The nineteenth century saw an explosion of sensationalist pamphlet literature in America due to technological innovations in printing, the development of a national book market, and increasing literacy rates. A significant portion of pamphlet literature (roughly 35%) featured a range of rogue figures including “pirates, corsairs, freebooters, swaggering ship’s captains, [and] mythic monsters” [1]. A prominent figure among these characters was the white woman who cross-dressed or assumed a male guise [2], providing a potential early archive of trans masculine characters and experiences. Yet, locating these crossdressing narratives and trans or trans-like characters has proven difficult because language describing gender and gendered transformation within this genre of pamphlet literature is often vague or euphemistic.

Study of gender is not new to cultural analytics, with previous research exploring characterization [3], agency [4], and gendered depictions of the body [5]. However, cross-dressing as a literary phenomenon has not yet been explored at-scale via computation, lacking a large-scale bibliography of volumes and characters relevant to cross-dressing. Our project is intended to fill this gap and take the first steps into leveraging computation to analyze such narratives.

We began our analysis by programmatically mining select records from the HathiTrust Research Center’s (HTRC) full list of MARC records. We limited our candidate volumes to items that included less than 130 pages and were published within the 19th century and geographic United States. This resulted in a potential pool of 199,687 relevant volumes. From these MARC records, we requested access to the public domain, full text of these volumes from the HathiTrust Digital Library.

Leveraging a set of expert-curated phrases that depict cross-dressing sourced from exemplar ephemeral literature volumes, we automated an identification and extraction pipeline to find sentences containing instances of these keyphrases. We piloted two main approaches for this pipeline: a more cutting-edge but abstract method using word vectors and text similarity and a more simplistic string matching approach that looked for keyphrases in the text explicitly.

The string searching pipeline was also implemented in Python using spaCy for natural language processing tasks [6]. Our data resembles keywords-in-context (KWIC) in which each sentence contains a keyphrase, sentences directly before and after, and associated bibliographic metadata and our results totaled 13,505 potential KWIC. Partial manual review hinted at a high rate of ambiguous or false results, highlighting the complexity of our task. However, positive results were also identified, and further manual review would be both fruitful for finding results and for potentially finding training data for advanced automated approaches.

Examples of positive KWIC results

We found passages of cross-dressing across a number of types of volumes and writing, ranging from legal testimony about perpetrators disguising themselves, literature depicting women fleeing in disguise and poetry depicting women warriors posing as men in battle.

We also implement a vector search model in which we chunk at the paragraph-level and index our text using Facebook AI Similarity Search (FAISS) [7]. We then search for our expert-curated phrases across all extracted books. Because FAISS is not fine-tuned for 19th-century texts, we implement a human-in-the-loop system that allows us to re-introduce more phrases based on KWIC information, as well as, subject-matter knowledge.

With a verified dataset of KWIC results, we can investigate, for the first time using a data-centric approach, and report on the following research questions (to name only a few):

What types of gendered characters are depicted as cross-dressing?

What actions do they undertake while disguised and how do other characters speak and interact with them?

Which authors, time periods and genres of texts are most likely to contain cross-dressing and most likely to exist today in digital libraries?

We’ll report on methods prototyped to identify 19th-century American ephemeral literature and detail our pipeline for leveraging exemplar volumes and phrases to identify new, unknown texts from our candidate pool in the HathiTrust Digital Library (HTDL). Additionally, we will outline our human-in-the-loop workflow that allows us to identify new entries for our bibliography. Our hope is this method for identifying understudied literary genres will be useful for others who are hoping to compile bibliographies with incomplete bibliographic data.