From Texts to Networks

Worshop in Network Analysis

Maciej Eder

University of Tartu | Polish Academy of Sciences

06/05/2026

Introduction

Aims of the study

  • Assessing large number of texts
  • Leveraging existing explanatory methods
  • Automatic analysis of textual relations…
  • … that will reveal groups of similar texts

Mapping literature

  • Chronology
  • Genre
  • Imitation
  • Authorship
  • …

Basic concepts

  • “Computation into criticism” (John Burrows)
  • “Distant reading” (Franco Moretti)
  • “Macroanalysis” (Matthew Jockers)
  • Big Data
  • Authorship attribution

Cluster analysis dendrograms

  • Highly dependent on feature selection
  • Highly dependent on similarity measure
  • Highly dependent on linkage algorithm
  • No validation provided

136 most frequent words

137 most frequent words

Unstable results

  • Noise?
  • Unreliability of the linkage procedure?
  • More than one stylistic layer?

Layers

Networks to rescue

A new technique: assumptions

  • Explanatory power of dendrograms
  • Stable results and/or validation
  • Scalability: 500 texts? 1000 texts? 31M texts?

Authorship attribution

  • An anonymous text is tested against a set of “candidates”
  • Goal: to find the nearest neighbor
  • Effective style-marker: most frequent words (grammatical words)

Establishing a distance

Another distance…

Checking each sample

Deciding which distance is shortest

From attribution to stylometry

  • If it works with anonymous texts, …
  • … what about extending the same procedure?
  • What about applying it to all texts?

Distances to txt1

Distances to txt2

Distances to txt3

A resulting network

From stylometry to networks

  • Nearest neighbors can be represented as connected nodes of a network.
  • A variety of layout algorithms can be applied.

Example

Reliability, stupid!

  • Instead of one analysis (e.g. 100 MFWs)…
  • … a whole range (100, 200, 300 MFW, etc.).
  • Next, particular “snapshots” summarized.

100 MFWs

200 MFWs

300 MFWs

Consensus network

Importance of runners-up

  • Nearest neighbors = the most similar
  • Runners-up: do they really deserve to be filtered out?
  • Solution: more connections!

3 neighbors to txt1

3 neighbors to txt2

3 neighbors to txt3

A resulting network

Two algorithms togegher

Latin literature at a glance

Traces of chronology?

  • “Ciceronianus es, non Christianus!” (God accusing St Jerome)
  • “Renovatio antiquitatis” (Renaissance humanists)
  • hypothesis: noticeable traces of chronology

How influential was Cicero?

  • Ciceronian Quarrel: the single most important debate of the Renaissance.
  • Imitation of the admirable Ciceronian style.
  • hypothesis: visible traces of Cicero in early modern Latin writings.

Conclusions

  • No clear chronological pattern.
  • Genre signal: predominant.
  • Masters of style tend to keep outside the network.

Further research

  • More texts!
  • More genres!
  • More “Attic” writers (e.g. Lipsius).