📄 The Afterlives of Shakespeare and Company in Online Social Readership

Summary

This paper asks how far reading behavior from a historical community can be compared with reading on a modern social platform. It links the digitized borrowing records of Shakespeare and Company (SC), the English-language lending library Sylvia Beach ran in interwar Paris, to Goodreads book pages. Of 6,018 SC works, 4,460 were matched to Goodreads entries through API search and manual verification. The authors then scraped metadata (ratings, reviews, shelves, genres, lists) and released it publicly.

They compare popularity on a normalized rank scale, find the books whose standing rose or fell most, and contrast co-readership networks. Author popularity correlates moderately across the two sources. Older classics such as Austen and Alcott gained ground, while experimental contemporaries of SC, such as Dorothy Richardson, lost it. Some authors, including Faulkner and Hemingway, are still read alongside the same peers. At the whole-network level, centrality measures mainly highlight two prolific borrowers rather than a literary core. The study shows both the value and the limits of computational reception history.

Key points

• 4,460 of 6,018 SC books (74.1%) were matched to Goodreads, leaving 4,454 books by 1,726 authors after scraping. Of these, 3,940 have at least one rating and 3,223 at least one review. The data is released at gyauney/shakespeare-and-company-social-readership on GitHub.
• The two scales differ enormously: the most-borrowed SC book has 56 borrows, while the most-reviewed Goodreads book (Pride and Prejudice) has 77,817 text reviews.
• Ranks scaled to [0, 1] give an author-level Pearson r = 0.51 (p < 0.05). Among shared books, 18 of the SC top 100 are also in the Goodreads top 100.
• The biggest risers (about +0.96 in rank) include Little Women, Dracula and The Wonderful Wizard of Oz, mostly written before SC opened. All of the biggest fallers were contemporary to SC, for example Katherine Mansfield and Dorothy Richardson.
• Network centrality and core-periphery analysis of SC mostly magnified two heavy borrowers, Killen and Raphaël, whose shared reading habits suggest they influenced each other.

Interesting ideas

• I1. Converting raw counts to ranks scaled to [0, 1] makes two sources comparable even when their scales differ by three orders of magnitude, and it damps the dominance of a few huge items.
• I2. The most informative items are often the outliers of a correlation: works whose rank moved most between two sources, rather than those that agree.
• I3. Automatic catalog search was followed by manual checking, with explicit tie-break rules: prefer the most specific edition over collected works, then the one with the most ratings.
• I4. A whole-network metric that seemed to describe a community actually reflected two prolific individuals, a reminder to check whether an aggregate is driven by a few contributors.
• I5. Whether a work keeps the same co-reading neighbors across eras is used as a measure of whether its context of reception survived.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): scaled ranks could compare how often constructions or verb frames occur in small treebanks versus large silver corpora, without the big corpus swamping the comparison.
• I2 (learn): ranking constructions by how much their rank differs between two witness parsers or two treebanks would point directly at where repair rules or UD dialect conversion are needed.
• I3 (world): the same policy could guide linking text entities to Wikidata, preferring the specific item over a collection and using popularity only as a tie-breaker, with the chosen rule logged for explainability.
• I4 (learn): evaluation reports could include a control that removes the most prolific texts or sources and checks whether a score or induced rule still holds.
• I5 (world): comparing an entity's or verb's graph neighbors in old fables versus modern texts could flag sense or usage drift before it is compiled into shared knowledge.