Multi-Agent System for AI-Assisted Extraction of Narrative Arcs in TV

Series

Roberto Balestri

and Guglielmo Pescatore

Department of the Arts, Universit

a di Bologna, Italy

Keywords:

Multi-Agent Systems, Narrative Analysis, TV Series, Computational Narratology, LLM.

Abstract:

Serialized TV shows are built on complex storylines that can be hard to track and evolve in ways that defy

straightforward analysis. This paper introduces a multi-agent system designed to extract and analyze these

narrative arcs. Tested on the ﬁrst season of Grey’s Anatomy (ABC 2005-), the system identiﬁes three types

of arcs: Anthology (self-contained), Soap (relationship-focused), and Genre-Speciﬁc (strictly related to the

series’ genre). Episodic progressions of these arcs are stored in both relational and semantic (vectorial)

databases, enabling structured analysis and comparison. To bridge the gap between automation and criti-

cal interpretation, the system is paired with a graphical interface that allows for human reﬁnement using tools

to enhance and visualize the data. The system performed strongly in identifying Anthology Arcs and character

entities, but its reliance on textual paratexts (such as episode summaries) revealed limitations in recognizing

overlapping arcs and subtler dynamics. This approach highlights the potential of combining computational

and human expertise in narrative analysis. Beyond television, it offers promise for serialized written formats,

where the narrative resides entirely in the text. Future work will explore the integration of multimodal inputs,

such as dialogue and visuals, and expand testing across a wider range of genres to reﬁne the system further.

1 INTRODUCTION

The analysis of narrative structure has long been a fo-

cus of inquiry in both literary studies and media anal-

ysis. From Formalism and Structuralism to the more

contemporary methodologies of Classical and Post-

Classical Narratology, the study of narrative has cen-

tered on understanding the underlying relationships

and components that constitute storytelling (Ionescu,

2019). The process of narrative understanding mir-

rors other forms of comprehension: it involves iden-

tifying constituent elements and integrating them into

a coherent whole.

Television series pose unique challenges, with sto-

rylines interweaving over seasons, resisting straight-

forward categorization (Mittell, 2015). These ”nar-

rative arcs” are not singular or linear but distributed

across episodes and interconnected with other story-

lines. Some arcs conclude within an episode, while

others develop gradually, spanning seasons. Under-

standing these patterns requires approaches that can

identify connections and developments across diverse

temporal scales.

https://orcid.org/0009-0000-5008-2911

https://orcid.org/0000-0001-5206-6464

This paper introduces a multi-agent system de-

signed to assist in the extraction and mapping of nar-

rative arcs from serialized television content. The

system parses episode summaries to identify arcs and

their progressions, categorizing them into types such

as self-contained, interpersonal, or genre-speciﬁc

arcs. These arcs are stored in a relational database

and enriched with semantic embeddings in a vector

database. To test the system, we analyzed the ﬁrst sea-

son of Grey’s Anatomy (ABC 2005-), comparing the

extracted arcs to those identiﬁed by a human scholar

familiar with the series.

The remainder of this paper is organized as fol-

lows. Section 2 situates this study within the broader

context of narrative analysis and computational meth-

ods. Section 3 introduces the conceptual framework

for modeling and storing narrative arcs. Section 4

describes the technical infrastructure supporting the

system. Section 5 outlines the data collection and

preprocessing pipeline of the episodes’ plots. Sec-

tion 6 presents the multi-agent framework for extract-

ing and organizing narrative arcs. Section 7 explores

the graphical interface designed to facilitate human

oversight and reﬁnement of the system’s outputs. Sec-

tion 8 evaluates the system’s performance, discusses

Balestri, R. and Pescatore, G.

Multi-Agent System for AI-Assisted Extraction of Narrative Arcs in TV Series.

DOI: 10.5220/0013369600003890

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 1, pages 663-670

ISBN: 978-989-758-737-5; ISSN: 2184-433X

663

its strengths and limitations, while 9 closes the paper

identifying potential improvement.

The GitHub repository containing the code and

the instructions to run the software are avail-

able at: https://github.com/robertobalestri/MAS-AI-

Assisted-Narrative-Arcs-Extraction-TV-Series .

