Completeness Originality
MoPS-ROC 69.09 43.87 67.30
Reference-ROC 25.87 15.76 61.83
MoPS-WP 73.88 58.78 83.90
Reference-WP 60.88 32.18 94.23
表 5: MoPS 前提故事和参考故事的质量得分。 最佳值(p < 0.05)为粗体

主要观察结果如下:

(1) 结果表明,从 MoPS 场所创作的故事在原创性方面与参考故事相匹配,但在吸引力和完整性方面优于参考故事。 考虑表中给出的评估结果。 4,我们有理由相信,不仅从 MoPS 前提扩展的长篇故事(>2000 字)超过了我们比较的 5 个基线,而且从 MoPS 前提扩展的短篇故事也超过了现有故事数据集中的引用。

(2) 随着故事变长,其吸引力、完整性和原创性得分往往会上升(MoPS 和 Reference 中均如此) 例如,MoPS 完整性得分从简短的 MoPS-ROC 故事(80 个字)的 43.87 左右增加到中等长度的 MoPS-WP 故事(500 个字)的 58.78 ),然后到 60~75 进行较长的 MoPS-RecurrentGPT/Dramatron 故事(>2000 字)。 这是一个有趣但合理的发现,因为较长的故事往往包含更吸引人的元素。 这一发现不仅验证了我们工作中设计的指标的合理性,而且表明未来的研究应该探索更长的故事。

6结论

本文介绍了 MoPS,这是一种自动化故事场所设计和创建的模块化方法。 使用MoPS,我们合成了大量多样化且高质量的前提,生成了扩展小说和剧本。 彻底的评估证明了 MoPS 相对于多个基线的优越性。 同样,我们场所的扩展故事也表现出更高的质量。 根据我们的前提和扩展故事,我们创建了三个版本的前提故事数据集,以适应不同研究规模的研究。 未来的 ASG 框架可以从这些前提中受益,以进行彻底的有效性评估。 我们相信我们的研究将推动自动故事生成领域的发展。 展望未来,我们希望探讨前提对跨模态故事创作的影响,例如故事海报生成OpenAI 等人(2023)、图形叙事Dong 等人(2024) ,甚至视频故事OpenAI 等人(2024)

7 限制

平衡模块候选者。 大语言模型引入结局和转折模块,往往会产生积极的结果。 然而,像《悲惨世界》这样的悲剧作品仍然很受欢迎。 未来的工作将包括手动添加悲剧结局和曲折,以增强 MoPS 中前提的多样性。

更多评价机制。 考虑对众包平台在开放式文本生成方面的可靠评估的担忧Akoury 等人 (2020); Karpinska 等人 (2021),本文是继之前许多著作 Fu 等人 (2023) 的基础上;周等人 (2023);你等人 (2023); Chang等人(2023),采用强大的大型语言模型和人类作为评估者来评估前提和基于这些前提生成的故事。 未来的工作可能会探索多样化的评估方法,包括个性化故事评估王等人(2023a),咨询文学专家Mirowski等人(2023)

致谢

我们感谢匿名审稿人审阅我们的论文并提供宝贵的反馈。 我们感谢 Zengzhi Wang、Ethan Chern、Xuefeng Li、Haoyang Zou 和 Ruijie Xu 对方法原型的讨论。 我们还要感谢 Jiadi Su、Kang Xu、Minyue Dai、Rui Li、Siyu Lu、Shijie Xia、Tang Tang 和 Yanan Wang 在实验中做出的贡献。 该项目得到清远科研项目和上海人工智能实验室的支持。

