Revisiting John Stuart Mill's "The Subjection of Women": A computer-assisted stylometric analysis

According to John Stuart Mill’s Autobiography, his mature work should be thought of as “the product not of one intellect and conscience but of three” (J. S. Mill [1873] 1981, 265). He claimed that The Subjection of Women (1869) was co-authored by himself, Harriet Taylor Mill, and Helen Taylor. Most of J. S. Mill’s readers have been largely unconvinced both by his claims of co-authorship and by his encomiums of his co-authors. Rather than strengthening the claims of a common “fund of thought,” collaboration, and co-authorship, his testimony to their abilities undermined them. Those who are most reluctant to take these claims at face value reject the idea that not only did Harriet Mill have an active, pervasive, and everlasting part in John Stuart Mill’s writings, but also that she was the originator of some of his most characteristic ideas. Others, however, readily admit her influence and her originality. Unlike her mother, Helen Taylor has never actually gotten any consideration as her stepfather’s co-author. Given the challenges of assessing authorship for this text through traditional methods, we apply computational stylometric analysis. Should we accept a key tenet of stylometric studies, that an author’s mind engrafts itself onto the text, then we might be able to test J. S. Mill’s claims of co-authorship. This paper presents the state of the question and the results of a supervised machine learning-based authorship identification analysis of the Subjection. We train three classifiers (SVM, K-NN, DT) on a dataset of essays by all three potential authors. These models are then used to attribute segments of the Subjection to individual authors. The most effective models attribute the majority of the text to John Stuart Mill, though stylistic traces suggest contributions from Harriet Mill and, to a lesser extent, Helen Taylor. This is a particularly difficult authorship identification issue to address. This zip file includes the corpus used for the training and test sets, the application code, summary results and comprehensive results of the three sets of tests.