2 RELATED WORKS

The study of narrative structures and their evolution

has been a foundational area of research across disci-

plines, from literary theory to computational linguis-

tics. Movements such as Formalism and Structural-

ism in the 20th century explored the mechanics of sto-

rytelling, emphasizing recurring patterns and struc-

tures. Pioneering works like Propp’s analysis of Rus-

sian folktales (Propp, 2012) and Todorov’s structural

analysis (Todorov and Weinstein, 1969) established

frameworks that continue to inform both theoretical

and computational methods for narrative understand-

ing. In recent years, advances in natural language pro-

cessing (NLP) and machine learning have expanded

the scope of computational narrative studies (Piper

et al., 2021).

2.1 Advances in Computational

Narrative Analysis

Early computational methods, including Support Vec-

tor Machines (Hearst et al., 1998) and logistic regres-

sion (Vimal, 2020), struggled to capture the complex

relationships and temporal dynamics inherent in nar-

ratives (Young et al., 2018). The advent of deep learn-

ing, particularly recurrent neural networks (RNNs)

and convolutional neural networks (CNNs), marked

signiﬁcant progress by enabling the modeling of se-

quential and local features in text (Yin et al., 2017).

Transformer-based models, such as GPT (Rad-

ford, 2018) and BERT (Vaswani, 2017), further rev-

olutionized the ﬁeld with self-attention mechanisms

that capture long-range text dependencies. These

models leverage embeddings to represent semantic

relationships mathematically, enabling deeper narra-

tive understanding and generation (Haywood et al.,

2024). More recently, Retrieval-Augmented Genera-

tion (RAG) has improved narrative interpretation by

integrating external knowledge sources (Lewis et al.,

2020).

2.2 Temporal and Structural Analysis

of Narratives

Methods like sentiment analysis and topic model-

ing have been used to identify emotional arcs and

thematic progressions in narratives (Reagan et al.,

2016; Schmidt, 2015). However, advanced models

are needed to address non-linear temporal dynamics

and complex story turning points (Piper and Toubia,

2023). Concepts such as narrative ﬂow, suspense, and

causality remain underexplored despite their impor-

tance for audience engagement (Wilmot and Keller,

2020).

Television series, with their layered and evolving

storylines, pose unique challenges for narrative analy-

sis. Unlike standalone stories, TV series often feature

arcs spanning multiple episodes or seasons, requir-

ing robust methods to identify and track these con-

nections (Mittell, 2015). While NLP techniques have

been employed to extract narrative elements and char-

acter interactions (Dalla Torre et al., 2023; Janosov,

2021; Bost et al., 2016), these methods often rely

on signiﬁcant manual effort (Rocchi and Pescatore,

2022).

2.3 Multi-Agent Systems in Narrative

Understanding

Large language models (LLMs) have shown poten-

tial in narrative generation and analysis (Zhao et al.,

2023). Multi-agent systems (MAS) further enhance

these capabilities by enabling collaborative task pro-

cessing. MAS have been used to advance ﬁelds such

as legal reasoning (Yuan et al., 2024) and scien-

tiﬁc theory development (Ghafarollahi and Buehler,

2024), as well as narrative generation (Aoki et al.,

2023).

By assigning specialized tasks to individual

agents, MAS facilitate the extraction of narrative arcs

and the tracking of their progression across episodes,

making them well-suited for analyzing serialized nar-

ratives.

2.4 Narrative Arcs (Plotlines)

TV series diverge signiﬁcantly from the classical nar-

ration model, relying on networks of interconnected

storylines rather than single linear plots (P

erez and

Ortiz, 2021). Serialized narratives operate through a

multi-layered structure that evolves across episodes

and seasons, creating a unique sense of depth and

continuity (Pescatore and Rocchi, 2019; Rocchi and

Pescatore, 2022; Degli Esposti and Pescatore, 2023).

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

664

Narrative arcs in TV series can be categorized

into Anthology Plots, which resolve within a sin-

gle episode, and Running Plots, which span multi-

ple episodes or seasons. Running Plots include Soap

Plots, focusing on character relationships, and Genre-

Speciﬁc Plots, tied to the show’s thematic or profes-

sional elements, such as survival challenges in a zom-

bie series.

These categories build on the framework outlined

in (Pescatore and Rocchi, 2019), which uses the con-

cept of isotopy to ensure textual coherence (Greimas

et al., 1982). Episodic isotopies deﬁne Anthology

Plots, interpersonal and intrapersonal isotopies sus-

tain Soap Plots, and thematic isotopies characterize

Genre-Speciﬁc Plots.