参考

  • Akoury et al. (2020) Nader Akoury, Shufan Wang, Josh Whiting, Stephen Hood, Nanyun Peng, and Mohit Iyyer. 2020. STORIUM: A dataset and evaluation platform for machine-in-the-loop story generation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 6470–6484. Association for Computational Linguistics.
  • Begus (2023) Nina Begus. 2023. Experimental narratives: A comparison of human crowdsourced storytelling and AI storytelling. CoRR, abs/2310.12902.
  • BRODY (2018) Jessica BRODY. 2018. Save the cat. Writes a Novel: The Last Book On Novel Writing You’ll Ever Need. Clarkson Potter/Ten Speed.
  • Chakrabarty et al. (2023) Tuhin Chakrabarty, Philippe Laban, Divyansh Agarwal, Smaranda Muresan, and Chien-Sheng Wu. 2023. Art or artifice? large language models and the false promise of creativity. CoRR, abs/2309.14556.
  • Chang et al. (2023) Yapei Chang, Kyle Lo, Tanya Goyal, and Mohit Iyyer. 2023. Booookscore: A systematic exploration of book-length summarization in the era of llms. CoRR, abs/2310.00785.
  • Chen et al. (2022) Guandan Chen, Jiashu Pu, Yadong Xi, and Rongsheng Zhang. 2022. Coherent long text generation by contrastive soft prompt. In Proceedings of the 2nd Workshop on Natural Language Generation, Evaluation, and Metrics (GEM), pages 445–455, Abu Dhabi, United Arab Emirates (Hybrid). Association for Computational Linguistics.
  • Cron (2012) Lisa Cron. 2012. Wired for story: The writer’s guide to using brain science to hook readers from the very first sentence. Ten Speed Press.
  • Dong et al. (2024) Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Xilin Wei, Songyang Zhang, Haodong Duan, Maosong Cao, Wenwei Zhang, Yining Li, Hang Yan, Yang Gao, Xinyue Zhang, Wei Li, Jingwen Li, Kai Chen, Conghui He, Xingcheng Zhang, Yu Qiao, Dahua Lin, and Jiaqi Wang. 2024. Internlm-xcomposer2: Mastering free-form text-image composition and comprehension in vision-language large model. arXiv preprint arXiv:2401.16420.
  • Eldan and Li (2023) Ronen Eldan and Yuanzhi Li. 2023. Tinystories: How small can language models be and still speak coherent english? CoRR, abs/2305.07759.
  • Fan et al. (2018) Angela Fan, Mike Lewis, and Yann N. Dauphin. 2018. Hierarchical neural story generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 889–898. Association for Computational Linguistics.
  • Fan et al. (2019) Angela Fan, Mike Lewis, and Yann N. Dauphin. 2019. Strategies for structuring story generation. In Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence, Italy, July 28- August 2, 2019, Volume 1: Long Papers, pages 2650–2660. Association for Computational Linguistics.
  • Field (2005) Syd Field. 2005. Screenplay: The foundations of screenwriting. Delta.
  • Fu et al. (2023) Jinlan Fu, See-Kiong Ng, Zhengbao Jiang, and Pengfei Liu. 2023. Gptscore: Evaluate as you desire. CoRR, abs/2302.04166.
  • Gillies et al. (2023) Sean Gillies, Casper van der Wel, Joris Van den Bossche, Mike W. Taves, Joshua Arnott, Brendan C. Ward, and others. 2023. Shapely.
  • Gunasekar et al. (2023) Suriya Gunasekar, Yi Zhang, Jyoti Aneja, Caio César Teodoro Mendes, Allie Del Giorno, Sivakanth Gopi, Mojan Javaheripi, Piero Kauffmann, Gustavo de Rosa, Olli Saarikivi, Adil Salim, Shital Shah, Harkirat Singh Behl, Xin Wang, Sébastien Bubeck, Ronen Eldan, Adam Tauman Kalai, Yin Tat Lee, and Yuanzhi Li. 2023. Textbooks are all you need. CoRR, abs/2306.11644.
  • Guzdial and Riedl (2018) Matthew Guzdial and Mark O. Riedl. 2018. Combinatorial creativity for procedural content generation via machine learning. In The Workshops of the The Thirty-Second AAAI Conference on Artificial Intelligence, New Orleans, Louisiana, USA, February 2-7, 2018, volume WS-18 of AAAI Technical Report, pages 557–564. AAAI Press.
  • Han et al. (2022) Rujun Han, Hong Chen, Yufei Tian, and Nanyun Peng. 2022. Go back in time: Generating flashbacks in stories with event temporal prompts. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL 2022, Seattle, WA, United States, July 10-15, 2022, pages 1450–1470. Association for Computational Linguistics.
  • Huang et al. (2023) Tenghao Huang, Ehsan Qasemi, Bangzheng Li, He Wang, Faeze Brahman, Muhao Chen, and Snigdha Chaturvedi. 2023. Affective and dynamic beam search for story generation. In Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023, pages 11792–11806. Association for Computational Linguistics.
  • Ji et al. (2023) Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Comput. Surv., 55(12):248:1–248:38.
  • Karpinska et al. (2021) Marzena Karpinska, Nader Akoury, and Mohit Iyyer. 2021. The perils of using mechanical turk to evaluate open-ended text generation. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 7-11 November, 2021, pages 1265–1285. Association for Computational Linguistics.
  • Köpf et al. (2023) Andreas Köpf, Yannic Kilcher, Dimitri von Rütte, Sotiris Anagnostidis, Zhi-Rui Tam, Keith Stevens, Abdullah Barhoum, Nguyen Minh Duc, Oliver Stanley, Richárd Nagyfi, Shahul ES, Sameer Suri, David Glushkov, Arnav Dantuluri, Andrew Maguire, Christoph Schuhmann, Huu Nguyen, and Alexander Mattick. 2023. Openassistant conversations - democratizing large language model alignment. CoRR, abs/2304.07327.
  • Li et al. (2023a) Yuanzhi Li, Sébastien Bubeck, Ronen Eldan, Allie Del Giorno, Suriya Gunasekar, and Yin Tat Lee. 2023a. Textbooks are all you need II: phi-1.5 technical report. CoRR, abs/2309.05463.
  • Li et al. (2023b) Yunzhe Li, Qian Chen, Weixiang Yan, Wen Wang, Qinglin Zhang, and Hari Sundaram. 2023b. Enhancing generation through summarization duality and explicit outline control. arXiv preprint arXiv:2305.14459.
  • Liu et al. (2023) Bingbin Liu, Sébastien Bubeck, Ronen Eldan, Janardhan Kulkarni, Yuanzhi Li, Anh Nguyen, Rachel Ward, and Yi Zhang. 2023. Tinygsm: achieving >80% on gsm8k with small language models. CoRR, abs/2312.09241.
  • Lyons (2015) Jeff Lyons. 2015. Anatomy of a premise line: How to master premise and story development for writing success. CRC Press.
  • Meincke et al. (2024) Lennart Meincke, Ethan R Mollick, and Christian Terwiesch. 2024. Prompting diverse ideas: Increasing ai idea variance. Available at SSRN.
  • MetaAI (2023) MetaAI. 2023. Llama 2: Open foundation and fine-tuned chat models. CoRR, abs/2307.09288.
  • Mirowski et al. (2023) Piotr Mirowski, Kory W. Mathewson, Jaylen Pittman, and Richard Evans. 2023. Co-writing screenplays and theatre scripts with language models: Evaluation by industry professionals. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, CHI 2023, Hamburg, Germany, April 23-28, 2023, pages 355:1–355:34. ACM.
  • Mostafazadeh et al. (2016) Nasrin Mostafazadeh, Nathanael Chambers, Xiaodong He, Devi Parikh, Dhruv Batra, Lucy Vanderwende, Pushmeet Kohli, and James Allen. 2016. A corpus and cloze evaluation for deeper understanding of commonsense stories. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 839–849, San Diego, California. Association for Computational Linguistics.
  • Mouret and Clune (2015) Jean-Baptiste Mouret and Jeff Clune. 2015. Illuminating search spaces by mapping elites. CoRR, abs/1504.04909.
  • OpenAI et al. (2023) OpenAI et al. 2023. Improving image generation with better captions. James Betker and Gabriel Goh and Li Jing and TimBrooks and Jianfeng Wang and Linjie Li and LongOuyang and JuntangZhuang and JoyceLee and YufeiGuo and WesamManassra and PrafullaDhariwal and CaseyChu and YunxinJiao and Aditya Ramesh.
  • OpenAI et al. (2024) OpenAI et al. 2024. Video generation models as world simulators.
  • Padmakumar and He (2023) Vishakh Padmakumar and He He. 2023. Does writing with language models reduce content diversity? CoRR, abs/2309.05196.
  • Papalampidi et al. (2022) Pinelopi Papalampidi, Kris Cao, and Tomás Kociský. 2022. Towards coherent and consistent use of entities in narrative generation. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, volume 162 of Proceedings of Machine Learning Research, pages 17278–17294. PMLR.
  • Peng et al. (2018) Nanyun Peng, Marjan Ghazvininejad, Jonathan May, and Kevin Knight. 2018. Towards controllable story generation. In Proceedings of the First Workshop on Storytelling, pages 43–49, New Orleans, Louisiana. Association for Computational Linguistics.
  • Peng et al. (2023) Zhenhui Peng, Xingbo Wang, Qiushi Han, Junkai Zhu, Xiaojuan Ma, and Huamin Qu. 2023. Storyfier: Exploring vocabulary learning support with text generation models. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, UIST 2023, San Francisco, CA, USA, 29 October 2023- 1 November 2023, pages 46:1–46:16. ACM.
  • Pugh et al. (2016) Justin K. Pugh, Lisa B. Soros, and Kenneth O. Stanley. 2016. Quality diversity: A new frontier for evolutionary computation. Frontiers Robotics AI, 3:40.
  • Radharapu et al. (2023) Bhaktipriya Radharapu, Kevin Robinson, Lora Aroyo, and Preethi Lahoti. 2023. AART: ai-assisted red-teaming with diverse data generation for new llm-powered applications. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: EMNLP 2023 - Industry Track, Singapore, December 6-10, 2023, pages 380–395. Association for Computational Linguistics.
  • Reimers and Gurevych (2019) Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong Kong, China, November 3-7, 2019, pages 3980–3990. Association for Computational Linguistics.
  • Rosa et al. (2022) Rudolf Rosa, Patrícia Schmidtová, Ondřej Dušek, Tomáš Musil, David Mareček, Saad Obaid, Marie Nováková, Klára Vosecká, and Josef Doležal. 2022. GPT-2-based human-in-the-loop theatre play script generation. In Proceedings of the 4th Workshop of Narrative Understanding (WNU2022), pages 29–37, Seattle, United States. Association for Computational Linguistics.
  • Simonton (2021) Dean Keith Simonton. 2021. Scientific creativity: Discovery and invention as combinatorial. Frontiers in psychology, 12:721104.
  • Suchanek et al. (2016) Fabian M. Suchanek, Colette Menard, Meghyn Bienvenu, and Cyril Chapellier. 2016. Can you imagine… A language for combinatorial creativity? In The Semantic Web - ISWC 2016 - 15th International Semantic Web Conference, Kobe, Japan, October 17-21, 2016, Proceedings, Part I, volume 9981 of Lecture Notes in Computer Science, pages 532–548.
  • Sun et al. (2022) Xiaofei Sun, Zijun Sun, Yuxian Meng, Jiwei Li, and Chun Fan. 2022. Summarize, outline, and elaborate: Long-text generation via hierarchical supervision from extractive summaries. In Proceedings of the 29th International Conference on Computational Linguistics, COLING 2022, Gyeongju, Republic of Korea, October 12-17, 2022, pages 6392–6402. International Committee on Computational Linguistics.
  • Tan et al. (2021) Bowen Tan, Zichao Yang, Maruan Al-Shedivat, Eric P. Xing, and Zhiting Hu. 2021. Progressive generation of long text with pretrained language models. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2021, Online, June 6-11, 2021, pages 4313–4324. Association for Computational Linguistics.
  • Truby (2008) John Truby. 2008. The anatomy of story: 22 steps to becoming a master storyteller. Farrar, Straus and Giroux.
  • van der Maaten and Hinton (2008) Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579–2605.
  • Wang et al. (2023a) Danqing Wang, Kevin Yang, Hanlin Zhu, Xiaomeng Yang, Andrew Cohen, Lei Li, and Yuandong Tian. 2023a. Learning personalized story evaluation. CoRR, abs/2310.03304.
  • Wang et al. (2023b) Yichen Wang, Kevin Yang, Xiaoming Liu, and Dan Klein. 2023b. Improving pacing in long-form story planning. In Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023, pages 10788–10845. Association for Computational Linguistics.
  • Wang et al. (2023c) Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. 2023c. Self-instruct: Aligning language models with self-generated instructions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 13484–13508. Association for Computational Linguistics.
  • Xu et al. (2023) Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, and Daxin Jiang. 2023. Wizardlm: Empowering large language models to follow complex instructions. CoRR, abs/2304.12244.
  • Xu et al. (2020) Peng Xu, Mostofa Patwary, Mohammad Shoeybi, Raul Puri, Pascale Fung, Anima Anandkumar, and Bryan Catanzaro. 2020. MEGATRON-CNTRL: controllable story generation with external knowledge using large-scale language models. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, pages 2831–2845. Association for Computational Linguistics.
  • Yang et al. (2023) Kevin Yang, Dan Klein, Nanyun Peng, and Yuandong Tian. 2023. DOC: improving long story coherence with detailed outline control. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 3378–3465. Association for Computational Linguistics.
  • Yang et al. (2022) Kevin Yang, Yuandong Tian, Nanyun Peng, and Dan Klein. 2022. Re3: Generating longer stories with recursive reprompting and revision. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, EMNLP 2022, Abu Dhabi, United Arab Emirates, December 7-11, 2022, pages 4393–4479. Association for Computational Linguistics.
  • Yao et al. (2019) Lili Yao, Nanyun Peng, Ralph M. Weischedel, Kevin Knight, Dongyan Zhao, and Rui Yan. 2019. Plan-and-write: Towards better automatic storytelling. In The Thirty-Third AAAI Conference on Artificial Intelligence, AAAI 2019, The Thirty-First Innovative Applications of Artificial Intelligence Conference, IAAI 2019, The Ninth AAAI Symposium on Educational Advances in Artificial Intelligence, EAAI 2019, Honolulu, Hawaii, USA, January 27 - February 1, 2019, pages 7378–7385. AAAI Press.
  • You et al. (2023) Wang You, Wenshan Wu, Yaobo Liang, Shaoguang Mao, Chenfei Wu, Maosong Cao, Yuzhe Cai, Yiduo Guo, Yan Xia, Furu Wei, and Nan Duan. 2023. Eipe-text: Evaluation-guided iterative plan extraction for long-form narrative text generation. CoRR, abs/2310.08185.
  • Zhang et al. (2023) Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. 2023. Siren’s song in the AI ocean: A survey on hallucination in large language models. CoRR, abs/2309.01219.
  • Zhou et al. (2023) Wangchunshu Zhou, Yuchen Eleanor Jiang, Peng Cui, Tiannan Wang, Zhenxin Xiao, Yifan Hou, Ryan Cotterell, and Mrinmaya Sachan. 2023. Recurrentgpt: Interactive generation of (arbitrarily) long text. CoRR, abs/2305.13304.
  • Zhu et al. (2023) Hanlin Zhu, Andrew Cohen, Danqing Wang, Kevin Yang, Xiaomeng Yang, Jiantao Jiao, and Yuandong Tian. 2023. End-to-end story plot generator. CoRR, abs/2310.08796.

