Data: Quantitative Perspectives on Fifty Years of the Journal of the History of Biology

Data from the forthcoming paper:

Peirson, B. R. Erick, Erin Bottino, Julia L. Damerow, and Manfred D. Laubichler. 2017. Quantitative perspectives on fifty years of the Journal of the History of Biology 50(4).

Unless otherwise specified, data are comma-delimited. Missing values are left empty.

How to Cite

Peirson, B. R. Erick, Erin Bottino, Julia L. Damerow, and Manfred D. Laubichler. 2017. Data: Quantitative Perspectives on Fifty Years of the Journal of the History of Biology (revision 1). https://github.com/erickpeirson/jhb-data/releases/tag/1.0. doi:10.5281/zenodo.893499

Questions?

Erick Peirson (orcid:0000-0002-0564-9939)

• Web: https://erickpeirson.github.io
• Twitter: @undercaffeinatd

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

Article metadata

article_metadata.csv

From JSTOR. DOIs can be used for joins across other tables in this dataset.

• Title: string
• Date: integer
• Volume: integer
• Issue: integer
• StartPage: integer
• EndPage: integer
• DOI: string

author_names.csv

Maps author indices (used in other tables) to readable names.

• Author: integer (author index)
• Name: string (readable name)

document_authors.csv

Relations between article_metadata.csv (by DOI) and document_authors.csv (by author index).

• DOI: string
• Author: integer

Geography

We examined each of the articles in JHB over its entire run, and attempted to identify the physical location of the author at the time of publication, and to determine locations that were discussed in the article content. To locate references to locales, we took a single visual pass over each article and noted any references to municipalities, regions, or states, taking care to spend an equitable amount of time on each article. We assume that we found a subset of the total references to locations. We then found the closest match to that location in the GeoNames geographical database (http://www.geonames.org/), and recorded the corresponding Uniform Resource Identifier (URI).

For the sake of consistency, we used current location identifiers and geopolitical boundaries, which is in some cases highly anachronistic. For example, the name "Czechia" was only officially adopted by the Czech Republic in 2016, and the Republic itself has only existed since 1993, yet we have used the current term to tag articles published as early as the 1960s that refer to (historical) Czechoslovakia. For the present high-level analysis this does not have a substantial impact on our results or conclusions. In other studies, however, historians incorporating digital geographic data in their research may find it fruitful to use regional databases of historical place names (e.g. The Historical Gazatteer of England's Place Names [http://placenames.org.uk/]). GeoNames itself also has increasing support for historical place names.

article_localizations.csv

Location tags for article content and and authors.

• DOI: string
• Relation: string (content or author)
• GeoNamesID: integer

location_metadata.csv

Location information retrieved from GeoNames.

• GeoNamesID: integer
• Latitude: float (degrees north of the equator)
• Longitude: float (degrees east of the prime meridian)
• Name: string
• CountryCode: string (ISO 3166-1 alpha-2)
• CountryID: integer (GeoNames country identifier; missing values are -1)

Organisms

We used the LINNAEUS NER model (Gerner et al 2010) to tag references to organisms on each page of JHB. NER is a problem in information retrieval in which the goal is to identify words or phrases in a text that refer to instances of a particular class of entities, such as people, places, institutions, or dates. NER is usually achieved through supervised machine learning, in which a "training set" of human-annotated documents is used to train a classifier. LINNAEUS is a dictionary-based NER application, in which a large collection of documents from Medline and PubMed Central that had already been tagged with entries from the NCBI Taxonomy database were used to generate a lexicon of phrases that refer to specific taxa. LINNAEUS matches both formal taxonomic terms (e.g. species binomial names) and common names (e.g. "mouse").

organisms.csv

Entities recognized on each page.

• Entity: integer (NCBI Taxonomy identifier)
• DOI: string
• Page: integer (0-indexed)
• Text: string (matching word or phrase in document)
• Start: integer (character offset of phrase start)
• End: integer (character offset of phrase end)

taxon_labels.csv

Human-readable labels for entities in organisms.csv. Taken from the ScientificName field in NCBI Taxonomy database.

• Entity: integer (NCBI Taxonomy identifier)
• Label: string (usually a binomial name)

Topics

Prior to model fitting we applied several preparatory transformations to the paginated full text provided by JSTOR. Within each document, we removed running headers from each page. We used an MCMC simulation to locate the bibliography within each article (if one was present) based on the distribution of specific punctuation characters and key terms; if a bibliography was identified, that content was excluded from analysis. We tokenized each page at the level of individual words, removing all punctuation, whitespace, and numeric characters. Since many concepts of interest in JHB are represented by multi-word phrases, we extracted two- to six-word phrases by applying (in three sequential passes) the criterion $\frac{N_{ij}-5}{N_i+N_j }>0.1N$ where Nij is the number of occurrences of the bigram (word i followed by word j), Ni is the total number of occurrences of word i, and N is the total number of tokens in the whole corpus (Řehůřek and Sojka 2010). After conjoining word-parts into phrases as described, we removed tokens of any individual words that (a) occur in the Natural Language ToolKit stopwords list (Bird et al 2009), or (b) occur on more than 6,000 pages (about 1/4 of the corpus).

We fit a series of topic models to the articles in JHB, treating each page as a separate document and varying the number of topics with . We used the parallelized collapsed Gibbs sampler implemented in the InPhO Vector Space Model package (Murdock 2015), and ran each simulation for 10,000 iterations. In each case the simulation converged successfully.

word_labels.csv

This is the corpus vocabulary; maps "word" indices (integer) to human-readable labels. Note that many of the "words" in the vocabulary are actually multi-word phrases.

• Word: integer
• Label: string

topic_page_assignments__k[N].csv

Posterior probability of topic assignments for each page.

• Topic: integer (0-N)
• DOI: string
• Page: integer
• Probability: float
• Assignments: float (number of words on page assigned to topic)
• Characteristic: float (number of times more likely topic is to occur on this page than on a random page in the corpus).

word_topic_assignments__k[N].csv

Posterior probability of words given each topic.

• Topic: integer (0-N)
• Word: integer (word index)
• Probability: float

topic_article_assignments__k[N].csv

Data from topic_page_assignments__k[N].csv aggregated and re-normalized at the article level.

• Topic: integer (0-N)
• DOI: string
• Probability: float
• Assignments: float (number of words in article assigned to topic)
• Characteristic: float (number of times more likely topic is to occur in this article than on a random page in the corpus).