To generalize this framework beyond medical dra-

mas, where Running Plots were initially deﬁned as

Sentimental and Professional Plots (Pescatore and

Rocchi, 2019), we propose broader deﬁnitions appli-

cable to other serialized genres.

3 MODELING NARRATIVE ARCS

FOR STORING

To analyze and document these arcs effectively, we

conceptualized a model that captures their key at-

tributes and episodic progressions in a structured for-

mat. Each arc is categorized into one of three types

based on its nature and scope:

• Anthology Arcs, which are self-contained stories

resolved within a single episode, often centered

on genre-speciﬁc events or cases.

• Genre-Speciﬁc Arcs, focusing on professional or

thematic elements tied to the series, such as work-

place dynamics or medical conﬂicts.

• Soap Arcs, which explore interpersonal relation-

ships, personal growth, and emotional conﬂicts.

Each arc is deﬁned by a title and description that

encapsulates its central theme or conﬂict. The main

characters driving the arc are identiﬁed, along with

the arc type, which provides context for its role within

the series. To track how arcs evolve over time, we

document their progressions, which capture episodic

developments relevant to each storyline. A progres-

sion speciﬁes the episode and season where it occurs,

the interfering characters inﬂuencing the arc in that

context, and a concise description of the events ad-

vancing the storyline.

Summarizing, in our system, the object Narrative

Arc includes several ﬁelds. Each arc is assigned a

unique identiﬁer (arc

id), along with a title and de-

scription encapsulating its central theme or conﬂict.

The arcs also contain progressions, which are a list of

Progression objects representing their developments

across episodes. Additionally, each arc identiﬁes the

main characters driving the storyline, speciﬁes the arc

type (Anthology, Soap, or Genre-Speciﬁc), and indi-

cates the series to which it belongs.

Each Progression is similarly structured with key

attributes. It is assigned a unique identiﬁer (progres-

sion id) and is linked to its corresponding narrative

arc through the arc id. The progression includes a

description of the events related to the arc in the spe-

ciﬁc episode (Content), as well as information about

the series, season, and episode in which it occurs.

4 TECHNICAL DETAILS

In this paper, ”LLM” refers speciﬁcally to the Ope-

nAI GPT-4o model, accessed via API. Data storage

was implemented using an SQLite database, while

embeddings were generated with the Cohere-embed-

english-v3.0 model and stored in a Chroma vector

database. The backend was developed in Python (ver-

sion 3.11), with FastAPI facilitating communication

between the backend and frontend, which was built

using the React JavaScript framework. Character en-

tities were extracted from episode plots using the

Spacy NLP model en core web trf.

5 MATERIAL COLLECTION AND

PREPROCESSING

Our software was designed to be generalizable across

various TV series genres. For testing, we focused on

the ﬁrst season of Grey’s Anatomy, a choice motivated

by our team’s prior research on medical dramas. This

familiarity provided a solid foundation for evaluating

the software’s performance. The system requires only

episode summaries of a season to function effectively.

5.1 Episodes’ Plots Gathering

To develop and test the software, we sourced episode

summaries from the fan-maintained Grey’s Anatomy

Wiki (https://greysanatomy.fandom.com), which op-

erates under the Creative Commons Attribution-Share

Alike License (CC BY-SA). This license permits re-

search and commercial use with proper attribution.

The raw data was preprocessed to ensure standard-

ization and usability for analysis.

Multi-Agent System for AI-Assisted Extraction of Narrative Arcs in TV Series

665

5.2 Data Cleaning and Simpliﬁcation

The preprocessing began with cleaning and simpli-

fying the episode summaries. Using the LLM, sen-

tences were rewritten in simpler, more structured

forms, emphasizing clarity by focusing on a single

character or event per sentence. Direct quotations

were avoided, and complex sentence structures were

replaced with concise descriptions. This process en-

sured the plots were easily interpretable and consis-

tent.

5.3 Character Entity Normalization

Character name variability was addressed by replac-

ing pronouns with corresponding character names,

guided by contextual analysis within a ﬁfteen-

sentence window. The spaCy Transformer model was

used to extract character entities, and the LLM further

reﬁned this output by resolving similar names and al-

ternative appellations. A character database was cre-

ated, containing unique identiﬁers, preferred names,

and alternative names for each character. The plots

were then standardized by replacing all character ref-

erences with their preferred names.