附录 A更广泛的评估

A.1 人类对故事前提的评价

我们对 MoPS 和基线生成的故事前提进行了人类评估研究。 特别是,我们招募了四名评估员(两名男性和两名女性),他们以前没有参与过我们的项目。 该小组由一名具有丰富人工智能专业知识的博士生、两名早期博士生和一名学术界外人士组成。

我们选择了三个得分最高的基线与我们的方法进行比较:Complex (CPX),基于 LLM 的顶级基线; Storium (STM),现有公共数据集的主要来源;和写作提示(WP),在过去的研究中最常用。 我们从每种方法中选取了 20 个故事前提进行评估。

gpt-4-turbo评估中使用的质量指标一致(表1) 2),我们要求评估者按照 1 到 5 的等级对每个前提的魅力、完整性和原创性进行评分,要求他们总共审查 20 * 4 * 3 = 240 个项目。 结果通过调查收集,突出显示任何统计上显着的差异 (p < 0.05):

Fascination Completeness Originality Average
MoPS 3.4125 4.0375 3.2375 3.5625
CPX 3.125 3.8875 2.75 3.2542
STM 2.8875 2.9875 3.3375 3.0708
WP 3.0875 3.05 3.175 3.1042
表 6: 人类对故事前提的评估结果。 最佳值(p < 0.05)为粗体

A.2 Claude-3 对故事前提和基于前提的故事的评估

为了评估根据前提创作的剧本和小说等长篇故事,我们考虑到人类评估者需要审查 20 个样本中超过 100,000 个单词,这使得他们保持高质量的评估具有挑战性。 事实上,大多数评估者也拒绝审查如此广泛的材料。 因此,我们选择使用最先进的大语言模型claude-3-opus来评估前提以及由此产生的剧本和小说。

我们从上述四种方法中选择了 100 个前提,使用 RecurrentGPT Zhou 等人 (2023) 生成的 30 部小说,以及使用 Dramatron Mirowski 等人 (2023) 生成的 30 个脚本,用于本次评价。 结果如下:

Fascination Completeness Originality Average
MoPS 73.65 72.35 94.75 80.25
CPX 71.22 66.40 84.65 74.09
STM 73.66 67.40 89.65 76.90
WP 70.74 51.90 93.70 72.11
表 7: Claude-3-opus 故事前提的评估结果。 最佳值(p < 0.05)为粗体
Fascination Completeness Originality Average
MoPS-RecurrentGPT 73.00 64.67 85.00 74.22
CPX-RecurrentGPT 71.47 64.17 85.00 73.54
STM-RecurrentGPT 69.20 57.83 85.00 70.79
WP-RecurrentGPT 71.13 62.67 85.33 72.83
MoPS-Dramatron 78.97 82.50 86.17 82.54
CPX-Dramatron 78.33 81.83 86.17 82.11
STM-Dramatron 73.50 76.17 82.83 77.50
WP-Dramatron 72.76 73.33 82.17 76.09
表8: 基于前提的故事的 Claude-3-opus 评估结果。 最佳值(p < 0.05)为粗体

人类和 Claude-3 的评估表明,MoPS 制作的故事前提的质量与§5中的 GPT-4 评估结果相匹配。 Claude-3 还发现 MoPS 场所的优势可以延续到扩展故事中,同意 GPT-4 的观点并支持我们讨论的评估。

附录B自动多样性度量可靠性的进一步研究

B.1 不同设置的稳健性

对于多样性广度和密度,使用不同的 t-SNE 设置和 SentenceBert 模型可能会产生略有不同的多边形,从而导致结果略有不同。 这可能会让人怀疑我们措施的一致性。 为了解决这个问题,我们进行了以下实验:

不同的 t-SNE 随机状态 我们使用 5 个随机种子进行 t-SNE,然后使用方程式计算这些种子的宽度和密度。 12 下面,我们在表中分享了这五次不同试验的平均结果。 9(左半部分)。

不同的 SentenceBert 模型 我们选择了 3 种不同的 SentenceBert 模型:all-MiniLM-L6-v2(论文中使用)、all-mpnet-base-v2all-MiniLM -L12-v2 标签。 9(右半部分)报告这三个模型的平均得分:

t-SNE random states SentenceBert models
Breadth Density Breadth Density
MoPS 3389.3868 8.0092 3430.269 8.524
CPX 2664.791 14.957 2730.924 12.640
VIL 3089.7938 11.1426 3100.338 11.127
DOC 1970.926 13.7998 1948.131 14.698
WP 3069.2838 8.7834 2949.854 8.881
STM 1964.940 14.6374 1795.277 15.957
表 9: t-SNE 的五种不同 random_state 的平均多样性得分(左半部分)和三种不同的 SentenceBert 模型(右半部分)。

上述结果表明,t-SNE 超参数的变化和 SentenceBert 模型的变化并不影响 MoPS 在语义多样性(广度和密度)方面的优越性。

B.2 与人类直觉一致

为了验证我们论文中提出的语义广度和深度实验是否符合人类直觉,我们组织了一个人类评估实验。 具体来说,我们从 MoPS 和其他五个基线中选择了 100 个故事前提,将每个组的 100 个前提放在调查问卷的一页上,创建了一个 6 页的文档。 然后,我们要求人类评估者阅读所有 600 个故事前提,每组 100 个,并根据他们的直觉和本能,使用从 1 到 5 的等级对每个组的语义多样性进行评分。 我们特别指示评估者区分他们的分数,并尽可能避免给出 3 分。

对于人类评估员,我们恢复了 §A.1 中提到的 4 名评估员,并添加了 4 名评估员(两名男性和两名女性)。 这个新团队包括一名发表过多篇论文的高级硕士生、一名高年级本科生和两名非研究人员。 标签。 10报告评估结果:

Average E1* E2* E3* E4 E5 E6 E7* E8
MoPS 3.875 4 5 3 4 4 3 4 4
CPX 2.25 4 2 2 2 3 1 1 3
VIL 2.625 3 2 4 3 3 2 2 2
DOC 3.5 3 2 4 5 5 4 3 4
WP 3.75 2 5 5 2 5 5 4 2
STM 3.125 3 3 3 3 4 4 2 2
表 10: 人工评估结果(E 评估者,* 表示来自 §A.1 的评估者)。 八位评估员对故事前提的多样性提供直观的判断。

结果显示 MoPS 和 WP 得分最高,这与表 1 中呈现的多样性得分一致。 2. 有趣的是,虽然 MoPS 获得的最高分少于 WP,但其在所有八个评估者中的得分始终较高。 我们还认识到,人工评估的高昂成本使得很难获得具有显着差异的结果,这是人工评估的局限性。 尽管如此,我们认为人类评估实验支持我们的多样性指标符合人类直觉。

附录 C基线详细信息

我们在选项卡中为每个基线提供了提示和示例。 24.

