CHAE Fine-Grained Controllable Story Generation with Characters Actions and Emotions Xinpeng Wang1 Han Jiang1 Zhihua Wei1 Shanlin Zhou2_2

2025-04-30 0 0 690.42KB 12 页 10玖币

侵权投诉

CHAE: Fine-Grained Controllable Story Generation with Characters,

Actions and Emotions

Xinpeng Wang1, Han Jiang1, Zhihua Wei1∗

, Shanlin Zhou2

1Department of Computer Science and Technology, Tongji University, Shanghai, China

2School of Computer Science and Technology, Shanghai University of Electric Power, Shanghai, China

{wangxinpeng, 2230780, zhihua_wei}@tongji.edu.cn

zhoushanlin@mail.shiep.edu.cn

Abstract

Story generation has emerged as an interest-

ing yet challenging NLP task in recent years.

Some existing studies aim at generating ﬂu-

ent and coherent stories from keywords and

outlines; while others attempt to control the

global features of the story, such as emotion,

style and topic. However, these works focus

on coarse-grained control on the story, neglect-

ing control on the details of the story, which

is also crucial for the task. To ﬁll the gap,

this paper proposes a model for ﬁne-grained

control on the story, which allows the gen-

eration of customized stories with characters,

corresponding actions and emotions arbitrar-

ily assigned. Extensive experimental results

on both automatic and human manual eval-

uations show the superiority of our method.

It has strong controllability to generate sto-

ries according to the ﬁne-grained personalized

guidance, unveiling the effectiveness of our

methodology. Our code is available at https:

//github.com/victorup/CHAE.

1 Introduction

Story generation, one of emergent tasks in the ﬁeld

of natural language generation, requires following

sentences given the beginning of the story. For hu-

man beings, it is believed that storytelling requires

strong logical thinking ability and organizational

competence, and for machines it may be even more

intractable. Nonetheless, works on story generation

can help machines communicate with humans and

drive improvements in natural language processing

(Alabdulkarim et al.,2021).

At present, most works on story generation fo-

cus on the coherence of the story generated ac-

cording to keywords, outlines and commonsense

knowledge (Yao et al.,2019;Guan et al.,2019;

Rashkin et al.,2020;Guan et al.,2020;Ji et al.,

2020). Some other works aim at generating sto-

ries controlled by overall emotion, style, and topic

∗Corresponding author

(Keskar et al.,2019;Xu et al.,2020;Brahman and

Chaturvedi,2020;Kong et al.,2021). However, in

reality, people often expect more detailed designs

catering to their needs rather than a simple theme or

topic in the generated story. For example, a novel

with more complete elements, i.e., plot, character,

theme, viewpoint, symbol, and setting is usually

preferred to those made up out of thin air.

Taking the control in story generation as the cut-

ting point, GPT-2 (Radford et al.,2019) can fulﬁll

the story according to the beginning, but the pro-

cess of generation cannot be controlled by people,

resulting in unlogical outputs that lack practical-

ity. CTRL (Keskar et al.,2019) can specify the

generation of articles with different styles through

some style words, but such control stays at the

coarse-grained level, and makes a relatively weak

inﬂuence. CoCon (Chan et al.,2020) introduces

natural language to guide text generation. Fang

et al. (2021) propose a new task that guides para-

graph generation through a given sequence of out-

line events. However, the above two studies just

explicitly add some contents to the generated sen-

tences, which is similar to forming sentences with

given phrases, not using the input as a condition

guide for the generative models. SoCP presented by

Xu et al. (2020) can generate stories under change-

able psychological state control, while it does not

govern the detailed contents of the story.

In this paper, we consider more ﬁne-grained

control on story generation, and propose a model,

CHAE

for ﬁne-grained controllable story gen-

eration, allowing the generation of stories with

customized

aracters, and their

ctions and

motions. Characters are the core of the story.