5.4 Season Plot Generation

To create a cohesive season-level summary, individ-

ual episode plots were ﬁrst summarized using the

LLM. These summaries were then concatenated and

further summarized to produce a streamlined narra-

tive overview of the season. This two-step process

ensured the essential details of the narrative were re-

tained while providing a comprehensive summary.

This preprocessing pipeline transformed raw tex-

tual data into a structured format, enabling detailed

narrative analysis and ensuring the software’s appli-

cability across various TV series genres.

6 THE MULTI-AGENT SYSTEM

FOR ARC EXTRACTION

The narrative arc extraction process utilizes a multi-

agent system to analyze the episode plots. It iden-

tiﬁes, categorizes, and reﬁnes narrative arcs, inte-

grating episodic storylines into a cohesive seasonal

framework. The workﬂow is sequential, with spe-

cialized agents performing tasks. The resulting arcs

and their progressions are stored in both a relational

database and a vector database, as detailed in Section

6.2. Figure 1 provides a visual representation of the

system.

Figure 1: Narrative Arc Extraction Process.

6.1 Workﬂow Design

The extraction workﬂow is divided into stages, with

each stage handled by an autonomous agent. Each

agent focuses on a speciﬁc aspect of narrative analy-

sis, ensuring arcs are both episode-speciﬁc and sea-

sonally coherent. Results from prior episodes are in-

corporated to maintain continuity and account for the

evolving nature of serialized storytelling.

6.1.1 Agent 1 - Existing Season Arcs Identiﬁer

This agent evaluates if arcs from previous episodes (if

present) are present in the current episode. If detected,

they are ﬂagged as ”possibly present” and serve as

reference points for subsequent agents.

6.1.2 Agent 2 - Anthology Arc Extractor

This agent identiﬁes self-contained, standalone story-

lines unique to the current episode. These arcs are

processed independently, as they do not contribute to

season-wide continuity.

6.1.3 Agent 3 - Soap and Genre-Speciﬁc Arc

Extractor

This agent analyzes the episode plot to identify new

soap and genre-speciﬁc arcs and to validate arcs

ﬂagged by Agent 1.

6.1.4 Agent 4 - Seasonal Arc Optimizer

This agent minimizes redundancy by analyzing soap

and genre-speciﬁc arcs for overlaps. It performs

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

666

stricter checks on ”possibly present” season arcs and

newly identiﬁed arcs, merging or reﬁning them as

needed to maintain distinct scopes.

6.1.5 Agent 5 - Arc Deduplicator

All arcs extracted from the episode, including Anthol-

ogy Arcs, are reviewed for similarity. Overlapping

arcs are resolved through a disambiguation process.

For example, arcs identiﬁed as both Anthology and

Genre-Speciﬁc by separate agents are clariﬁed by this

deduplicator.

6.1.6 Agent 6 - Detail Enhancer

This agent enriches each arc with detailed contextual

information, including: main characters driving the

arc, supporting or interfering characters inﬂuencing

the arc in the episode, a concise description of the

arc’s episodic events (the ”progression”).

6.1.7 Agent 7 - Progression Veriﬁer

The progressions of each arc are reviewed to ensure

speciﬁcity and relevance. This step ensures that sig-

niﬁcant developments are captured without overlap-

ping with unrelated narratives.

6.1.8 Agent 8 - Character Role Veriﬁer

This agent veriﬁes and, if necessary, corrects the clas-

siﬁcation of characters as either main or interfering.

6.1.9 Agent 9 - Final Reviewer

A ﬁnal veriﬁcation ensures narrative consistency. Val-

idated arcs are stored in the database for future

episodes’ analysis.

6.2 Semantic Comparison and Arc

Embedding

After analyzing each episode, embeddings are gener-

ated for the arcs. The embeddings represent the ”Ti-

tle” and ”Description” of each arc, while the ”Con-

tent” of the progression provides additional context.

These embeddings are compared against the vec-

tor database to identify semantically similar arcs. If

similarities are detected, an LLM determines whether

the arcs represent the same storyline or not. Based on

this determination, the system either stores them as

new arcs or links them as continuations of overarch-

ing plotlines.

7 GRAPHIC INTERFACE

Human input remains essential for reﬁning the out-

puts of the multi-agent system. The graphical inter-

face is designed to facilitate this oversight and correc-

tion, balancing simplicity with advanced features. It

enables users to visualize, edit, regenerate, and ana-

lyze narrative arcs effectively.

