Multi-Graph Encoder-Decoder Model for Location-Based Character

Networks in Literary Narrative

Avi Bleiweiss

BShalem Research, Sunnyvale, U.S.A.

Keywords:

Narrative Intelligence, Geo-Location, Graph Theory, Community Discovery, Neural Networks.

Abstract:

In the past decade, an extended line of research developed a broad range of methods for reasoning about nar-

rative from a social perspective. This often revolved around transforming literary text into a character network

representation. However, there remain inconsistent traits of narrative structure produced computationally by

either neural language technology or network theory tools. In this paper, we propose an encoder-decoder

model with a main objective to mitigate the apparent computational divergence. Our encoder novelty lies in

generating hundreds of location-based network graphs to render a ﬁne-grained narrative. We further formal-

ize a decoder task for detecting character communities and analyze modularity and membership afﬁliation.

Through empirical experiments, we present visualization of stages in our computational process for four liter-

ary ﬁction novels.

1 INTRODUCTION

Over a decade ago, the study of literature underwent

a major shift from close reading of individual texts to

the construction of abstract models. The quantitative

approach to literature has Moretti (2005) graphically

map out text according to history, geography, and

social connections. Moreover, by turning time into

space, a narrative plot can be further represented as a

social network of characters and interactions (Moretti,

2011). Network analysis also offers a powerful mode

of intrinsic criticism, by providing empirical mea-

sures of the novel social scale and density (Alexander,

2019).

A character network is a graph describing a narra-

tive by representing the characters as its vertices, and

their structural relationships with others through its

edges. Edges are often attributed a weight and direc-

tion properties to express multidimensional related-

ness. Because much of what the characters do or say

is narrated, a direct discourse only covers a small part

of the plot and thus the transformation of plots into

networks is a lot less consistent. Our work centers

around automating network extraction from ﬁctional

novel text by applying advanced natural language pro-

cessing (NLP) technology.

The emergence of sociological approaches to a

narrative, cast characters as points of social intersec-

tion rather than centered subjects. In her seminal

work, Levine (2009) suggests the networked novel ex-

tends alternatives to conventional constructions and

offers a perceptive account of how disease reveals so-

cial networks. Characters in text are drawn into one

great distributed network, but often they act as nodes

on two or more different networks. Our study is mo-

tivated by having the unfolding plot revolve around

multiple principles of interconnections to capture so-

cial experience.

Understanding how spoken language is repre-

sented in a novel over time is a key question in the

digital humanities. In particular, representing social

relationships between characters is an important com-

ponent of literature, as Elson et al. (2010) took the

ﬁrst steps toward automating the task of mention-level

quote attribution for literary text. In our work, in-

stead of conversational networks we render the rel-

atively understudied narrative dimension of named

places into a literary network structure that captures

character-speciﬁc physical settings and geo-locations.

We present the network as a collection of many small

location-based graphs, thus encoding an explicit spa-

tial logic of the narrative. Our study leverages recent

advances in neural NLP for the tasks of named entity

recognition (NER) and coreference resolution.

This paper offers the following key contributions:

(1) we propose a graph encoder-decoder model that

consistently retains a deﬁned narrative structure at

each of its computational stages, (2) we introduce a

790

Bleiweiss, A.

Multi-Graph Encoder-Decoder Model for Location-Based Character Networks in Literary Narrative.

DOI: 10.5220/0011776400003393

In Proceedings of the 15th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2023) - Volume 3, pages 790-797

ISBN: 978-989-758-623-1; ISSN: 2184-433X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

NER/

Coref

Network

Encoder

Novel

Text

GPE

PERSON

()

()

()

index

Graph

Union

Narrative

Graph

(a) Encoder.

Community

Decoder

Character

Affiliation

()

()

()

Membership

Narrative

Graph

Weighting

(b) Decoder.

Figure 1: A graph encoder-decoder model for extracting and analyzing a location network that represents a narrative in

literature. The encoder input is a literary text sequence that follows location and character NER along with coreference

resolution of aliases. In the next stage, the encoder computes n small geo-location graphs g

1:n

, each linking a handful of

characters. The graphs are then merged into a single narrative network G. The decoder operates on graph G and performs

pairwise graph weighting and community discovery to produce m communities c

1:m

. This process compresses the narrative

representation of hundreds of geo-location graphs down to under a dozen communities. Lastly, the decoder extracts character

memberships from each community.

formal decoder that is tasked with detecting charac-

ter communities that reduce the literary network com-

plexity from hundreds of small graphs to less than a

dozen entities, and (3) through extensive qualitative

and quantitative analysis we provide visualization of

the narrative network drawn from different perspec-

tives.

2 RELATED WORK

Detecting character relations plays an important role

in narrative understanding built upon social network

theory. In this context, one of the more vital con-

tributions to literary theory in the past two decades

is the concept of character-space relation, coined by

Woloch (2009). In his account of ﬁctional characteri-

zation, a ﬁgure, central or minor, emerges in a prede-

termined position within the narrative. The conﬁgura-

tion and intersections between many character-spaces

within a single narrative are essential to the dynam-

ics of literary representation. In our work, we fol-

low this core abstraction and generate many location-

based network graphs to render a novel. We then

unionize the graphs to form a single social network

and analyze the distribution of character communi-

ties.

