Fig 1.
Growth of scientific articles published in scholarly journals, 1870–2015.
Table 1.
List of inputs to the character-based convolutional deep neural network.
Table 2.
Illustration of character embedding.
Fig 2.
Deep neural network architecture.
Fig 3.
Herfindahl scores for three classification techniques at article and journal levels.
The curves show a distribution of all the 379,413 Herfindahl scores that were computed for each classification technique, while the pale grey thicker vertical lines are outliers.
Fig 4.
Article-level averaged differences between Herfindahl indexes of papers.
The scores in rows are subtracted from those in columns, each time involving scores computed on all 379,413 papers having more than 100 matched references. A green cell indicates that the technique listed in the row yielded greater concentration than the technique listed in the column, and a red cell indicates the opposite.
Table 3.
A sample of the gold standard.
Fig 5.
Agreement (%) between computational and manual article classification.
Two datasets (disciplinary vs. multidisciplinary) of 100 papers were classified by five and six experts respectively to find the modal classification of each article.