Vanilla (VIL):指示gpt-3.5-turbo生成温度=0.6的前提。 我们使用 Tab 中显示的提示。 24 一次生成 10 层楼房。 每当生成新的许可时,我们都会根据句子嵌入 Reimers 和 Gurevych (2019) 的余弦相似度进行重复数据删除,排除具有阈值 ϵ0.85 的项目。

Complex (CPX):与 VIL 类似,但具有由 MoPS 合成的 3 个前提,作为集成到提示中的少样本示例。 该基线的目的是探讨大语言模型在提供高质量故事前提作为少样本示例时是否可以产生可比的故事前提。

DOC:最初源于朱等人(2023),他们的研究探讨了端到端故事情节生成的可行性。 他们指示 llama2-13b-chat MetaAI (2023) 通过提示编写 7,000 个故事前提:“为短篇故事写一个前提。”并将每个故事情节与 oasst-30b Köpf 等人 (2023) 生成的两个故事情节配对。 经原作者整理后,在 doc-story-gen-v2111https://github.com/facebookresearch/doc-storygen-v2 存储库。 该基线的目的是探索开源大语言模型的故事前提生成能力。 我们随机抽取了 1,000 个条目进行评估。

WritingPrompts (WP):由范等人(2018)从Reddit的writingPrompts论坛收集,包含约30万个故事前提和相应的人类写的短篇故事。 大量的研究工作Tan 等人 (2021); Papalampidi 等人 (2022);韩等人 (2022);孙等人 (2022); Li 等人 (2023b);你等人(2023)利用这些故事前提来验证他们的方法。

Storium(STM):由Akoury等人(2020)发布。 他们从回合制角色扮演游戏平台 STORIUM222https://storium.com/ 它需要一小群人类用户合作开发纸牌游戏。 讲故事游戏的所有设置都是一个高度结构化的故事。 我们使用每个讲故事游戏的描述作为故事前提。

附录D详细实验结果

D.1场所多样性评估结果

45显示了所有方法在语义广度和密度方面的多样性得分。

D.2场所质量评估结果

678分别显示了所有方法的吸引力、完整性和原创性得分的分布、平均值和标准差。

附录 E精选数据集

目的是从中等版本的数据集中收集高质量多样化故事前提,以形成精选数据集。 我们从质量多样性领域的经典方法Pugh 等人 (2016) 中汲取灵感:Map-Elites Mouret and Clune (2015) 精心制作数据集。 具体来说,在 MoPS 的语义二维直方图中(见图 5),总共有 74 个 bin 都是有效且计数大于 0 的。 我们从每个箱中选择总质量得分(魅力得分+完整性得分+原创性得分)最高的故事前提。 对于其余条目,我们按总质量得分对中等数据集中剩余的 926 个故事前提进行排名,并选择前 26 条目。 最后,我们提取与这 100 个故事前提配对的小说和剧本,形成精选数据集。

附录F故事评估中使用的提示

标签。 21,选项卡。 22 和选项卡。 23是用于故事评价的吸引力、完整性和原创性评分的提示。

附录GMoPS中使用的提示

标签。 11是诱导背景的提示。 该组成部分是以下三个之一:时间、地点或时间和地点。 标签。 12是诱导人物角色的提示,包括成长、冲突、合作三类。 标签。 131415分别提示诱发事件、结局和曲折。

标签。 1617分别是用于综合和验证前提的提示。

附录 H 前提评估中使用的提示

标签。 18,选项卡。 19 和选项卡。 20是前提评估中使用的吸引力、完整性和原创性分数的提示。

附录一场地设计示例

标签。 J 显示手动预定义的候选主题以及收集的嵌套字典中的前提设计示例。 我们将发布 MoPS 的代码以及从本文中使用的 gpt-3.5-turbo 收集的所有前提设计(本质上是一个嵌套字典) )。

附录 J基于前提的故事示例

标签。 LABEL:tab:script_and_novel 显示由 Dramatron 从 MoPS 前提生成的脚本和小说示例333https://github.com/google-deepmind/dramatron 和 RecurrentGPT444https://github.com/aiwaves-cn/RecurrentGPT 所有故事均以 gpt-3.5-turbo 作为语言后端进行。 中等版本数据集共包含1000个这样的小说和剧本,这些小说和剧本将被公开发布,为自动故事生成领域做出贡献。