The critical review by Labatut and Bost (2019)

presents extensive scientiﬁc research related to ex-

tracting and analyzing ﬁctional character networks.

We follow in more detail on the methods most rele-

vant to our study.

Considered by many the more inﬂuential work,

Elson et al. (2010) characterize a text of literary

ﬁction by extracting a network of social conversa-

tions. They obtain character mentions from conver-

sational segments in nineteenth-century British nov-

els by analyzing their dialogue interaction. Using

the NER pipeline in CoreNLP to discover charac-

ter names, they expand on aliases by using manually

drawn coreference chains. The use of explicit quo-

tations in text for inferring links between characters

remains however ambiguous at times.

To identify interacting agents, Agarwal and Ram-

bow (2010) build character networks using tree

kernel-based relation extraction from structured parse

tree information. A text snippet may thus describe

social relations between two individuals explicitly by

the type of a relation, or implicitly by a social event.

Whereas Lee and Yeung (2012) assert that quoted

speech need not assumed to be the main course of

encoding interpersonal relations (Elson et al., 2010).

Besides people and events, they also integrate loca-

tions in their networks. Our work expands on their

model and a geographical node is more than just a

global staging position, instead we generate many

network graphs to represent the novel text, each of

a star topology with a hub that captures a location

chronologically.

A machine learning model devised by Celikyil-

maz et al. (2010) describes a probabilistic approach

for detecting conversations between actors in a novel,

and analyzing networks built based on topical similar-

ity in actor speech. Their method follows the linguis-

tic intuition that rich contextual information can be

useful in understanding dialogues. On the other hand,

He et al. (2013) proposed an alternate venue for iden-

tifying speaker references in novels, using a proba-

bilistic model that exploits lexical and syntactic clues

in the text itself. However, they manually construct a

list of characters and their aliases, unlike Vala et al.

(2015), who proposed an eight stage pipeline for de-

tecting characters automatically, which builds a graph

where nodes are names and edges connect names be-

longing to the same character.

Compelling is Edwards et al. (2020) study

that model and compare manual to automatic co-

occurrence and unsupervised machine learning meth-

ods for extracting social networks from narratives.

In their ﬁndings, automatic extraction methods pro-

duced commensurable results for density and cen-

Multi-Graph Encoder-Decoder Model for Location-Based Character Networks in Literary Narrative

791

trality measures with the more accurate but by far

more time consuming manual approach. While edge

weights were only moderately correlated with a 0.8

Spearman coefﬁcient for NLP networks. Although

their experiments conducted on a television show,

their conclusions are likely to extend to literary nar-

ratives.

More recently, Schmidt et al. (2021) applied both

a rule-based pipeline and an end-to-end deep learning

model to NER and coreference resolution (Lee et al.,

2018), and showed that neural networks outperform

the rule-based approach on most evaluation settings.

Lastly, the effort by Piper et al. (2021) seeks to pro-

vide a coherent theoretical foundation for implement-

ing NLP computational solutions and identify narra-

tive as an important basis for understanding human

behavior.

3 MODEL

In Figure 1, we provide an overview of our proposed

graph encoder-decoder model.

The encoder process for transforming novel text

to a list of location-based networks of characters is

straightforward. We scan the entire novel text and

extract words deﬁned as either Geo-Political (GPE)

or PERSON entities, using the spaCy named entity

recognition (NER) and the neural coreference reso-

lution (Coref) tools.

A GPE occurrence triggers

both terminating the creation of the current network

and also starting to construct a new network. After

that we collect all the PERSON names that are de-

limited between a pair of GPEs or a GPE and end of

text. The graphs we construct are each of an undi-

rected star topology— they have the GPE as the root

node, and all the PERSON nodes are the leaves. In

addition, we draw timestamp attributes from the run-

ning word index of either the GPE or PERSON en-

tities. The indices aid in retaining the chronology of

the story telling.

More formally, using a colon notation, we denote

a collection of k PERSON entities p

1:k

= (p

,..., p

Given k characters in a location, the graph g that we

construct has k + 1 vertices and k edges emanating

from 1 : k all to vertex k + 1. The graph cardinal-

ity is thus

= k. PERSON names p

and location

identity l are attached to their corresponding vertex

labels. We unionize all n graphs g

( j)

into one network

G =



(1)

,...,g

(n)



that represents the narrative, ex-

pecting location graphs g

( j)

to connect each a small

number of characters— about a handful on average.

https://spacy.io/

The algorithm for integrating subgraphs g

( j)

has lin-

ear node complexity with the total number of charac-

ters in a narrative. Thus, the same character residing

in multiple locations is assigned an identical node ID

to effectively address large-scale and dynamic narra-

tives.

Unlike the encoder components that mostly oper-

ate in NLP space, the decoder computationally func-

tions entirely in the network theory domain. The nar-

rative network G links the encoder with the decoder,

and the latter produces a set of m communities, where

m << n. Each community c is a disjoint union of

separable geo-location graphs. We denote a commu-

nity set C =



(1)

,...,c

(m)



. While there are several

methods for community detection in networks, in our

work, we chose the Louvain (Blondel et al., 2008) hi-

erarchical clustering algorithm, owing to both its abil-

ity to detect high-modularity community partitions

and its exceptional computational efﬁciency with a

runtime of O(nlogn). The last stage of the decoder

generates character membership in each community,

we further use for visual analysis.

