Why should we care?

• Interlinked circumstances of text production
• Printers’ selection, agency, and motivation
• Understanding clandestine printing

Building the Network

Working with Imprints

Approaches to Parsing Imprints

• Human reading
• Regular expressions
• Traditional natural language processing
• Fuzzy search
• Large Language Models?

Local LLMs for Imprint Parsing

Working with Local LLMs (On Your Own Computer!)

All this only gets us the known printers!

And it doesn’t tell us what kinds of texts they were printing.

Using the Network

Finding similar texts by content

• Normalization and regularization
• Term Frequency–Inverse Document Frequency
• UMAP and t-SNE
• Cosine Distance

These simple methods power a range of tools for exploring text corpora.

Using this approach to map restoration printing.

What is a document embedding?

Viewing Braddyll and Everingham with Tf-Idf

Viewing Braddyll and Everingham with LLM Embeddings

Setting up a test

1. If printers have lines, we would expect that texts printed by the same person are more similar than randomly-selected sets of texts.
2. Using a resampling procedure, we can compare sets of texts printed or published by specific individuals to the distribution of random sets of texts.
3. The metric used for comparison is the ratio of the number of texts in each set that are among the top ten nearest neighbors of any other text in the set to the total number of texts. In other words, what percentage of texts in each set are close neighbors with one another?

Exploring Braddyll and Everingham Individually

Viewing Braddyll and Everingham as a Unit

Total Braddyll/Everingham texts: 98
Texts with close neighbors from the same set: 23
Ratio of close neighbors: 0.23

If printers do have lines, why does this happen?

Possible next steps

• Printed poetry and text parts
• Graph learning and link prediction

LLMs represent an extension of cultural technologies rather than a radical break.

A human-in-the-loop process that combines researcher expertise with computational approaches at varying scales.

LLM-based methods can fit into existing workflows for computational early modern scholarship.

Spelling regularization, structured language parsing, and document embeddings are a solid start.

Understanding how LLMs work is worthwhile even if your research isn’t about LLMs.

The culture-mapping abilities of modern language models are useful beyond the novelty of the methods.

Thank you!