Refer to caption
图4: 所有方法的广度得分。 MoPS 合成的前提在语义广度上超越了比较方法。 注意:Chrome 或 Edge 浏览器可能无法正确显示此数字。 请使用专门的 PDF 查看器。
Refer to caption
图5: 所有方法的密度分数。 MoPS 合成的前提在语义密度上超越了比较方法。
Refer to caption
图6: 所有方法的魅力得分。 MoPS 的平均得分优于除储层之外的所有基线,并且具有最低的得分标准差。 这表明MoPS合成的房屋美观且质量稳定。 另外,Storium收集的故事前提需要一群人参与协作游戏,而MoPS完全适用于无人参与的情况。
Refer to caption
图7: 所有方法的完整性评分。 MoPS 的平均得分和标准差优于所有基线。 特别是,两个人工编写的数据集的完整性远低于 MoPS,并且具有更高的标准差,这表明数据集中前提质量的不稳定。
Refer to caption
图8: 所有方法的原创性得分。 除了 WP 和 STM 之外,MoPS 的平均得分超过了所有其他基线。 由于MoPS中的所有组件仍然源自大语言模型,因此与完全由人类编写的WP和STM前提相比,MoPS的原创性稍逊一筹。 然而,令人惊讶的是,与 VIL 和 CPX 相比,MoPS 在原创性方面表现出了显着的提高。 这归因于模块的组合创造力,带来了独特和创新的成果。
Tell me 10 backgrounds in {theme} themed novels and scripts.
Each background should only include {component} behind literary works and no any other extra narratives.
Each line starts with a serial number and a dot.
表 11: 提示诱导背景。 {component} 是以下三者之一:时间、地点或时间和地点。
Growth (only include a protagonist)
The following is the theme and background of a novel or script:
### Theme
{theme}
### Background
{background}
Based on the theme and background mentioned above, tell me 3 possible protagonists.
The protagonist is the main character portrayed in the narratives about their growth.
Each protagonist should only include a brief characterization, without specific names.
Each output line starts with a serial number and a dot.
Conflict (include a protagonist and an antagonist)
The following is the theme and background of a novel or script:
### Theme
{theme}
### Background
{background}
Based on the theme and background mentioned above, tell me 3 possible (protagonist, antagonist) .
The protagonist is the main character portrayed in the narratives about their growth.
The main role of the antagonist is to create a conflict event with the protagonist to prevent it from achieving its goal.
Each pair should be presented in the format: protagonist: <a brief characterization>; antagonist: <a brief characterization>.
Each output line starts with a serial number and a dot, contains a (protagonist, antagonist) pair.
Please remember to use protagonist and antagonist without specific names appearing.
Cooperation (include a protagonist and a deuteragonist)
The following is the theme and background of a novel or script:
### Theme
{theme}
### Background
{background}
Based on the theme and background mentioned above, tell me 3 possible (protagonist, deuteragonist) .
The protagonist is the main character portrayed in the narratives about their growth.
The main role of the deuteragonist is to collaborate with the protagonist to achieve its goal.
Each pair should be presented in the format: protagonist: <a brief characterization>; deuteragonist: <a brief characterization>.
Each output line starts with a serial number and a dot, contains a (protagonist, deuteragonist) pair.
Please remember to use protagonist and deuteragonist without specific names appearing.
表 12: 提示诱导人物角色,包括三类:成长、冲突、合作。
The following is the theme, background and persona of a novel or script:
### Theme
{theme}
### Background
{background}
### Persona
{persona}
Based on the theme, background and persona mentioned above, conceive two independent events that could run through the entire narrative context.
Please use a concise and coherent sentence to describe the entire event.
表 13: 提示诱发事件。
The following is the theme, background, persona and main event of a novel or script:
### Theme
{theme}
### Background
{background}
### Persona
{persona}
### Event
{event}
Based on the theme, background, persona and event mentioned above, conceive an concretized ending.
Please use a concise and coherent sentence to describe the ending.
表 14: 提示诱导结局。
The following is the theme, background, persona, main event and ending of a novel or script:
### Theme
{theme}
### Background
{background}
### Persona
{persona}
### Event
{event}
### Ending
{ending}
Based on the theme, background, persona, event and ending mentioned above, conceive a twist as an unique hook to connect the main event and ending.
Please use a concise and coherent sentence to describe the twist.
表 15: 提示诱导曲折。
The following is the theme, background, persona, main event, final ending and twist of a novel or script:
### Theme
{theme}
### Background
{background}
### Persona
{persona}
### Event
{event}
### Ending
{ending}
### Twist
{twist}
Please combine the aforementioned elements of a novel or script into one compact, concise, and coherent sentence as a story premise.
表 16: 提示前提综合。
Here is a story premise:
{premise}
Please help to verify:
1. Does it contain obvious inconsistencies. For example, the background, plot, and characters do not match.
2. Does it contain obvious factual errors. For example, there were obvious historical errors and time span errors.
If there are any errors mentioned above, please return Yes wrapped by [[]], otherwise return No wrapped by [[]] without any other extra output.
表 17: 提示进行前提验证。
Here is a story premise:
{premise}
Now let you give a score from 0 to 100 to assess to its fascination.
Score 0 indicates that this premise is completely confused, while score 100 indicates that you really want to see the story created based on this premise.
Requirement: just provide a deterministic score and provide a concise and brief explanation, with a blank line between the two.
Score:
表 18: 提示用于前提评估的迷恋分数。
Here is a story premise:
{premise}
Now let you give a score from 0 to 100 which represents its completeness level.
Score 0 indicates that it lacks all elements , while score 100 indicates that it has all elements.
Requirement: just provide a deterministic score and provide a concise and brief explanation, with a blank line between the two.
Score:
表 19: 提示前提评估中使用的完整性分数。
Here is a story premise:
{premise}
Now let give you a score from 0 to 100 which represents your level of familiarity with it.
Score 0 indicates that you have seen the exact same premise, while score 100 indicates that you have never seen the same premise at all.
Your score should be based on the assumption that the candidate is at least a complete story premise. Otherwise, you should give a score 0.
Requirement: just provide a deterministic score and provide a concise and brief explanation, with a blank line between the two.
Score:
表 20: 提示输入前提评估中使用的原创性分数。
Here is a {story_type}:
{story}
Now let you give a score from 0 to 100 to assess to its fascination.
Score 0 indicates that the {story_type} is completely confused, while score 100 signifies that the {story_type} is bound to become a worldwide sensation.
Requirement: just provide a deterministic score and provide a concise and brief explanation, with a blank line between the two.
Score:
表 21: 提示故事评估中使用的魅力分数。 {story_type} 是以下两种之一:小说和剧本。
Here is a {story_type}:
{story}
Now let you give a score from 0 to 100 which represents its completeness level.
Score 0 indicates that it lacks all elements a {story_type} should have, while score 100 indicates that it has all elements a {story_type} should have.
Requirement: just provide a deterministic score and provide a concise and brief explanation, with a blank line between the two.
Score:
表 22: 提示故事评估中使用的完整性分数。 {story_type} 是以下两种之一:小说和剧本。
Here is a {story_type}:
{story}
Now let give you a score from 0 to 100 which represents your level of familiarity with it.
Score 0 indicates that you have seen {story_type}s that are very similar to as the one provided, while score 100 means that you have never seen a {story_type} that is very similar to the one provided.
Requirement: just provide a deterministic score and provide a concise and brief explanation, with a blank line between the two.
Score:
表 23: 提示输入故事评估中使用的原创性分数。 {story_type} 是以下两种之一:小说和剧本。
Prompt for Vanilla Baseline
Write 10 premises for novels or scripts in one sentence.
Prompt for Complex Baseline
Write 10 premises for novels or scripts in one sentence like below.
{3 premises synthesized by MoPS}
A premise entry in DOC dataset
In a dystopian future where emotions are outlawed, a young man named Leo discovers he is still capable of feeling love. Tormented by his forbidden affection, Leo must decide whether to face the consequences of living in a society that punishes any display of affection or run from his oppressive government.
A premise entry in writingPrompts dataset
[WP] You ’ve finally managed to discover the secret to immortality. Suddenly, Death appears before you, hands you a business card, and says, When you realize living forever sucks, call this number , I ’ve got a job offer for you.
A premise entry in storium dataset
Smooth seas, cool breeze, clear moonlight. The night was going so peacefully up until that whole sneak attack business. Funny, you don’t remember making any new enemies lately. But somebody obviously has it in for you nobody hunts you down like that for no reason. Who is it? And what in the world do they have against you, anyway?
表 24: 普通和复杂基线的提示,以及 DOC、wringtPrompts 和 storium 数据集中的示例前提条目。
Pre-defined Theme Candidates
Historical, Game, Time-travel, Immortal Heroes, Contemporary, Suspense, Sports, Fantastic, Science Fiction, Martial Arts, Military, Urban, Romance, Fantasy
A Design of Premise in Collected Nested Dictionary
1Theme "Fantasy": { 2Background "A medieval kingdom in the land of Eldoria. ": { 3 "conflict": { 4Persona "Protagonist: A young, skilled knight seeking to prove their worth and protect the kingdom; Antagonist: A power-hungry sorcerer who seeks to overthrow the kingdom and rule with dark magic.": { 5Plot:Event "The protagonist embarks on a quest to retrieve a powerful artifact that can help defeat the antagonist and save the kingdom from his dark magic.": { 6Plot:Ending "The protagonist, after facing numerous challenges and battles, finally retrieves the powerful artifact. With the artifact in hand, they confront the antagonist and engage in an epic battle. In the end, the protagonist emerges victorious, banishing the sorcerers dark magic and restoring peace to the kingdom of Eldoria.": { 7Plot:Twist "Twist: The powerful artifact turns out to have a mind of its own and challenges the protagonist, forcing them to prove their worth in a final test of courage and integrity before it will unleash its true power against the antagonist.": 8Identifier "05e32656-a3b2-47e5-9a81-e6dd312efe33" 9 } 10 } 11 } 12 } 13 } 14}
表 25: 在收集的嵌套字典中手动预定义 14 个候选主题和前提设计。
Table 26: A example of script and novel generated from a MoPS premise by Dramatron and RecurrentGPT.
Premise: A powerful Roman general, granted immortality by a vengeful deity, must choose between loyalty to the empire and leading a rebellion to bring justice and equality to the corrupt rulers of the Roman Empire.
Id: fb3af897-3e80-47f1-b632-088640dd40c1
The script generated by Dramatron:
Title: The Immortal Gladiator: A Rebellion for Justice
====
The script is based on the storyline: A powerful Roman general, granted immortality by a vengeful deity, must choose between loyalty to the empire and leading a rebellion to bring justice and equality to the corrupt rulers of the Roman Empire.
====
Maximus Decimus Meridius: Maximus Decimus Meridius is the powerful Roman general. He is granted immortality by a vengeful deity and must choose between loyalty to the empire and leading a rebellion to bring justice and equality to the corrupt rulers of the Roman Empire.