4 EVALUATION

In this section, we provide for our experiments visual-

ization of network modeling and report computational

metrics using igraph (Csardi and Nepusz, 2006).

Table 1: Our test set of ﬁction literary novels.

Title Chapters Tokens

Sign of the Four 12 43,736

Portrait of a Lady 27 112,661

Emma 55 159,950

David Copperﬁeld 64 361,938

Table 2: Location character density across test novels.

Title Characters Locations Density

Sign of the Four 558 106 5.26

Portrait of a Lady 1,436 319 4.50

Emma 3,980 443 8.98

David Copperﬁeld 5,845 1,106 5.28

Novel Test Set. We obtained unicode encoding of

the ﬁctional literature text from Project Gutenberg,

and carried our work on four 19-century British

novels including The Sign of the Four by Co-

https://igraph.org

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

792

Achmet

Agra

Akbar

Astrakhan

England

India

London

Singh

Tonga

Geo-Location

(a) Sign of the Four.

Albany

America

Bedfordshire

Boston

Brooklyn

England

Florence

Gardencourt

Great Britain

Italy

Liverpool

London

New York

Paris

Rome

the United States

Turkey

Uffizi

Venice

Washington

Geo-Location

(b) Portrait of a Lady.

100

Birmingham

England

Enscombe

Fairfax

Hartfield

Henry

Highbury

Ireland

Kingston

London

Maple Grove

Randalls

Richmond

Selina

Swisserland

tête

Weston

Geo-Location

100

Australia

Blunderstone

Britannia

Canterbury

Copperfield

Devonshire

Dora

Dover

Emily

England

France

India

Julia Mills

know’d

London

Murdstone

mutton

Plymouth

St. Paul’s

St. Paul’s Churchyard

Suffolk

Switzerland

Traddles

Trotwood

Geo-Location

(d) David Copperﬁeld.

Figure 2: Named geo-location distribution in logarithmic scale over our test novels.

Abdullah

Abdullah Khan

Abel White

Achmet

Agra

Alison

Athelney Jones

Atheney Jones

Aurora

Baker

Baker Street

Bartholomew

Bartholomew Sholto

Benares

Bernstone

Bishopgate

Blondin

boatman

Bouguereau

Bromley Brown

Brown

Buddha

Bunsen

Camberwell

carboy

Cawnpore

Cecil Forrester

Colin

Corot

Dawson

Dir

Dost Akbar

Englishman

Feringhee

Forrester

Greenwich

Gregson

Hall Lane

Hindoo

Holmes

Hudson

Jack

Jean Paul

Jewels

Jezail

Jim

John Holder

John Sholto

Jonathan

Jonathan Small

Jones

Lal Chowdar

Lal Rao

Lambeth

Langham

Lestrade

Lie

Mann

Martini

Mary

Mary Morstan

McMurdo

meek

Mensch

Millbank Penitentiary

Miss Morstan

Mohammedan

Mordecai Smith

Morstan

Natur

Nelson

Newfoundland

Nine Elms

Norwood

Pinchin Lane

Pondicherry Lodge

Queen

Robert Street

Roof

Salvator Rosa

Sam Brown

Sherlock

Sherlock Holmes

Sherman

Sholto

Sholtos

Sikh

Small

Smith

Somerton

Stockwell Place

Surrey

Thaddeus

Thaddeus Sholto

This Thaddeus

Upper Norwood

Vincent Square

Watson

Wiggins

Williams

Woolwich

Worcestershire

(a) Sign of the Four.

Almanach de Gotha

Amy Osmond

Annie

Annie Climber

Archer

Ariel

Bantling

Bob Bantling

brown holland

Buddha

Bunchie

Caspar

Caspar Goodwood

Catherine

Clapham Junction

Climber

Count

Daniel Touchett

Daniel Tracy Touchett

Dickens

ecus bien

Edith

Edmund

Edmund Ludlow

Edward

Edward Rosier

Elizabeth

Empire

England

Englishman

Ezekiel Jenkins

Florence

Florence Gilbert Osmond

Florence Ralph Touchett

Florentine

Gardencourt

Genoa

George Eliot

Gilbert Osmond

Goldsmith

Goodwood

Gounod

Haycock

Henrietta Stackpole

Hilary

hitherto Ralph

Isabel

Isabel Archer

Jermyn Street

jeune fille

Johnson

Jove

Juno

Justine

Keen

Keyes

Lady Jane Grey

Lady Pensil

Lancret

Lilian

Lily

Louis Philippe

Louis Quinze

Luce

Ludlow

ma fille

Machiavelli

Madonna

Matthew

Matthew Hope

Merle

Metastasio

Michael Angelo

Miss Archer

Molyneux

Monsieur Merle

Murray

Nelson

Osmond

Palazzo Crescentini

Pansy

Pau

Perugino

Philistine

Poor Ralph

Poor Ralph Touchett

Pope

Pratt

Queen Anne

Ralph

Ralph Touchett

Rosier

Saint Sophia

San Remo

Schubert

Stackpole

the Countess Gemini

Touchett

Turk

Varian

Versailles

Vittoria Colonna

Warburton

(b) Portrait of a Lady.

