Repository location

https://doi.org/10.6084/m9.figshare.21166171.v1

Steps

The steps for our dataset construction were the following. Books were manually selected according to the sampling strategies described in Table 1; digitized and manually cleaned; processed using the “large model” of bookNLP (Bamman, 2022); and manually and computationally annotated for features indicated in Table 2.

Sampling strategy

All books were chosen to represent “popular” writing across 12 different genres of contemporary publishing spanning a 20-year timeframe dating from 2001 through 2021. We define “popular” through multiple criteria that include user-generated awards or lists, elite prize committee lists or book reviews, or bestseller tags on platforms like Amazon or the New York Times. As a further way to validate popularity, we provide two measures drawn from the platform Goodreads.

We define genre through three different kinds of institutional framing: cultural capital (bestsellers, prizewinners, elite book reviews), stylistic affinity (mysteries, science fiction, biography, etc.), and age-level (middle-grade and young adult (YA)). This allows researchers a high degree of flexibility to better understand stylistic behavior of professionally published books targeting different kinds of readerships. We also segment our genres by the “instrumentality” of the information contained (“fiction” or “non-fiction”).

While our genre categories are not mutually exclusive (mysteries may appear in Bestsellers and vice versa), no books appear in two separate categories. It is important to note that our larger genre categories (cultural capital, style, age) are not necessarily commensurate with one another and thus researchers should use caution when comparing across these categories. Experimentation with alternative genre labeling systems can be a further affordance of this dataset. Finally, we aimed to select ca. 200 works per category, which we have found is sufficient for training robust text classification algorithms. Due to text availability, list sizes, and cleaning, some categories have more or less than this number. In the case of those books reviewed in the New York Times, we iterated twice on this process. In total, we assemble 2,754 books representing 2,234 unique authors across 12 genres.

To further understand our data, we provide figures of the distribution of publication dates (Figure 1), the average user rating on Goodreads (Figure 2), and the log-transformed number of ratings on Goodreads (Figure 3) to capture book popularity. Finally, while no attention was given to the selection of books based on author gender, our gender distribution across all books is 49.76% women and 49.94% men with only eight books written by self-identified non-binary authors. We note, however, that there are meaningful within-genre differences (Figure 4) as predicted by prior research (Argamon et al., 2003).

Quality Control

All texts were manually cleaned of front and end matter. Metadata such as publication date, authorial gender, author name and title were all manually entered. The dataset was manually reviewed for the appropriateness of genre labels for every book. Finally, duplicates were removed and any books that were not at least 15,000 tokens in length were also removed. No maximum length was set.

Limitations

Our data is limited by intellectual property restrictions that do not allow access to full text data. To overcome this limitation, we provide a robust set of derived data that has served in prior research as a reliable foundation for the stylistic understanding of creative writing. Our data is also limited by focusing on a single language. Future work will want to emphasize multilingual data construction to facilitate our understanding of cross-cultural stylistic behavior. Finally, for both manually and computationally derived features, we expect there to be some level of error. For the manual features, we have undertaken two-levels of review. For the computational features, the bookNLP documentation provides estimates on the expected error rates of different predictive models. Nevertheless, it is important for researchers to be aware that our derived features are always estimates. We would flag “Character Count” and “Event Counts” as two features that are worth further research due to the challenging nature of their prediction.

Object name

CONLIT

Format names and versions

.CSV

Creation dates

Start date: 2015-03-10; End date: 2022-06-22.

Dataset Creators

Andrew Piper (McGill University) was responsible for the overall design of the dataset. Eve Kraicer (McGill University) and Joey Love (McGill University) assisted with cleaning and processing the data.

Language

English

License

Attribution-NonCommercial 4.0 International (CC BY-NC 4.0).

Repository name

Figshare

Publication date

2022-09-22