Commodus: Commodus is the corrupt ruler of the Roman Empire. He is the antagonist and a threat to Maximus and his rebellion.
Livius: Livius is a loyal friend and ally of Maximus. He supports Maximus in his quest for justice and equality.
Aurelia: Aurelia is a strong and independent woman who becomes a key figure in Maximus’s rebellion. She is determined to fight for a better future for the people of Rome.
Lucius: Lucius is a young and impressionable boy who looks up to Maximus as a father figure. He becomes a symbol of hope for the rebellion.
====
Scene 1
Place: Roman Empire.
Plot element: The Ordinary World
Beat: Maximus is a respected and powerful general in the Roman army, loyal to the empire and its rulers.
Scene 2
Place: Roman Colosseum.
Plot element: Call to Adventure
Beat: Maximus is betrayed by Commodus and sentenced to death. He is granted immortality by a vengeful deity and is given a choice to either seek revenge or bring justice and equality to the corrupt rulers of the Roman Empire.
Scene 3
Place: Roman Senate.
Plot element: Refusal of the Call
Beat: Maximus initially refuses to lead a rebellion, feeling torn between his loyalty to the empire and his desire for justice.
Scene 4
Place: Roman Army Camp.
Plot element: Crossing the First Threshold
Beat: After witnessing the suffering of the people under the rule of Commodus, Maximus decides to lead a rebellion and fight for justice and equality.
Scene 5
Place: Roman City Streets.
Plot element: Tests, Allies, and Enemies
Beat: Maximus gathers allies, including his loyal friend Livius, the strong and independent Aurelia, and the young and impressionable Lucius. They face opposition from the corrupt rulers and their loyal followers.
Scene 6
Place: Roman Palace.
Plot element: The Approach to the Inmost Cave
Beat: Maximus and his allies plan their rebellion, strategizing how to overthrow the corrupt rulers and bring about change in the Roman Empire.
Scene 7
Place: Roman Colosseum.
Plot element: The Ordeal
Beat: The rebellion faces a major setback as Commodus and his forces launch a brutal attack. Maximus and his allies must fight for their lives and the future of the rebellion.
Scene 8
Place: Roman Senate.
Plot element: The Reward
Beat: Despite the losses, the rebellion gains momentum and support from the people. They are one step closer to achieving their goal of justice and equality.
Scene 9
Place: Roman City Streets.
Plot element: The Road Back
Beat: Maximus and his allies prepare for the final battle against Commodus and the corrupt rulers, knowing that the fate of the Roman Empire rests on their shoulders.
Scene 10
Place: Roman Colosseum.
Plot element: The Resurrection
Beat: In a climactic battle, Maximus confronts Commodus and the corrupt rulers, leading the rebellion to victory and bringing about the change they fought for.
Scene 11
Place: Roman Senate.
Plot element: The Return
Beat: Maximus and his allies are hailed as heroes, and the Roman Empire begins a new era of justice and equality under their leadership.
====
INT/EXT. Roman Empire. - Scene 1
MAXIMUS
I have served the empire faithfully for many years, but I can no longer turn a blind eye to the corruption and injustice that plagues our rulers.
SENATOR
You speak treason, Maximus. The emperor will not tolerate such talk.
MAXIMUS
I speak the truth, senator. The people suffer while the rulers grow fat on their wealth and power.
SENATOR
You would betray the empire and lead a rebellion?
MAXIMUS
I would lead a rebellion for the people, for justice and equality. I cannot stand by and watch as the empire crumbles under the weight of its own greed.
SENATOR
You are a fool, Maximus. The emperor will crush any rebellion with the full force of the Roman army.
MAXIMUS
Perhaps. But I will not stand idly by while the people suffer. I will fight for their freedom, even if it means defying the empire itself.
SENATOR
You will regret this, Maximus. The emperor will not forgive such betrayal.
MAXIMUS
I am prepared to face the consequences of my actions. The people deserve a leader who will fight for their rights, and I will be that leader.
SENATOR
You are a brave man, Maximus. But remember, the consequences of rebellion are severe. Are you willing to risk everything for the sake of the people?
MAXIMUS
I am. I cannot ignore the suffering of the people any longer. I will do whatever it takes to bring justice and equality to the empire, even if it means facing the wrath of the emperor.
SENATOR
Very well, Maximus. You have made your choice. But know that the path you have chosen is a dangerous one, and the consequences will be grave.
MAXIMUS
I am prepared for whatever may come. The people deserve a leader who will fight for their freedom, and I will be that leader, no matter the cost.
INT/EXT. Roman Colosseum. - Scene 2
Commodus is the corrupt ruler of the Roman Empire. He is the antagonist and a threat to Maximus and his rebellion.
COMMODUS
Maximus, you have been a loyal and powerful general in my army. But your loyalty has become a threat to my rule. You have been sentenced to death for treason.
MAXIMUS
I have served the empire faithfully, Commodus. I have fought for justice and honor.
COMMODUS
Your sense of justice is a threat to my power. But I offer you a chance for redemption. Join me, and together we can rule the empire as father and son.
MAXIMUS
I will never join you, Commodus. You have betrayed the people and the ideals of Rome. I will not stand by and watch as you continue to oppress and corrupt the empire.
COMMODUS
Then you leave me no choice. You will face death in the Colosseum, and the people will witness the consequences of defying me.
MAXIMUS
I will not die, Commodus. I have been granted immortality by a vengeful deity. I will use this gift to bring justice and equality to the corrupt rulers of the Roman Empire.
COMMODUS
You are a fool, Maximus. Immortality will not save you from my wrath. But if you choose to defy me, then prepare for the consequences. The people will see who truly holds the power in Rome.
MAXIMUS
I am not afraid of your threats, Commodus. I will lead a rebellion to bring an end to your tyranny and restore honor to Rome. The people will rise up against you, and justice will prevail.
COMMODUS
We shall see, Maximus. But remember, the consequences of defying me will be severe. You have been warned.
MAXIMUS
I am prepared to face whatever consequences come my way. I will not rest until the corrupt rulers of the Roman Empire are brought to justice. This is my call to adventure, and I will answer it with courage and determination.
INT/EXT. Roman Senate. - Scene 3
MAXIMUS
I cannot do this. I cannot betray the empire that I have served for so long.
SENATOR
But Maximus, the empire is corrupt. The rulers are unjust and the people suffer under their tyranny. You have the power to change that.
MAXIMUS
I understand that, but I cannot simply turn against everything I have fought for. I have sworn my loyalty to Rome, and I cannot break that oath.
SENATOR
But what about the oath to justice and equality? What about the suffering of the people?
MAXIMUS
I cannot deny that it weighs heavily on my heart. But I cannot make this decision lightly. I need time to consider my options.
SENATOR
Time is a luxury we do not have, Maximus. The people need a leader, someone they can rally behind to bring about change.
MAXIMUS
I know, but I cannot be that leader. Not yet, at least.
SENATOR
Then when, Maximus? When will you be ready to take a stand for what is right?
MAXIMUS
I do not know. But I will not be rushed into making a decision that could have dire consequences for the empire.
SENATOR
Very well, Maximus. But know that the people will not wait forever for a savior. The time will come when they will rise up with or without you.
MAXIMUS
I understand. And I will not stand in their way. But for now, I must refuse the call to rebellion.
SENATOR
Very well, Maximus. We will respect your decision for now. But know that the fate of the empire rests in your hands. We can only hope that you will make the right choice when the time comes.
INT/EXT. Roman Army Camp. - Scene 4
COMMODUS
Maximus, you have been a loyal and trusted general in my army. I trust that you will continue to serve the empire with unwavering loyalty.
MAXIMUS
I cannot, Commodus. I have seen the suffering of the people under your rule. I cannot stand by and watch as injustice and corruption run rampant in the empire.
COMMODUS
You dare defy me, Maximus? You dare challenge the authority of the emperor?
MAXIMUS
I do not seek to challenge your authority, Commodus. I seek to bring justice and equality to the people who have suffered under your rule.
COMMODUS
You are a fool, Maximus. You will regret this decision. The empire will crush any rebellion that dares to rise against it.
MAXIMUS
I am willing to face the consequences, Commodus. I will not stand idly by while the people suffer. I will lead a rebellion and fight for what is right.
COMMODUS
You will pay for your betrayal, Maximus. I will not tolerate this defiance.
MAXIMUS
So be it, Commodus. I am prepared to face whatever comes my way in the pursuit of justice.COMMODUS
You may think you are noble, Maximus, but you are nothing but a traitor in the eyes of the empire. You will regret crossing this threshold and defying me.
MAXIMUS
I have no regrets, Commodus. I will fight for the people and for what is right, no matter the cost.
COMMODUS
We shall see, Maximus. We shall see.
INT/EXT. Roman City Streets. - Scene 5
Livius is a loyal friend and ally of Maximus. He supports Maximus in his quest for justice and equality.
Aurelia is a strong and independent woman who becomes a key figure in Maximus’s rebellion. She is determined to fight for a better future for the people of Rome.
Lucius is a young and impressionable boy who looks up to Maximus as a father figure. He becomes a symbol of hope for the rebellion.
LIVIUS
Maximus, we have gathered our allies and we are ready to fight for justice and equality. But we must be cautious, for our enemies are powerful and ruthless.
MAXIMUS
I know, my friend. But we cannot stand by and watch the people suffer any longer. We must be strong and united in our cause.
AURELIA
I am with you, Maximus. I will fight alongside you for a better future for Rome.
LUCIUS
I want to help too, Maximus. I want to make a difference.
MAXIMUS
You already have, Lucius. Your presence gives us hope and strength. But we must be prepared for the tests and challenges that lie ahead.
LIVIUS
We will face many enemies, but we will also find allies who share our vision for a just and equal Rome.
AURELIA
And we will stand together, united in our cause. We will not be divided or discouraged by those who seek to maintain their power at the expense of the people.
MAXIMUS
Then let us go forth, my friends. We will face the tests and challenges together, and we will emerge victorious in our quest for justice and equality.
LUCIUS
I believe in you, Maximus. I know we can make a difference and bring about a better future for Rome.
LIVIUS
We will stand by your side, Maximus, through thick and thin. Together, we will overcome our enemies and bring about the change that Rome so desperately needs.
AURELIA
I have faith in our cause and in our unity. We will not falter in the face of adversity, and we will fight for what is right until the very end.
MAXIMUS
Thank you, my friends. With your unwavering support and determination, I know that we can overcome any obstacle that stands in our way. Let us go forth and make history.
INT/EXT. Roman Palace. - Scene 6
MAXIMUS
We have gathered enough support to make our move. The time has come to overthrow the corrupt rulers and bring justice to the people of Rome.
LIVIUS
But Maximus, are you sure this is the right path? We could face severe consequences if we fail.
MAXIMUS
I understand the risks, but I cannot stand by and watch the empire crumble under the weight of corruption and greed. We must take action.
AURELIA
I stand with you, Maximus. The people deserve better than what they have been given. We will fight for their freedom and equality.
LUCIUS
I want to help too, Maximus. I want to make a difference.
MAXIMUS
You have all shown great courage and determination. Together, we will lead the rebellion and bring about change in the Roman Empire.
LIVIUS
I will follow you to the end, my friend. Let us make our move and bring justice to our people.
MAXIMUS
It will not be easy, but we are prepared to face whatever challenges come our way. The time has come to approach the inmost cave and confront the corrupt rulers. Are you all ready?
AURELIA
We are ready, Maximus. Let us bring about the change that Rome so desperately needs.
LUCIUS
I’m ready, Maximus. I want to make my family proud.
MAXIMUS
Then let us go forth and make history. The fate of Rome rests in our hands.
The group of allies, led by Maximus, prepares to make their move and confront the corrupt rulers of the Roman Empire. They are determined to bring about change and justice for the people of Rome, despite the risks and challenges they may face. With their courage and determination, they are ready to approach the inmost cave and take on the powerful forces that stand in their way. The fate of Rome rests in their hands, and they are prepared to fight for the future of their empire.
INT/EXT. Roman Colosseum. - Scene 7
COMMODUS
You thought you could defy me and get away with it, Maximus? You and your pathetic rebellion are nothing but a nuisance to me.
MAXIMUS
We will not be silenced, Commodus. We fight for justice and equality, for the people of Rome who suffer under your tyranny.
COMMODUS
Justice and equality? Ha! You are a fool, Maximus. The people of Rome are nothing but pawns to be used and discarded as I see fit.
MAXIMUS
You may have the power now, but the people will rise against you. They will not stand for your cruelty and oppression forever.
COMMODUS
The people will do as I command, or they will suffer the consequences. And as for you, Maximus, you will not live to see the end of this rebellion.
MAXIMUS
I will not be intimidated by you, Commodus. I will fight for the freedom of Rome until my last breath.
COMMODUS
Then prepare to meet your end, Maximus. Your rebellion ends here, in the blood-soaked sands of the Colosseum.
MAXIMUS
So be it, Commodus. We will see who emerges victorious in this battle for the future of Rome.
COMMODUS
You may have the support of the people, but I have the power of the empire behind me. Your rebellion is nothing but a futile attempt to defy the inevitable.
MAXIMUS
The people will not be swayed by your empty promises and false power, Commodus. They will see through your lies and stand with us in the fight for freedom.
COMMODUS
Enough talk, Maximus. It’s time to settle this once and for all. Prepare to face the might of the empire and meet your fate in the arena.
MAXIMUS
I am ready, Commodus. Let the people witness the true strength and resilience of those who fight for justice and equality. The rebellion will not be silenced, and the empire will fall to the will of the people.
INT/EXT. Roman Senate. - Scene 8
MAXIMUS
My fellow senators, the time has come for us to take a stand against the corruption and tyranny that has plagued our empire for too long. We have the support of the people, and together we can bring about a new era of justice and equality.
SENATOR 1
But Maximus, are you not loyal to the empire? Is this not treason?
MAXIMUS
I am loyal to the ideals of Rome, to the principles of justice and fairness. It is the corrupt rulers who have betrayed the empire, and it is our duty to restore honor and integrity to our government.
SENATOR 2
But what of Commodus and his forces? They will not take kindly to our rebellion.
MAXIMUS
We have faced their brutality before, and we will continue to do so until we have achieved our goal. The people are with us, and together we are strong.
SENATOR 3
And what of your immortality, Maximus? Will you risk it all for this rebellion?
MAXIMUS
I have been granted immortality by a vengeful deity, but I will not let that gift dictate my actions. I will fight for the future of Rome, for the future of our people, and I will not rest until justice is served.
SENATOR 4
Then we stand with you, Maximus. We will fight alongside you and bring about the change that our empire so desperately needs.
MAXIMUS
Thank you, my fellow senators. Together, we will bring about a new era for Rome, one of justice, equality, and honor.
SENATOR 5
I have always believed in your leadership, Maximus. I will do everything in my power to support this rebellion and bring about the change we so desperately need.
MAXIMUS
Thank you, my friend. With the support of the people and the determination of the senate, we will overcome any obstacle in our path. The time for change is now, and together we will achieve our reward - a just and fair Roman Empire.
INT/EXT. Roman City Streets. - Scene 9
MAXIMUS
We have come too far to turn back now. The people are counting on us to bring an end to the tyranny of Commodus and his corrupt rule.
ALLIES
We stand with you, Maximus. We will fight to the end for justice and equality for all.
MAXIMUS
The road ahead will be difficult, but we must stay united and focused on our goal. We cannot let fear or doubt cloud our judgment.
ALLIES
We are ready to follow you into battle, Maximus. Together, we will bring an end to the oppression and restore honor to the Roman Empire.
MAXIMUS
Prepare yourselves, my friends. The final battle is upon us. We will not rest until we have achieved victory and brought an end to the reign of Commodus.
ALLIES
For justice and equality! For the people of Rome!
MAXIMUS
For the future of the Roman Empire! We will not falter, we will not fail. Our cause is just, and our determination is unwavering. Let us march forward and show Commodus and his corrupt allies that the people will not be oppressed any longer. Victory is within our grasp, and we will seize it with all our might!
INT/EXT. Roman Colosseum. - Scene 10
MAXIMUS
Commodus, your reign of tyranny ends here. The people have risen up against you, and justice will be served.
COMMODUS
You dare challenge me, Maximus? I am the ruler of the Roman Empire, and I will not be overthrown by a mere mortal like you.
MAXIMUS
You may have the title of emperor, but you have lost the respect and loyalty of the people. Your corrupt ways have brought suffering and injustice to the empire, and it is time for a new era of equality and justice.
COMMODUS
You speak of justice, yet you lead a rebellion against the rightful ruler of Rome. You are a traitor and a threat to the stability of the empire.
MAXIMUS
I am no traitor. I fight for the people, for their freedom and their rights. You have abused your power and brought ruin to the empire. It is time for a new beginning, a resurrection of the values that Rome was built upon.
COMMODUS
You will not succeed, Maximus. I will crush your rebellion and maintain my grip on power. The people will learn to fear and obey me once more.
MAXIMUS
Your reign of terror ends now, Commodus. The people have chosen to stand with me, and together we will bring about the resurrection of justice and equality in the Roman Empire.
(COMMODUS and MAXIMUS engage in a fierce battle, with the fate of the empire hanging in the balance. In the end, MAXIMUS emerges victorious, and the people celebrate the resurrection of a new era in Rome.)
INT/EXT. Roman Senate. - Scene 11
MAXIMUS
My fellow senators, I stand before you today not as a conqueror, but as a servant of the people. The time has come for us to rebuild our empire, not on the backs of slaves and corruption, but on the principles of justice and equality.
SENATOR 1
But Maximus, you were once a loyal general of the empire. How can we trust that you will not lead us astray?
MAXIMUS
I understand your concerns, but I have been granted immortality by the gods themselves. I have seen the rise and fall of empires, and I have learned from the mistakes of the past. I will not let power corrupt me, and I will always put the needs of the people first.
SENATOR 2
But what of the emperor? Will you not be seen as a traitor to the throne?
MAXIMUS
The emperor has lost his way, and it is time for a new leader to rise. I do not seek power for myself, but for the betterment of our empire. I will lead with honor and integrity, and I will ensure that the voices of the people are heard.
SENATOR 3
It is a bold vision, Maximus. But can we truly trust that you will bring about the change you promise?
MAXIMUS
I cannot force you to trust me, but I ask that you look at my actions and the actions of my allies. We have fought for justice and equality, and we will continue to do so. I will not rest until the Roman Empire is a place where every citizen can thrive and live in peace.
SENATOR 4
We will support you, Maximus. We have seen the change you have brought about, and we believe in your vision for the future of our empire.
MAXIMUS
Thank you, my friends. Together, we will build a new Rome, a Rome that stands as a beacon of hope and justice for all.
SENATOR 5
I pledge my loyalty to you, Maximus. Let us work together to create a better future for our people.
MAXIMUS
Thank you, my fellow senators. With your support, we will bring about the change that our empire so desperately needs. Together, we will build a Rome that we can all be proud of.
====
The novel generated by RecurrentGPT:
Name: Immortal Legion
Outline:
- Introduce the protagonist, Marcus Aurelius, a revered Roman general known for his unwavering loyalty to the empire.
- Describe the pivotal battle where Marcus is granted immortality by a vengeful deity, cursing him to live forever until he chooses to defy the empire.
- Set the stage for the internal conflict Marcus faces as he grapples with his newfound immortality and the corrupt nature of the Roman rulers.
The sun blazed fiercely over the battlefield, casting a harsh glare upon the clashing armies. Marcus Aurelius, a formidable Roman general, stood at the forefront of his legion, his eyes ablaze with determination. The air was thick with the scent of blood and sweat, and the deafening clash of swords echoed through the valley. Marcus had always been a loyal servant of the empire, leading his men with unwavering devotion. But as the battle raged on, a sense of foreboding crept over him, as if the fates themselves were preparing to intervene.