a Harriet Smith

a Miss

a Miss Hawkins

Abbey

Anna Weston

Anne Cox

Astley

Augusta

Augusta Hawkins

baker

Bates

Bath

Bella

Bickerton

Bird

Box Hill

Bragge

Bragges

Brunswick Square

Campbell

Campbells

Candles

Captain Weston

Catherine

Chaperon

Churchill

Churchills

Clara Partridge

Clifton

Cole

Coles

Cowper

Cox

Coxes

Dear Harriet

Dear Jane

Dixon

Donwell

Donwell Abbey

Donwell Lane

Elizabeth

Elizabeth Martin

Elton

Eltons

Emma

Emma for Harriet

Emma Woodhouse

Enscombe

Escape

Fairfax

Fetch Miss Bates

Ford

Frank

Frank Churchill

Frank to Emma

Garrick

George

George Otway

Gilbert

Goddard

Goldsmith

Graham

Green

Hannah

Harriet

Harriet Smith

Harry

Hawkins

Henry

Hetty

Hill

Hitherto

Hodges

Hughes

Hymen

Isabella

James

James Cooper

Jane

Jane Bates

Jane Fairfax

Jeffereys

John

John and

John and Isabella

John Knightley

John ostler

John Saunders

Kings Weston

Kingston

Knightley

Knightleys

La Baronne

La Comtesse

Lady Patroness

Maple Grove

Martin

Martins

Miniatures

Miss

Miss Bates

Miss Bickerton

Miss Hawkins

Miss Nash

Miss Smith

Miss Taylor

Miss Woodhouse

Nash

Nay

Otway

Partridge

Patty

Perry

Perrys

Philip Elton

Prince

Quantities

Randall

Randalls

Richardson

Robert

Robert Martin

Robin

Shakespeare

Smallridge

Smith

Stokes

Suckling

Surry

Taylor

the Miss Martins

Tom

Tupman

Uncle Knightley

Vicarage Lane

Wallis

Weston

Westons

Weymouth

William

William Cox

William Coxe

William Larkins

Windsor

Wingfield

Woodhouse

Woodhouses

Wright

Yorkshire

a Hymn Book

a Model Prisoner

a Queen Bee

Abraham

Admiralty

Agnes

Agnes Wickfield

Aliens

Angel

Animal

Annie

Babley

Baby

Bailey

baker

Bare

Bark

Barkis

Beadle

Bearers

Beauty

Bedlam

Beein

beggar

Bench

Benjamin

Betsey

Betsey Trotwood

Betsey Trotwood Copperfield

bin

Blackboy

Blackheath

blew Dora

Blood

Blossom

Blunderstone Rookery

Bob

Bodgers

Bond

Booty

Bows

Brooks

Burke

Bush

Buxton

Cain

Calais

Canning

Captain

Captain Hopkins

Caroline

Charles the

Charles the First

Charley

Chestle

Chillip

Clara

Clara Peggotty

Clara Peggotty BARKIS

Clarissa

Colleges

Commons

Cook

Copperfield

Copperfull

Creakle

Crewler

Crupp

Daisy

Dan

Daniel

Dare

Dartle

David

David Copperfield

Demon

Devil

Dick

Distant

Dixon

Doady

Dobbin

Dolloby

Dolphin

Don Quixote

Doom

Dora

Dora first

Dora matches

Dora one night

Dora sang

Dora Spenlow

Dragon

Ecstasy

Edward

Edward harm

Ely Place

Emily

Emma

Englishman

Farmer

Fatima

Fibbitson

Fingers

Folkestone

Fox

Foxe

Francis

Franklin

Future

Gaul

genie

George

George Demple

Gil Blas

Gipsy

Glass

God

Grainger

Gravesen

Grayper

Greenwich

Greenwich Fair

Gregory

Grey

Grinby

Guernsey

Gulpidge

Gummidge

Guy Fawkes

Hall

Heep

Heeps

Hem

Henry Spiker

Hindoo

Holborn

Hoorah

Hopkins

Hor

Horace Crewler

Hornsey

Humphrey Clinker

Hungerford Stairs

Hush

Hymeneal

Hymn Book

Idol

Imps

Ipswich

Jack

Jack Ketch

Jack Maldon

Jackson

Jail

James

James Steerforth

Jane

Jane Murdstone

Janet

Jellips

Jip

Jip WOULD

Joe

John

Johnson

Jones

Joram

Jorkins

Julia

Julia Mills

Kidgerbury

King Charles the

Kitt

La ra la

Lady Mithers

laid heer

Larkins

Lavinia

Lazarus

lee

Lie

Lion

Littimer

Little Blossom

Little Mowcher

London Bridge

Lookee

Louisa

Lowestoft

Lucy

Macbeth

Madam

Maldon

Mama

Mangle

Margaret

Markham

Markleham

marry Dora

Martha

Martha Endell

Mary Anne

Mates

Mawther

Mealy

Mealy Potatoes

meek

Mell

MELL

Memorial

Messrs Wickfield

Micawber

MICAWBER

Michaelmas Term

Mick Walker

Mills

Minnie

Miss

Miss Betsey

Miss Crewler

Miss Dartle

Miss Dora

Miss Helena

Miss Larkins

Miss Lavinia

Miss Mills

Miss Mowcher

Miss Murdstone

Miss Trotwood

Mistress Peggotty

Mortimer

Mortimers

Moving

Mowcher

Murdering

Murdstone

mutton

Ned

Ned Beadwood

Nettingalls

Noah

Norfolk

Norwich

Omer

Outcasts

Pagoda

Passnidge

Peggotty

Peter

Pickle

Pidger

Pitt

Poor Dora

Port Middlebay

Pretty

Puss

Put Jip

Putney

Quinion

Richard Babley

RIDGER BEGS

Rip

Robin

Robinson Crusoe

Roderick Random

Rosa

Rosa Dartle

Sarah

Sarah Jane

Scotch Croesus

Sermuchser

Settle

Shakespeare

Sharp

Sheridan

Sister Lavinia

Skylark

Snobs

Sol

Sometimes Dora

Sophy

Spaniel

Speak

Spenlow

Spenlows

Spiker

Steerforth

Strong

Suffice

Suffolk

Surrey

Ta ra la

tarpaulin

Theerfur

Thomas

Thomas Benjamin

Thomas Traddles

Tidd

Tiffey

Tom

Tom Jones