Their actions drive the story along, and their emo-

tions make the story lively and interesting. Con-

sequently, we take the characters along with their

actions and emotions as the control conditions. It is

a challenge that our model needs to control multiple

characters with their actions and emotions respec-

arXiv:2210.05221v1 [cs.CL] 11 Oct 2022

tively in a story, especially under the guidance in

the form of natural language. To crack the nut, a

novel input form conducive to ﬁne-grained control

on story generation is introduced into CHAE. Con-

cretely, we use various prompts for ﬁne-grained

control conditions in different aspects. Moreover,

we design different methods for different control

conditions to improve the control effect. Inspired

by multi-task learning, we incorporate a character-

wise emotion loss while training, thus enforcing

the relevance between the characters and their emo-

tions respectively.

The contributions of our work can be summa-

rized as follows:

•

We ﬁrst take the characters with their actions

and emotions of the story into account to con-

duct more ﬁne-grained controllable story gen-

eration.

•

We propose a model CHAE with a novel in-

put form that helps the model control the story

in various aspects, and a character-wise emo-

tion loss to relate the characters and the corre-

sponding emotions.

•

The results of both automatic and human eval-

uation show that our model has strong control-

lability to generate customized stories.

2 Related Work

Story Generation

Story generation has attracted

more and more researchers to explore in recent

years. There are many challenges in the task, such

as context coherence and control. For context co-

herence, some works are devoted to introducing

a series of keywords (Yao et al.,2019), outlines

(Rashkin et al.,2020), or incorporating external

knowledge (Guan et al.,2019,2020;Ji et al.,2020)

into the story. For style and sentiment control,

Kong et al. (2021) generate stories with speciﬁed

style given a leading context. However, it only fo-

cuses on the global attributes of the story. Brahman

and Chaturvedi (2020) work on generating stories

with desired titles and the protagonists’ emotion

arcs, and Xu et al. (2020) generate stories consid-

ering the changes in the psychological state, while

they just control the emotion lines instead of the

detailed contents.

Controllable Text Generation

We have wit-

nessed the great performance of SOTA models for

BART

Encoder

BART

Decoder

CLS Head

Softmax

Linear

Input

Context Chae

Context Vector

Attention

Distribution

×(1 − *$%#)×*$%#

Emotion Label

Final Distribution

Output

!!"#

...

Figure 1: The architecture of CHAE. The input is the

concatenation of two components, Context and Chae,

which will be further explained in Sec 3. The emotion

labels are used for calculating a character-wise emotion

loss to tie up the characters and their emotions respec-

tively.

text generation these years. Despite the progress

in coherence and rationality of the text generated,

controllability remains to be challenging, which

means generating text with speciﬁc attributes, such

as emotion, style, topic, format, etc. CTRL (Keskar

et al.,2019) can control the overall attributes such

as domain, style and topic of the generated text by

adding control codes and prompts. By plugging in

a discriminator, PPLM (Dathathri et al.,2020) can

guide text generation without further training the

language model. CoCon (Chan et al.,2020) ﬁne-

tunes an intermediate block with self-supervised

learning to control high-level attributes i.e., senti-

ment and topic. Compared to the previous works,

our work places the emphasis on more ﬁne-grained,

all-round control on the generating process, includ-

ing the control of characters with their emotions

and actions in the story.

3 Methodology

3.1 Problem Formulation

The process of ﬁne-grained controllable story gen-

eration in this work is deﬁned as follows.

The input of the task has two components. We

refer to the one as

Context

. Let

Context =

(x1, x2, ..., xp)

denote the beginning sentence of

the story, which will be the initial

Context

. The

<soa>

<soc>

<soe>

<sep>

<SEP>

  



<no_action>

…<sep>  <soe> 



Figure 2: The input form of CHAE. The input starts with hsiand ends with h/si, comprising Context and Chae.

The latter is a sequence of kcontrol conditions on the next sentence to be generated, and each condition controls a

character. We show two possible forms of the control conditions after the brace. The special tokens in Chae are

further explained in Table 1.

other component is

Chae

, a sequence of

ﬁne-

grained control conditions on the next sentence.

