Thread by @maria_antoniak

Can we map how literary genres are redefined by online book taggers and reviewers? 📚

In work #CSCW2021, we show how @LibraryThing reviewers work together using free-text tags to create a shifting folksonomy that powers many IRL libraries. Genres are blurry + context-dependent!

You can read the full paper with @mellymeldubs and @dmimno here, and I’ll be presenting this virtually at #CSCW2021: maria-antoniak.github.io/resources/2021…

LibraryThing is similar to Goodreads but is more independent + accessible. Where Goodreads throttles access to its reviews, LibraryThing shows all reviews to its users. @mellymeldubs and I write about this as an “algorithmic echo chamber” in our paper on the Goodreads classics.

Unlike Goodreads, LibraryThing makes money from library software and a catalog provided to libraries and bookstores - and these resources are built on the free-text tags created by users.

There is a non-zero chance that your local library is organized according to a “folksonomy” devised by LibraryThing users!

We scrape and clean a dataset of 300 books for each of 20 popular genres (or tags) on LibraryThing. These reviews are mostly in English, and we contacted all the reviewers quoted in our paper.

(Side Note: Contacting the reviewers was really fun! They take a lot of pride in their writing and were interested in and had feedback about our project. Most wanted to be named. This is obvious to qual researchers, but NLP researchers - don’t be afraid to talk to people!)

We show different mappings of genres, for example using the book and user overlap to highlight outliers. The tags “classics” + “graphic novels” share a high number of users given their low book overlap - while “classics” + “animals” share few books given their user overlap.

We also trained a classifier to predict genre based on review text, but we were interested in *mis*classifications. Interestingly, the classifier frequently confused some genres, like "psychology" and "historical fiction," even though they don't share many books in common.

When reviews for a genre are distinctive (easy to classify), is that because of the books or because of the reviewers? Do the reviewers for that genre specialize or do they have diverse reading habits?

For example, “young adult” and fantasy” have review texts that are hard to distinguish but their reviewers have similar tag sets, while “classics” and “children” are also hard to classify but have reviewers with dissimilar tag sets.

This work builds on research from @tedunderwood, @mattwilkens, and other DH scholars. Instead of using book texts written by authors, we focus on tags and reviews provided by readers. We consider genre not as something that books have but as something readers create.

Free-text tagging gives individuals creative license to diverge from traditional catalogs. And these users’ decisions may be directly shaping your local libraries!