7.1 Interface Overview

The interface is divided into three main sections:

• Narrative Arc Timeline: Provides a visual repre-

sentation of how storylines evolve across episodes

in a season. Users can apply ﬁlters based on arc

type (Anthology, Genre-Speciﬁc, Soap) or associ-

ated characters.

• Vector Store Explorer: Includes tools for visual-

izing arc embeddings using 3D PCA and displays

clusters of similar arcs.

• Character Section: Allows users to explore ex-

tracted character entities, manage appellations, or

merge similar characters.

7.2 Key Features and Functionalities

7.2.1 Arc and Progression Visualization

Users can navigate narrative arcs through a tabular in-

terface, where rows represent arcs, and columns cor-

respond to episodes. Each cell displays the arc’s pro-

gression within a speciﬁc episode (see Figure 2).

Figure 2: The main view of the graphical interface.

Multi-Agent System for AI-Assisted Extraction of Narrative Arcs in TV Series

667

Figure 3: Form to add a new arc.

7.2.2 Arc Creation and Editing

Arcs can be created or edited via a dialog box (see

Figure 3). Users specify attributes such as title, de-

scription, and arc type, as well as main characters and

episodic progressions. Progressions can also be auto-

generated using AI for efﬁciency.

7.2.3 Arc Merging and Deduplication

For overlapping or semantically similar arcs, users

can compare them side-by-side and merge duplicates

to maintain consistency.

7.2.4 Progression Management

Users can manually create or edit episodic progres-

sions, which include episode-speciﬁc descriptions

and interfering characters. Alternatively, progressions

can be auto-generated and reﬁned as needed.

7.2.5 Clustering and Semantic Analysis

Clustering tools group arcs based on semantic em-

beddings and display them in both tabular and 3D

PCA formats (see Figure 4). This feature helps users

identify thematic connections and detect arcs that the

multi-agent system incorrectly treated as distinct.

7.2.6 Character Management

The character panel enables users to view, edit, and

merge characters to ensure consistency across arcs. A

similarity threshold based on the Jaccard index (Ni-

Figure 4: 3D PCA visualizer for clustering arcs based on

semantic similarity.

Figure 5: Jaccard Index indicating potential duplicated

characters. In this example, the suggestion is a false pos-

itive.

wattanakul et al., 2013) highlights potential duplicate

characters for resolution. For example, Figure 5 illus-

trates a false positive where two different characters

with the same surname are identiﬁed as similar by the

system.

8 TESTING THE SYSTEM

8.1 Test Setup

The system’s performance was evaluated by com-

paring its extracted narrative arcs, based solely on

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

668

episode summaries (paratexts), with arcs identiﬁed

by a human scholar who analyzed the ﬁrst season of

Grey’s Anatomy after watching each episode at least

twice.

Paratexts (Genette, 1997), such as episode plots,

fan ﬁction, plot maps, and similar materials, have

gained increasing signiﬁcance in the Internet era

(Laukkanen, 2024). While these materials do not, by

their intrinsic nature, constitute the primary text it-

self, they provide scholars like us with valuable sup-

plementary resources for research and analysis.

The test was conducted across the entire season,

and the results were analyzed to assess both the sys-

tem’s capabilities and its limitations.

8.2 Results

The system demonstrated notable strength in its abil-

ity to perform LLM-aided character entity recognition

and linking. It successfully identiﬁed 62 character

entities within the test dataset, with 61 of these be-

ing correct. The single duplicate entry (”Frost” and

”Jerry Frost”) resulted from inconsistent naming in

the source material. Implementing a two-agent ver-

iﬁcation system could resolve such errors but would

increase computational costs.

In arc extraction, the system excelled in identify-

ing Anthology Arcs, achieving a precision of 89.3%

(25 out of 28 arcs correctly extracted).

However, challenges were observed with other arc

types. The system occasionally duplicated arcs or

failed to merge arcs representing the same storyline.

For example, the arcs:

• ”Izzie Stevens: Overcoming Past and Professional

Growth”

• ”Izzie Stevens: Balancing Personal Life and Pro-

fessional Ambitions”

were treated as separate, whereas human analysis

considered them part of a uniﬁed storyline.

Conversely, the system overlooked the shared arc

of Meredith Grey and Derek Shepherd’s relationship.

Instead, it identiﬁed only individual character arcs

for Meredith and Derek, effectively ”diluting” their

shared storyline within the personal arcs of each char-

acter.