Tom Pipes

Tom Tiddler

Tommy

Tommy Traddles

Topsawyer

Towzer

Traddles

TRADDLES

Trotwood

Tungay

turret

Uncle

Uncle Dan

Uriah Heep

Ury

Valentines

Viscount Sidmouth

Voyages

Waif

Walker

Waterbrook

Westminster Abbey

Westminster Bridge

Westminster Hall

Wheer

Whether Dora

Whittington

Whoo

Wickfield

Wilkins

Wilkins Micawber

William

Windsor Terrace

Yarmouth

(d) David Copperﬁeld.

Figure 3: Network representation of our test novels. Vertices are labeled with unique character names, as edges initially

connect multi-word coreferent variations of the name.

nan Doyle (1890),

The Portrait of a Lady by

Henry James (1881),

Emma by Jane Austen (1815),

and David Copperfield by Charles Dickens (1849).

These texts totaled more than half a million words

(Table 1). Our choice of the test corpus concurs with

the test set in Elson et al. (2010) and facilitates com-

pare with a baseline. Although, unlike randomly se-

lecting a handful of chapters from each book, we used

each novel text of its entirety.

Past the NER stage, the encoder has sufﬁcient data

to assess both geo-location density and distribution in

the novel narratives. In Table 2, we show location

density ranging from a handful up to nine characters

across our test novels. This is suggestive of the num-

ber of location networks, as well as the cardinality of

narrative graphs our encoder constructs. While in Fig-

ure 2, we show in logarithmic scale the distribution of

identiﬁed named geo-locations in each of our test nov-

els, excluding places that are mentioned only once.

We expect graphs with hubs of same-name locations

to be merged in the community discovery stage of the

decoder. To ensure high quality NER, we conducted

several iterations of increasing the train set offered by

spaCy and circumvent false-negative named entities

https://www.gutenberg.org/ﬁles/2097/2097-0.txt

https://www.gutenberg.org/ﬁles/2833/2833-0.txt

https://www.gutenberg.org/ﬁles/158/158-0.txt

https://www.gutenberg.org/ﬁles/766/766-0.txt

for failing to ﬂag an entity. However, a few false posi-

tives for wrongly tagging a non-entity as a place entity

were unavoidable. For example, the word ‘Know’d’

in the novel David Copperfield was singled out as

GPE (Figure 2).

Our encoder retains a skeletal character network

for each literary narrative to aid analyzing charac-

ter aliases visually. Aliases transpire for mentioned

multi-word names and are learned by the neural coref-

erence resolver. We outline aliases in each of the nar-

rative network representations in Figure 3. Graph ver-

tices are labeled with unique character names, and the

edges, shown at the bottom left of the network, con-

nect vertices of character aliases.

Community Discovery. Detecting communities in

a literary narrative network based on geo-locations

is useful to broaden character social relations and

complement the downstream task of quote attribu-

tion. To this extent, community discovery can an-

swer the question of association by singling out sub-

graphs with identical named locations and effectively

merging them together into a concise representation.

Similarly, a community may link subgraphs of dif-

ferent named locations while sharing the same char-

acter node attributes. To help understand the narra-

tive space concept, a community structure facilitates

events that are likely to disseminate across a multi-

Multi-Graph Encoder-Decoder Model for Location-Based Character Networks in Literary Narrative

793

Table 3: Statistical summarization of graph node distribution that characterizes the input to our community discovery stage.

Title Graphs Nodes Min Max Mean STD

Sign of the Four 66 320 1 26 4.84 5.06

Portrait of a Lady 216 635 1 10 2.93 2.06

Emma 349 1,707 1 20 4.89 3.59

David Copperﬁeld 672 1,890 1 20 2.81 2.50

(a) Sign of the Four.

(b) Portrait of a Lady. (c) Emma.

(d) David Copperﬁeld.

Figure 4: Character community discovery across our test novels.

tude of places. Furthermore, uncovering organiza-

tional principles in networks sets the sphere of activity

bounds where the narrative plot unfolds.

The input to the community discovery task con-

sists of independent geo-location subgraphs. In Table

3, we provide subgraph count and statistical summa-

rization of the subgraph character nodes for each of

our test novels. As expected, literary networks are

represented for the most part with hundreds of small

networks, each of a handful of characters, on average.