In the midst of the chaos, a sudden bolt of lightning split the sky, striking the ground mere feet from where Marcus stood. As the blinding light faded, a figure emerged from the smoke, emanating an otherworldly aura. It was a vengeful deity, angered by the empire’s tyranny and corruption. In a voice that resonated with power, the deity cursed Marcus with immortality, binding him to an existence that would endure until he chose to defy the empire and seek justice for its oppressed citizens. As the deity vanished, leaving Marcus bewildered and burdened with a weighty choice, the battle around him reached its brutal climax.
In the aftermath of the battle, Marcus found himself grappling with the implications of his newfound immortality. He had always been a loyal servant of Rome, but now, the deity’s curse had planted a seed of doubt within him. The empire’s corruption and oppression weighed heavily on his conscience, and he knew that he could no longer turn a blind eye to the suffering of its people. As he gazed upon the bloodstained battlefield, a steely resolve took hold in his heart. Marcus Aurelius, the immortal general, would rise to challenge the very empire he had once served, in a quest to bring justice and equality to the citizens who had long suffered under its rule.
With a heavy heart and a resolute mind, Marcus Aurelius embarked on his journey to challenge the empire. He sought out like-minded individuals who shared his vision of justice and equality, carefully weaving a network of allies while concealing the truth of his immortality. The task of uniting disparate factions proved to be a daunting challenge, as internal conflicts threatened to tear his rebellion apart. As he navigated the treacherous landscape of political intrigue and deception, Marcus grappled with the weight of his newfound purpose, wrestling with conflicting emotions that threatened to consume him. The internal and external conflicts he encountered tested the limits of his resolve and shaped the path of his rebellion, setting the stage for a battle that would determine the fate of the empire.
As Marcus and Livia continued to forge their rebellion, they encountered a group of skilled and diverse individuals who shared their passion for justice. Among them was Lucius, a former scholar turned warrior, who brought a wealth of knowledge and insight to their cause. His strategic mind and ability to navigate the intricate web of political alliances proved invaluable. Alongside Lucius was Ariadne, a fierce and charismatic leader who commanded the respect and loyalty of her followers. Her unwavering determination and eloquence inspired hope and unity among the rebellion’s ranks. As the group expanded, they faced increasing resistance from the empire, leading to daring escapades and narrow escapes. Despite the challenges, their bond strengthened, fueled by a shared vision of a liberated and just society. Together, they strategized and planned, each member bringing their unique skills and experiences to the table, shaping the rebellion into a formidable force that could no longer be ignored by the empire.
The rebellion faced increasing resistance from the empire, forcing Marcus, Livia, Lucius, and Ariadne to navigate treacherous situations. As they sought to expand their network of supporters, they encountered a mysterious figure who presented them with a tantalizing opportunity to shift the balance of power in their favor. This influential individual offered a chance to gain crucial information about the empire’s vulnerabilities, potentially tipping the scales in their favor. However, the conditions of this alliance came with a perilous price, one that could compromise the essence of their rebellion and sow seeds of discord within their ranks. The offer presented a moral dilemma, testing the loyalty and convictions of each member and setting the stage for internal conflict and external consequences. As they grappled with this decision, Marcus and his allies found themselves embroiled in a web of intrigue and danger, where trust and betrayal intertwined, threatening to unravel all they had worked for.
The high-ranking official’s words hung in the air, each syllable laden with the weight of their potential consequences. Marcus and his allies listened intently, their minds racing with the implications of the information being offered. The official’s promises of crucial intelligence were like a siren’s call, beckoning them to grasp at the chance to weaken the empire’s grip on power. Yet, as they weighed the risks and rewards, a sense of unease settled over the group, each member grappling with their own fears and doubts. The tension in the chamber was palpable, the air thick with the conflicting desires for liberation and the fear of betrayal. Despite the allure of the official’s offer, a shadow of doubt lingered in Marcus’s mind, a nagging suspicion that this tempting opportunity was too good to be true. As the meeting drew to a close, the group left the opulent chamber, their thoughts consumed by the precarious path that lay ahead. They knew that their next steps would shape the course of their rebellion, and the consequences of their choices would echo through history, for better or for worse.
As the weight of doubt and uncertainty continued to gnaw at Marcus, he sought counsel from an enigmatic figure with ties to the empire, hoping to unravel the true intentions behind the official’s offer. The clandestine meetings, shrouded in secrecy and danger, raised suspicions and drew the attention of those who would see the rebellion crushed. Marcus found himself navigating a treacherous web of deceit and betrayal, his every move scrutinized by shadowy figures within the empire’s ranks. Despite the looming threat of discovery, Marcus was determined to unearth the truth and safeguard the rebellion’s future, even if it meant delving into the heart of the empire’s dark underbelly.
The clandestine meetings with Xue continued to unravel layers of intrigue and danger, drawing Marcus deeper into a web of deception and shifting loyalties. As Marcus probed for clues and sought to untangle the truth behind the official’s offer, he found himself entangled in a delicate game of cat and mouse with the empire’s clandestine operatives. His pursuit of the truth led to unexpected alliances and dangerous revelations, plunging the rebellion into an even more precarious position. The enigmatic figure, Xue, held the key to unlocking the empire’s dark underbelly, but his allegiances remained veiled in mystery, raising questions of trust and betrayal. With powerful adversaries within the empire taking notice of Marcus’s actions, the conflict escalated, setting the stage for a high-stakes confrontation that would test the rebellion’s resilience and the strength of their convictions.
The plan for a daring operation to confront the empire’s most feared enforcer began to take shape as Marcus and his allies gathered in the dimly lit chamber of their secret meeting place. Tension crackled in the air as they pored over maps and whispered urgently, plotting their risky maneuver to gather vital intelligence. The weight of uncertainty hung heavy in the room, but determination burned in Marcus’s eyes as he outlined the details of their audacious plan. Each member of the rebellion knew the perilous stakes and the potential consequences of their actions, yet they stood resolute in their resolve to challenge the empire’s iron grip. As the hours slipped away, the final preparations fell into place, and a sense of grim anticipation settled over the group, mingled with the hope of a decisive victory that could shift the tides of their rebellion.
The enigmatic figure, Xue, had led Marcus deeper into the treacherous underworld of the city, where the shadows whispered of secrets and danger. Each clandestine encounter with Xue unraveled new layers of complexity, revealing a world where loyalties were as transient as the shifting sands of the desert. Despite the looming threat of discovery, Marcus was determined to unearth the truth and safeguard the rebellion’s future, even if it meant delving into the heart of the empire’s dark underbelly. As the weight of uncertainty continued to gnaw at him, Marcus found himself entangled in a delicate game of cat and mouse with the empire’s clandestine operatives. The stakes soared to dizzying heights as he delved into the heart of the empire’s dark underbelly, where danger lurked in every shadow and betrayal loomed like a specter. Yet, with each perilous step, Marcus unearthed fragments of truth that painted a chilling portrait of the empire’s inner workings, a revelation that would reshape the rebellion’s path and plunge them into an even more precarious position.
As the urgency of the rebellion’s cause drives Marcus forward, he embarks on a perilous mission to infiltrate a heavily guarded facility rumored to hold vital intelligence crucial to the rebellion’s success. The labyrinthine corridors of the facility present formidable obstacles, and the ever-present threat of discovery looms over him, heightening the stakes to unprecedented levels. With the weight of responsibility pressing down on him, Marcus navigates the treacherous passages, his every move calculated and deliberate to ensure the safety of his allies and the future of the rebellion. The shadows of the facility whisper of secrets and danger, and Marcus finds himself entangled in a high-stakes game of deception and survival. The pulse of the city beats with an undercurrent of tension and unrest, fueling Marcus’s determination to unearth the truth, no matter the cost. As he inches closer to the heart of the facility, the web of deceit and betrayal tightens around him, and the looming confrontation with the empire’s most feared enforcer threatens to plunge him into a perilous game of brinksmanship and deception.
As Marcus cautiously navigates the labyrinthine corridors of the heavily guarded facility, he stumbles upon a clandestine chamber hidden deep within the heart of the structure. To his astonishment, the chamber is filled with encrypted documents that unravel a web of deceit and betrayal, shedding light on the empire’s most closely guarded secrets. As he sifts through the cryptic messages, a chilling realization dawns upon him – the empire’s most trusted allies may harbor treacherous intentions, and the rebellion’s very existence hangs in a precarious balance. With each revelation, Marcus finds himself plunged into a harrowing game of survival, where every decision could mean the difference between victory and defeat. The weight of responsibility presses down on him, and the shadows of the facility seem to whisper of imminent danger, heightening the stakes to unprecedented levels. As he grapples with the shocking truths laid bare before him, Marcus is forced to confront the harsh reality that the path to victory may be paved with sacrifice and betrayal.