Additionally, the system misclassiﬁed the ”Room-

mates Dynamics” arc, which involves Meredith, Izzie,

and George becoming roommates in the second

episode. This storyline was grouped under a broader

arc, ”Intern Dynamics: Friendship And Rivalry,”

though human analysis deemed them distinct.

Progressions within arcs were generally consis-

tent with their titles and descriptions. However, they

were not always fully captured for speciﬁc episodes,

as episode summaries often omit minor developments

that a viewer might identify.

8.3 Impact of the Graphic Interface

The graphical interface proved highly effective in re-

ﬁning the system’s outputs. It enabled human ana-

lysts to review, correct, and enhance extracted arcs

efﬁciently, with the assistance of LLM tools stream-

lining the correction process.

9 CONCLUSIONS AND FUTURE

WORK

This project tackled the challenge of analyzing the in-

terwoven narrative arcs of serialized television. The

multi-agent system introduced represents a step to-

ward automating the traditionally manual process of

identifying and organizing narrative arcs for deeper

study.

While the results are promising, they also reveal

limitations, particularly the reliance on textual para-

texts, such as episode summaries, as the primary in-

put. Summaries focus on major plot-driving events,

neglecting subtle details found in visuals, dialogue,

or other storytelling layers. Consequently, the system

occasionally misses overlapping arcs or blends dis-

tinct storylines.

However, this limitation also presents an oppor-

tunity. The system is particularly well-suited for se-

rialized written narratives, where all story elements

are text-based. Serialized novels, episodic web ﬁc-

tion, and other text-driven storytelling formats elim-

inate the multimodal challenges of TV series, offer-

ing a clearer ﬁeld for arc extraction and progression

mapping. With minor adjustments, the workﬂow can

be adapted to handle chapters, segments, or seman-

tic chunks of text (Qu et al., 2024), treating them as

”episodes” to track storylines over an extended series.

Future work will enhance the system by integrat-

ing richer inputs such as subtitles, scene descriptions,

and multimodal data. Reﬁnements to the multi-agent

system are also needed to improve reliability, particu-

larly in managing overlapping arcs and deduplication.

Expanding testing across diverse genres will further

reﬁne the system’s versatility and applicability.

Ultimately, the combination of automation and

human oversight remains indispensable. While ma-

chines excel at processing and organizing large

datasets, human interpretation continues to provide

the understanding required for meaningful narrative

analysis.

Multi-Agent System for AI-Assisted Extraction of Narrative Arcs in TV Series

669

REFERENCES

Aoki, N., Mori, N., and OKada, M. (2023). Analysis of llm-

based narrative generation using the agent-based sim-

ulation. In 2023 15th International Congress on Ad-

vanced Applied Informatics Winter (IIAI-AAI-Winter),

pages 284–289. IEEE.

Bost, X., Labatut, V., Gueye, S., and Linar

es, G. (2016).

Narrative smoothing: dynamic conversational net-

work for the analysis of tv series plots. In 2016

IEEE/ACM International Conference on Advances in

Social Networks Analysis and Mining (ASONAM),

pages 1111–1118. IEEE.

Dalla Torre, P., Fantozzi, P., Naldi, M., and Pellegriti, P.

(2023). 3. looking for lexical signatures in gomorrah.

AUDIOVISUAL DATA, page 41.

Degli Esposti, M. and Pescatore, G. (2023). Exploring tv

seriality and television studies through data-driven ap-

proaches. In Audiovisual Data: Data-Driven Perspec-

tives for Media Studies. 13th Media Mutations Inter-

national Conference, pages 23–40. Media Mutations

Publishing.

Genette, G. (1997). Paratexts: Thresholds of Interpreta-

tion. Literature, Culture, Theory. Cambridge Univer-

sity Press.

Ghafarollahi, A. and Buehler, M. J. (2024). Scia-

gents: Automating scientiﬁc discovery through multi-

agent intelligent graph reasoning. arXiv preprint

arXiv:2409.05556.

Greimas, A. J., Court

es, J., Crist, L., and Patte, D. (1982).

Semiotics and language: An analytical dictionary,

volume 10. Indiana University Press Bloomington.

Haywood, S., Warren, G., and Wolf, A. (2024). Understand

tokens. Microsoft Learn.

Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J., and

Scholkopf, B. (1998). Support vector machines. IEEE

Intelligent Systems and their applications, 13(4):18–

28.

Ionescu, A. (2019). Postclassical narratology: Twenty years

later. Word and Text, A Journal of Literary Studies and

Linguistics, 9(01):5–34.

Janosov, M. (2021). Network science predicts who dies next

in game of thrones. arXiv preprint arXiv:2110.09856.

Laukkanen, M. (2024). Audience-authored paratexts: le-

gitimation of online discourse about game of thrones.

Frontiers of Narrative Studies, 10(1):35–55.

Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin,

V., Goyal, N., K

uttler, H., Lewis, M., Yih, W.-t.,

Rockt

aschel, T., et al. (2020). Retrieval-augmented

generation for knowledge-intensive nlp tasks. Ad-

vances in Neural Information Processing Systems,

33:9459–9474.

Mittell, J. (2015). Complex TV: The Poetics of Contem-

porary Television Storytelling. New York University

Press.

Niwattanakul, S., Singthongchai, J., Naenudorn, E., and

Wanapu, S. (2013). Using of jaccard coefﬁcient for

keywords similarity. In Proceedings of the interna-

tional multiconference of engineers and computer sci-

entists, volume 1, pages 380–384.

erez, H. J. and Ortiz, M. J. (2021). Multi-plot structure in

television serials. In Cognition, Emotion, and Aesthet-

ics in Contemporary Serial Television, pages 47–67.

Routledge.

Pescatore, G. and Rocchi, M. (2019). Narration in medical

dramas i. interpretative hypotheses and research per-

spectives. La valle dell’eden, 1:107–115.

Piper, A., So, R. J., and Bamman, D. (2021). Narrative the-

ory for computational narrative understanding. In Pro-

ceedings of the 2021 Conference on Empirical Meth-

ods in Natural Language Processing, pages 298–311.

Piper, A. and Toubia, O. (2023). A quantitative study of

non-linearity in storytelling. Poetics, 98:101793.

Propp, V. Y. (2012). The Russian Folktale by Vladimir

Yakovlevich Propp. Wayne State University Press.

Qu, R., Tu, R., and Bao, F. (2024). Is semantic chunk-

ing worth the computational cost? arXiv preprint

arXiv:2410.13070.

Radford, A. (2018). Improving language understanding by

generative pre-training.

Reagan, A. J., Mitchell, L., Kiley, D., Danforth, C. M., and

Dodds, P. S. (2016). The emotional arcs of stories

are dominated by six basic shapes. EPJ data science,

5(1):1–12.

Rocchi, M. and Pescatore, G. (2022). Modeling narra-

tive features in tv series: coding and clustering analy-

sis. Humanities and Social Sciences Communications,

9(1):1–11.

Schmidt, B. M. (2015). Plot arceology: A vector-space

model of narrative structure. In 2015 IEEE Inter-

national Conference on Big Data (Big Data), pages

1667–1672. IEEE.

Todorov, T. and Weinstein, A. (1969). Structural analysis

of narrative. In NOVEL: A forum on ﬁction, volume 3,

pages 70–76. JSTOR.

Vaswani, A. (2017). Attention is all you need. Advances in

Neural Information Processing Systems.

Vimal, B. (2020). Application of logistic regression in nat-

ural language processing. Int J Eng Res, 9.

Wilmot, D. and Keller, F. (2020). Modelling suspense in

short stories as uncertainty reduction over neural rep-

resentation. In Proceedings of the 58th Annual Meet-

ing of the Association for Computational Linguistics,

pages 1763–1788, Online. Association for Computa-

tional Linguistics.

Yin, W., Kann, K., Yu, M., and Sch

utze, H. (2017). Com-

parative study of cnn and rnn for natural language pro-

cessing. arXiv preprint arXiv:1702.01923.

Young, T., Hazarika, D., Poria, S., and Cambria, E. (2018).

Recent trends in deep learning based natural language

processing. ieee Computational intelligenCe maga-

zine, 13(3):55–75.

Yuan, W., Cao, J., Jiang, Z., Kang, Y., Lin, J., Song, K., Yan,

P., Sun, C., Liu, X., et al. (2024). Can large language

models grasp legal theories? enhance legal reasoning

with insights from multi-agent collaboration. arXiv

preprint arXiv:2410.02507.

Zhao, R., Zhang, W., Li, J., Zhu, L., Li, Y., He, Y., and

Gui, L. (2023). Narrativeplay: Interactive narrative

understanding. arXiv preprint arXiv:2310.01459.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

670