DLINA Working Group

• DLINA = Digital LIterary Network Analysis
• interinstitutional working group consisting of literary scholars and computer scientists
• members: Frank Fischer, Dario Kampkaspar, Christopher Kittel, Mathias Göbel, Hanna-Lena Meiners, Peer Trilcke, Andreas Vogel
• Documentation …
  • Blog
  • DLINA on GitHub
  • "dramavis" on GitHub
  • "Play(s)" on GitHub  ·  on Google Play Store

DLINA Working Group

• main working corpus: "dlina Corpus 15.07" (see blog post for detailed description)
• based on TextGrid Repository
• comprises 465 German-language plays from 1731 to 1929
• objective: automated, philologically curated extraction and analysis of network data from dramatic texts and their philological interpretation

DLINA Working Group

• network data = interactions between characters
  • operationalisation of 'interaction': "Two characters interact with one another if they perform a speech act within the same segment of a drama (usually a 'scene')."

• more info on the automated extraction and analysis of network data:
  • "Digital Network Analysis of Dramatic Texts" (our talk at DH2015, Sydney), slides and transcription
  • blog posts on our intermediary DLINA XML format, on editing rules, on our all-in-one tool dramavis
• main subject today: How do we interpret literary network data?

Moretti's Analysis of "Hamlet" (2011)

Source: newleftreview.org

"Mapping Shakespeare’s Tragedies"

Martin Grandjean's network visualisation of 6 (out of 11) Shakespearean tragedies (Dec., 2015). Full poster and explanations on Grandjean's website, also cf. his interview on PBS.org, April 22, 2016.

"Distant-Reading Showcase"

"Distant-Reading Showcase", poster presented at #DHd2016 in Leipzig (March 9, 2016).
Download in full-res (28.88 MB) via Figshare. DOI: 10.6084/m9.figshare.3101203.v1.

So let's focus on …

… the heterogeneity of drama networks.

Interpretation Scenarios

Historical	Typological
Cf. our blog post "200 Years of Literary Network Data"	Current research: types of drama networks

Types of Drama Networks

Background hypotheses:

• Dramatic texts are context-sensitive aesthetic models of social formations, which means that …
  • … dramas represent social formations (e.g., nuclear family, royal court, 'society');
  • … these social formations only exist in their aesthetic representation, as models;
  • … these models are potentially context-sensitive and interact with real social formations.

Types of Drama Networks:
'Small World' Concept

Approach (in reference to a particular branch of network theory):

• "small-world networks" (started by Watts & Strogatz 1998; cf. Watts 2004);
  • "widespread in biological, social and man-made systems" (Watts & Strogatz 1998, 442)
  • "highly clustered, like regular lattices, yet have small characteristic path lengths, like random graphs" (Watts & Strogatz 1998, 440)
• already applied on dramatic texts (Shakespeare): Stiller, Nettle & Dunbar 2003; Stiller & Hudson 2005
• we applied the concept on a much more diverse corpus by help of automatic extraction of network data

Types of Drama Networks:
'Small World' Concept

Detecting 'Small Worlds'

Regular	'Small world'	Random
	Clustering Coefficient (C)
0,600	0,852	0,131
	Average Path Length (APL)
6,261	2,346	2,258

Types of Drama Networks:
'Small World' Concept

'Small world' networks, definition

• Criterion 1: The clustering coefficient (C) of an observed network, in our case the character network of a dramatic text, is significantly higher than the C of a corresponding random network.
• Criterion 2: The average path length (APL) of an observed network does not differ significantly from the APL of a corresponding random network.

Types of Drama Networks:
'Small World' Concept

(Additional) 3rd Criterion: 'Scale Free'

• 'Scale free' networks: variant of 'small world' networks (described by Albert & Barabási 2002);
• these networks exhibit a node-degree distribution following a power law

Types of Drama Networks:
'Small World' Concept

Overview of All Criteria (Indicators)

• Criterion 1: The clustering coefficient (C) of an observed network, in our case the character network of a dramatic text, is significantly higher than the C of a corresponding random network.
• Criterion 2: The average path length (APL) of an observed network does not differ significantly from the APL of a corresponding random network.

• Criterion 3 ('scale free'): The node-degree distribution can be best described by a power-law regression.

