“What is Dead May Not Die”: Locating Marginalized Concepts Among Ordinary Biologists

Abstract

Historians and biologists identify the debate between mechanists and vitalists over the nature of life itself with the arguments of Driesch, Loeb, and other prominent voices. But what if the conversation was broader and the consequences deeper for the field? Following the suspicions of Joseph Needham in the 1930s and Francis Crick in the 1960s, we deployed tools of the digital humanities to an old problem in the history of biology. We analyzed over 31,000 peer-reviewed scientific papers and learned that bioexceptionalism participated in a robust discursive landscape throughout subfields of the life sciences, occupied even by otherwise unknown biologists.

Appendix: How We Established the Number of Topics in Each Model

We found sparse documentation on which method to use under what circumstances, and we wanted to embrace the hermeneutica ethos of iterative understanding (Rockwell and Sinclair 2015). This led us to focus on three methods:

• 1. Scholarly interpretation of top words output for multiple models of the same data set using LDA (Blei et al. 2003) and Collapsed Gibbs Sampling as provided by the lda R package (Chang 2015). We limited our use of this method for the smallest models.

• 2. Topic tuning using the ldatuning package in R, specifically the function FindTopicNumber using three metrics named for the associated publications (Griffiths 2004; Cao et al. 2009; Arun et al. 2010) with the Collapsed Gibbs Sampler method for creating the model (Nikita 2017).

• 3. Five-Fold Cross Validation using LDA and the Collapsed Gibbs Sampler method, also part of the ldatuning R package.

Each method suggested slightly different topic numbers, and the ldatuning tools provide a range, not a specific value. As a result, we built nearly 200 models to explore these variations. Other than K (the number of topics), the parameters were consistent: alpha 0.2, beta 0.2, burn-in iterations 5000. Stop words were compiled from the standard English list used by Voyant Tools, with the addition of Roman and Arabic numerals, special characters, and terms related to the genre of scientific writing: appendix, conclusion, discussion, figure, fig, graph, introduction, methods, references, results, table, volume, vol. Voyant’s list of stop words included common adverbs, auxiliary verb forms, comparatives, conjunctions, prepositions, and pronouns. When building the models, we used the set.seed() function in R for reproducibility, using the value of 357 in a demonstration of the LDAvis package (Sievert 2014). Note, we did not use LDAvis, but repurposed topic modeling code from a different project. While this exploratory method did point us to articles that answered our questions, and we document the details for reproducibility, in future projects we would revise this method for efficiency.

For the first method–scholarly interpretation–we relied on attributes inspired by Chang et al. (2009) as ways of quantifying human interpretability of topics: word intrusion and topic intrusion. We approached our reading of the top high-probability terms in the topics as though any word or topic could be an intrusion. We read the output of the modeling algorithms with this skepticism, balanced by knowledge of the field and period. That is, our exploration of the models focused on two aspects: semantic consistencies in top words in topics across models and the ability of the topics to identify relevant passages, as we describe in the main body of text.

As indicated, the second method–the tuning process–deployed three metrics for evaluation of perplexity, which in topic modeling is a measure of how well the probabilistic model describes an unseen sample. Each method produced different quantitative results, which pushed us to assess models built with the topic numbers suggested by the extrema of minimum perplexity and the ranges in which perplexity was consistent across other methods.

Five-Fold Cross Validation is a method that uses 80% of the corpus to build the model and then tests the model against the held-out passages. This occurs five times for each possible number of topics, so that passages are alternatingly part of the training corpus or part of the test data. The algorithm returns perplexity scores for each of the five models. Using a polynomial regression, we determined the topic number with the lowest overall perplexity, often a range.

Below is a summary of the output from the methods to establish the number of topics for the keywords presented in this article (Table 9).

Table 9 Suggested number of topics for models built on different size context around the four keywords presented in this paper using the methods described above