In Figure 4, we provide visualization of commu-

nity partitions. The Louvain algorithm we used un-

covers community structures at different resolutions

(Lambiotte et al., 2014), and uses modularity as an

objective function to optimize for ﬁnding the best sub-

division of a network. Identiﬁed communities are

both small and large and suggest a plausible repre-

sentation compression of at least an order of magni-

tude, for a narrative network with hundreds of geo-

subgraphs reduced to around ten communities. In Ta-

ble 4, we show the number of discovered communi-

ties, along with the modularity quality scale, a nu-

merical scalar that measures the relative density of

edges inside communities with respect to edges out-

side communities. Given the [−1,+1] range, our

modularity scores proved sufﬁciently compelling.

The geo-location distributions shown in Figure 2

present the ﬁrst order for predicting community par-

titions in a literary narrative. Qualitatively, we an-

ticipated subgraphs with the same named location

at their root to be merged into a single community.

Thus, the number of communities m for each of the

narratives in our novel test set ought not to exceed

m ∈

{

9,20,17,24

}

, respectively. The Louvain algo-

rithm effectively reduced the number of subdivisions

to m ∈

{

7,10,7,10

}

, respectively, with an impressive

2X compression ratio for the larger three narratives.

Character Afﬁliation. Community graph parti-

tions are a vital resource to learn about character dis-

tributions. Allocations of member afﬁliation with a

community gives the division of the vertices across

communities and is outlined in Table 4. Character ap-

portions for The Sign of Four novel are evidently

fairly balanced, however, the remainder of the literary

narratives rather offer diverse cluster sizes. In Fig-

ure 5, we show a dendrogram plot of the membership

extracted from a sampled narrative community. We

applied a hierarchy plane cutoff and rendered a hand-

ful of clearly distinct character clusters for improved

visualization. We note that named entities at the leaf

nodes of each subtree may consist of both persons and

geo-locations, of which the latter can be further ﬁl-

tered if so desired. The community dendrogram split

introduces a representation with a new and concise set

of character relationships and thus much more simpler

to understand.

Temporal Relations. Our encoder captures the

temporal aspect of narrative evolution by the asso-

ciation of named entities with a token running index

over a narrative text sequence. In Figure 6, we show

the decoder interpretation of timestamp intervals for

twenty named entities including both character and

geo-location samples obtained from The Portrait

of a Lady narrative. A few intervals shown with a

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

794

Table 4: Output community distribution across our test novels.

Communities 7

Modularity 0.43

1 2 3 4 5 6 7

Community

Size

(a) Sign of the Four.

Communities 10

Modularity 0.37

1 2 3 4 5 6 7 8 9 10

Community

Size

(b) Portrait of a Lady.

Communities 7

Modularity 0.32

1 2 3 4 5 6 7

Community

Size

Communities 10

Modularity 0.44

1 2 3 4 5 6 7 8 9 10

Community

Size

(d) David Copperﬁeld.

Thaddeus

Astrakhan

Roof

Bunsen

Bartholomew Sholto

Thaddeus Sholto

Ballarat

Paris

Holmes

Nelson

Broderick

Nine Elms

Kennington Lane

Jezail

morocco

Gregson

Lestrade

Watson

0 5 10 15

(a) Sign of the Four.

Englishman

Paris

Louis Philippe

Pau

Edward Rosier

Edward

Versailles

Elizabethan

Ralph Touchett

San Remo

Stackpole

Turkey

Bantling

Jermyn Street

Bob Bantling

Pope

Florence Ralph Touchett

Washington

0 5 10 15

(b) Portrait of a Lady.

Tom

Smallridge

Box Hill

Richmond

Miss Bickerton

Churchill

Windsor

Goldsmith

Randalls

Yorkshire

Frank

Frank to Emma

Mickleham

Chaperon

Elton

Maple Grove

Clayton Park

Oxford

Emma

Harriet

Madness

Stilton

0 5 10 15 20

English Grammar

Grayper

Blackboy

Clara Peggotty BARKIS

Topsawyer

George

Blunderstone

Devil

Clara Peggotty

Robin

Hoorah

Foxe

Joe

Mates

Clara

Sunderland

Dan

Uncle Dan

Speak

Blunderstone Rookery

Gummidge

Omer

Australia

Theerfur

Bows

Mawther

0 5 10 15 20 25

(d) David Copperﬁeld.

Figure 5: Dendrogram visualization of a sample of character afﬁliation with a community from each of our test novels.

1425

153

1886

197

204

2443

2539

2756

3696

4065

4394

4648

479

487

5721

590

5972

603

6681

6732

7966

8315

8531

9405

9821

9900

Albany

America

Archer

Archer’s

Bunchie

Edith

Edward

Elizabeth

Elizabethan

England

Florence

Jove

Liverpool

London

Miss Archer

New York

Ralph

Ralph Touchett

Touchett

Warburton

Named Entity

Timestep

Figure 6: Timestamp intervals of named entity samples.

horizontal bar are of an empty order, thus indicating a

single appearance instance in the entire plot. Instead

of using network theory features (Besnier, 2020), the

unfolding of role importance over time for either a

character or a location for this matter is rather im-

plicit in our framework— the longer the interval the

higher the signiﬁcance attached to the entity plot con-

tribution. In an absolute sense, an entity is given im-

portance priority that we qualify by how close is a

long-standing interval endpoint to the conclusion of

the narrative.

Table 5: Physical and virtual location-based narrative accu-

racy across our test set.

Title Physical Virtual Accuracy

Sign of the Four 54 12 0.82

Portrait of a Lady 150 66 0.70

Emma 295 54 0.85

David Copperﬁeld 412 260 0.61

Error Analysis. In our model, character entities

share a location based on the proximity of the per-

son mention to the location mention in the narrative

text. We distinguish between physical and virtual

plot bound geo-locations, with the latter represented

as single-node graph occurrences that render no so-

cial relationship (Table 3). While virtual locations

are perfectly valid entities in an automated graph con-

struction, we consider them an exception. This lets

us attach a quality measure for transforming narrative

text to a multitude of geo-location subgraphs, using

an accuracy form: l

/(l

+ l

), where l

and l

are

physical and virtual locations, respectively. In Table

5, we show narrative accuracy measures across our

test novels with a plausible corpus mean of 0.75.

Multi-Graph Encoder-Decoder Model for Location-Based Character Networks in Literary Narrative

795

5 DISCUSSION

In the context of a narrative space, we sought af-

ter practical graph-based computational methods that

match our proposed model. We attend to the role

space plays in a narration as a feature capacity of the

story plot, with places that make up the physical en-

vironment in which the characters of a narrative live

and move (Brasher, 2017).

Narrative Graph to Graphormer. In this section,

we contrast our graph encoder-decoder approach with

the well established Transformer (Vaswani et al.,

2017) architecture in modeling natural language data.

We reviewed whether the Transformer is suitable to

model graphs and make graph representation learn-

ing work for the task of narrative network understand-

ing end-to-end, while considering feeding the Trans-

former with our encoder output G.

Table 6: Train scores for GNN node classiﬁcation.

Title Precision Recall F1 Support

Sign of the Four 0.33 0.48 0.37 79

Portrait of a Lady 0.29 0.44 0.33 88

Emma 0.26 0.34 0.27 107

David Copperﬁeld 0.11 0.22 0.13 255

Table 7: Comparing performance to an external baseline.

System Precision Recall F1

Elson et al. (2010) 0.54 0.55 0.48

Ours 0.25 0.37 0.28

To this end, mainstream variants of graph neural

networks (GNNs; Scarselli et al., 2009) have shown

to outperform the Transformer on many graph-level

classiﬁcation tasks. Recently, Ying et al. (2021) intro-

duced the Graphormer,

built upon a standard Trans-

former neural network that encodes directly the struc-

tural information of graphs. Their proposed central-

ity and spatial encodings proved many GNN vari-

ants may be cast as special Graphormer cases, and

shown to lead the state-of-the-art performance on a

wide range of graph-level prediction tasks. How-

ever, in its current state the Graphormer mostly at-

tends to node classiﬁcation and less on loose sub-

graph clustering— vital for learning character-centric

narratives. Moreover, the training datasets for bench-

marking the Graphormer were primarily scraped from

https://github.com/Microsoft/Graphormer

physical and biological science graphs and would

have to be augmented with literary-speciﬁc network

data to beneﬁt the performance of narrative under-

standing.

Narrative Graph to GNN. To reason about our

system performance, we linked the narrative graph

structure G to a GNN. In Table 6, we outline

weighted-average train scores of GNN node classi-

ﬁcation across our literary narratives, noting that re-

call and accuracy measures are nearly identical. We

apportioned our graph data for each novel into train,

validation, and test splits of 70/10/20 percent, respec-

tively, and used degree as the node embedding rep-

resentation. Our neural model was constructed from

a two-layer graph convolutional network (GCN), and

we ran 500 training epochs on each of our narratives

using the Adam optimizer with a ﬁxed dropout of 0.2.

Although not intended to demonstrate performance,

the observation that predicting F1 scores decline with

the length of the narratives is compelling on its own.

We sought after comparing our classiﬁcation per-

formance to an external baseline. Using a test corpus

of four novels, with identical titles to the ones used in

Elson et al. (2010) for system evaluation, served our

purpose well, although the goals, tools, and resources

vary greatly between the implementations. For exam-

ple, they used a single large character network, where

we applied many small geo-location graphs. While

they explored 60 novels for their train set with over

ten million words, their novel test set is a considerable

small subset of our texts. In table 7, we compare our

average classiﬁcation scores across test novels to the

mean scores across the methods for detecting conver-

sations in Elson et al. (2010). We anticipated lower

scores for connecting our system to GNN, shown at

about 0.6X of the baseline.

6 CONCLUSIONS

In this paper, we propose an encoder-decoder com-

putational model to reason about a narrative spa-

tially. Surprisingly, many questions can be answered

by measuring the relationships between places men-

tioned in text. Geo-location networks are especially

useful to narrow down the search space of a narrative,

before deploying a realistic discourse structure.

Our analyses carve several avenues of future re-

search, such as alias resolution in mentioned location

names, introduce character proximity relationships by

using graph weights to represent physical distance be-

tween places, and address a more authentic temporal

evolution of a community over an unveiling narrative.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

796

ACKNOWLEDGEMENTS

We would like to thank the anonymous reviewers for

their insightful suggestions and feedback.

REFERENCES

Agarwal, A. and Rambow, O. (2010). Automatic detection

and classiﬁcation of social events. In Empirical Meth-

ods in Natural Language Processing (EMNLP), pages

1024–1034, Cambridge, MA.

Alexander, S. (2019). Social network analysis and the scale

of modernist ﬁction. Modernism/modernity, 3(4).

Besnier, C. (2020). History to myths: Social network

analysis for comparison of stories over time. In

Workshop on Computational Linguistics for Cultural

Heritage, Social Sciences, Humanities and Literature

(SIGHUM), pages 1–9, Online.

Blondel, V. D., Guillaume, J.-L., Lambiotte, R., and Lefeb-

vre, E. (2008). Fast unfolding of communities in large

networks. Journal of Statistical Mechanics: Theory

and Experiment, 2008(10):P10008.

Brasher, J. P. (2017). Narrating space/spatializing narra-

tive: Where narrative theory and geography meet.

The American Association of Geographers Review of

Books, 5(3):180–182.

Celikyilmaz, A., Hakkani-Tu, D., He, H., Kondrak, G., and

Barbosa, D. (2010). The actor-topic model for extract-

ing social networks in literary narrative. In Neural

Information Processing Systems (NIPS) in Machine

Learning for Social Computing Workshop, page 7,

Whistler, Canada.

Csardi, G. and Nepusz, T. (2006). The igraph software

package for complex network research. InterJournal,

Complex Systems:1695.

Edwards, M., Tuke, J., Roughan, M., and Mitchell, L.

(2020). The one comparing narrative social network

extraction techniques. In IEEE/ACM International

Conference on Advances in Social Networks Analysis

and Mining (ASONAM), pages 905–913.

Elson, D., Dames, N., and McKeown, K. (2010). Extracting

social networks from literary ﬁction. In Annual Meet-

ing of the Association for Computational Linguistics

(ACL), pages 138–147, Uppsala, Sweden.

He, H., Barbosa, D., and Kondrak, G. (2013). Identiﬁcation

of speakers in novels. In Annual Meeting of the As-

sociation for Computational Linguistics (ACL), pages

1312–1320, Soﬁa, Bulgaria.

Labatut, V. and Bost, X. (2019). Extraction and analysis of

ﬁctional character networks: A survey. ACM Comput-

ing Surveys, 52(5):1–40.

Lambiotte, R., Delvenne, J.-C., and Barahona, M. (2014).

Random walks, markov processes and the multiscale

modular organization of complex networks. IEEE

Transactions on Network Science and Engineering,

1(2):76–90.

Lee, J. and Yeung, C. Y. (2012). Extracting networks of

people and places from literary texts. In Paciﬁc Asia

Conference on Language, Information, and Computa-

tion, pages 209–218, Bali, Indonesia.

Lee, K., He, L., and Zettlemoyer, L. (2018). Higher-

order coreference resolution with coarse-to-ﬁne infer-

ence. In The North American Chapter of the Associa-

tion for Computational Linguistics: Human Language

Technologies (NAACL), pages 687–692, New Orleans,

Louisiana.

Levine, C. (2009). Narrative networks: Bleak house and

the affordances of form. Novel: A Forum on Fiction,

42(3):517–523.

Moretti, F. (2005). Graphs, Maps, Trees: Abstract Models

for a Literary History. Verso Books, Brooklyn, New

York.

Moretti, F. (2011). Network Theory, Plot Analysis. Stanford

Literary Lab, Palo Alto, California.

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

theory for computational narrative understanding. In

Empirical Methods in Natural Language Processing

(EMNLP), pages 298–311, Online.

Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M.,

and Monfardini, G. (2009). The graph neural net-

work model. IEEE Transactions on Neural Networks,

20(1):61–80.

Schmidt, D., Zehe, A., Lorenzen, J., Sergel, L., D

uker, S.,

Krug, M., and Puppe, F. (2021). The FairyNet cor-

pus - character networks for German fairy tales. In

Workshop on Computational Linguistics for Cultural

Heritage, Social Sciences, Humanities and Literature

(SIGHUM), pages 49–56, Online).

Vala, H., Jurgens, D., Piper, A., and Ruths, D. (2015). Mr.

bennet, his coachman, and the archbishop walk into

a bar but only one of them gets recognized: On the

difﬁculty of detecting characters in literary texts. In

Empirical Methods in Natural Language Processing

(EMNLP), pages 769–774, Lisbon, Portugal.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones,

L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I.

(2017). Attention is all you need. In Advances in Neu-

ral Information Processing Systems (neurIPS), vol-

ume 30. Curran Associates, Inc.

Woloch, A. (2009). The One vs. the Many: Minor Char-

acters and the Space of the Protagonist in the Novel.

Princeton University Press, Princeton, New Jersey.

Ying, C., Cai, T., Luo, S., Zheng, S., Ke, G., He, D., Shen,

Y., and Liu, T.-Y. (2021). Do transformers really per-

form badly for graph representation? In Advances

in Neural Information Processing Systems (neurIPS),

volume 34, pages 28877–28888. Curran Associates,

Inc.

Multi-Graph Encoder-Decoder Model for Location-Based Character Networks in Literary Narrative

797