Criterion 1: Clustering Coefficient

• The clustering coefficient (C) of an observed network, in our case the character network of a dramatic text, is significantly higher than the C of a corresponding random network.
• Procedure
  • calculated 1,000 random networks for each individual drama network
  • calculated the mean value for random networks → Clustering Coefficient (C Random)
  • formed the quotient of C and C Random → Clustering Coefficient deviation (C Dev)
  • identified all dramas, where quotient was significantly higher (i.e., bigger than Mean + 2 × SD)

Criterion 1: Clustering Coefficient

Criterion 2: Average Path Length

• The average path length (APL) of an observed network does not differ significantly from the APL of a corresponding random network.
• Procedure
  • calculated 1,000 random networks for each individual drama network
  • calculated the mean value for random networks → Average Path Length (APL Random)
  • formed the quotient of APL and APL Random → Average Path Length deviation (APL Dev)
  • excluded all plays fulfilling criterion 1, but exhibiting an average path length significantly different from that of the random network (i.e., smaller than Mean – 2 × SD or bigger than Mean + 2 × SD, respectively)

Criterion 2: Average Path Length

Criterion 2: Average Path Length

Intermediate Result

Dramas left after application of criteria 1 & 2:

Criterion 3: Power-Law Distribution

• The node-degree distribution can be best described by a power-law regression.
• Procedure
  • calculated node-degree distribution per drama
  • calculated coefficients of determination (R²) for several regressions (linear, quadratic, exponential, logarithmic, power law)
  • excluded all dramas fulfilling criteria 1 & 2, but failing to meet criterion 3 (i.e., no power-law regression)

Criterion 3: Power-Law Distribution

Criterion 3: Power-Law Distribution

Final Result

Five dramas meeting criteria 1, 2 & 3:

Discussion

Three (of many) questions:

• 1. What does it mean to assign the 'small world' status to a dramatic text?
• 2. If 'small world' dramas are an exception, a rare species: What, then, is the norm?
• 3. In addition to 'small world' dramas: are there other kinds of structural "deviations"?

1. "Theatre Plays as 'Small Worlds'",
What Does That Mean?

1. "Theatre Plays as 'Small Worlds'",
What Does That Mean?

Central character(s) and formation of cliques
(Goethe, "Götz")

1. "Theatre Plays as 'Small Worlds'",
What Does That Mean?

Arnim, "Jerusalem"	Raimund, "Der Barometermacher"
Soden, "Doktor Faustus"	Nestroy, "Der böse Geist"

2. What is the 'norm'?

2. What is the 'norm'?

'Deviation': power-law distribution in "Götz" (example)

plenty of "lower" characters – few "average" characters – very few "upper" characters

2. What is the 'norm'?

Usually ('norm'ally) we see other kinds of distributions:

plenty of "average" characters

3. Other kinds of structural "deviations"?

3. Other kinds of structural "deviations"?

E.g., reversed power-law regression:

3. Other kinds of structural "deviations"?

Goethe, "Götz" Drama of the "great individual"	Mühsam, "Judas" Drama of the crowds
Aristocratic model?	Communist model?

More Detailed Data Extraction

• Finer granulation of types of interaction
  • detection of individual entrances and exits of characters for more philologically exact interpretations
• Quantitative and qualitative enrichment of character and relation data
  • quantitative attributes (how much does a character actually say?) → data already extracted
  • quantitative interactional attributes (how often do characters interact?) → data already extracted
  • qualitative character attributes (gender, social status) → enrichment via "Play(s)" crowdsourcing app for Android
  • qualitative interactional attributes (kinship) → planned

Corpus and Workflow Optimisation

• DLINA tools
  • further development of all-in-one Python pipeline "dramavis": input of TEI file, DLINA XML format or structured network data (CSV) → output: graph visualisations, network and character values, random-graph values
  • GUI and online platform
  • social-editing app "Play(s)" currently in beta

New Corpora, e.g.: "Theatre classique", 842 French Plays

842 character networks from 250 years of French drama history at a glance (extracted from TEI-encoded files of the "Théâtre Classique" project, see GitHub). Highlighted authors: Pierre Corneille, Molière, Racine, Crébillon, Voltaire, Euripides (negative year numbers). Result of a w33k3nd h4ck with @goebel_m & @chris_kittel.