Each condition in

Chae

is the combination of the

name

Chari

actions

Acti1, Acti2, ..., Actin

, and

emotion

Emoi

of a character appearing in the next

sentence to be generated, where

is not ﬁxed and

is the index of the character. Note that we use Italic

Chae

to distinguish the special input component

from our model CHAE.

The model predicts one sentence denoted as

Y= (y1, y2, ..., yq)

at a time by estimating the con-

ditional probability

P(Y|Context, Chae)

. Here

we embody the idea of auto-regression by adopting

an iterative generation strategy, that is, the sen-

tence generated is then concatenated to

Context

for next prediction. Especially, at training time, we

concatenate the gold sentences instead of generated

sentences to

Context

incrementally like teacher

forcing.

The goal of this task is to generate a story where

each sentence adheres to the input condition

Chae

in terms of character, action, and emotion, through

which elevate the quality of generation in a ﬁne-

grained manner.

3.2 Model Architecture

The architecture of our model CHAE is shown

in Figure 1. CHAE is built upon a BART model

(Lewis et al.,2020). As mentioned, our model em-

bodies the idea of auto-regression by adopting an

iterative generation strategy. On the one hand, the

iteratively updated

Chae

helps control the content

of each sentence at a detailed level of granular-

ity. On the other hand, the strategy ensures that

the model can always see the foregoing. Consid-

ering the incremental

Context

can be extra long,

we employ BART rather than GPT-2. GPT-2 is an

auto-regressive model fully based on transformer

decoder, while BART has a bidirectional encoder,

which might make it better in understanding and

encoding long input sequences. To conﬁrm the

Special tokens Meaning

hSEP iThe start token of a condition.

hsociThe start token of a character’s name.

hsoaiThe start token of actions.

hsoeiThe start token of an emotion.

hsepiThe start token of a single action.

hno_actioniThe token representing no action.

Table 1: The meanings of the special tokens in Chae.

hypothesis, we also compared BART with GPT-2

on the benchmark dataset in Sec 4, and found that

BART outperformed GPT-2 in story generation.

3.3 Generation Based on Fine-Grained

Control Conditions

To generate a story with the characters, their ac-

tions and emotions speciﬁed, we need to remind

the BART model of the elements controlled cur-

rently from time to time. Inspired by the practice of

leveraging special tokens for controllable genera-

tion (Fang et al.,2021;Keskar et al.,2019;Tsutsui

and Crandall,2017), we propose a novel form of

input (titled

Chae

), which is a sequence of

ﬁne-

grained control conditions on the next sentence to

be generated. Each condition in

Chae

controls a

character, and each segment in the condition con-

trols an element (i.e., character’s name, action, and

emotion) of the corresponding character. Note that

any number of actions can be assigned in a condi-

tion. The nested sequence form of

Chae

facilitates

the neat combination of various ﬁne-grained con-

trol conditions.

We design several special tokens and add them

between each segment as the control prompts (see

Figure 2). In this study, 6 special tokens are used to

prompt the model. They are hSEP i,hsoci,hsoai,

hsoei

hsepi

, and

hno_actioni

. The meanings of

the tokens are shown in Table 1.

Then, we encode the input

Context

and

Chae

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

CHAE:Fine-GrainedControllableStoryGenerationwithCharacters,ActionsandEmotionsXinpengWang1,HanJiang1,ZhihuaWei1,ShanlinZhou21DepartmentofComputerScienceandTechnology,TongjiUniversity,Shanghai,China2SchoolofComputerScienceandTechnology,ShanghaiUniversityofElectricPower,Shanghai,China{wangxinpeng,2230...

展开>> 收起<<

CHAE Fine-Grained Controllable Story Generation with Characters Actions and Emotions Xinpeng Wang1 Han Jiang1 Zhihua Wei1 Shanlin Zhou2_2.pdf

共12页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

CHAE Fine-Grained Controllable Story Generation with Characters Actions and Emotions Xinpeng Wang1 Han Jiang1 Zhihua Wei1 Shanlin Zhou2_2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: