Patrick J. Burns (University of Texas at Austin / Quantitative Criticism Lab)
Latin prose composition courses remain a fixture of many classics programs and not without a wide-ranging, decades-spanning debate on the effectiveness and utility of such courses (Ball and Ellsworth 1989; Newman 1990; Saunders 1993; Meinking 2017). Alongside this debate has arisen a number of assignments and activities designed to bolster effectiveness and make the practice of writing in Latin more relevant to as many students as possible (Davisson 2000; Fogel 2002; Dugdale 2011; Trego 2014; Gellar-Goad 2015; Kershner 2019; Barrett 2020). Yet one way in which historical-language prose composition leaves open room for improvement is in providing additional avenues for supporting students, especially in the form of immediate responsive feedback (i.e. at the moment of writing). This paper argues for the use of computational language models and text-analysis methods as developed in classics-focused digital humanities research to supplement teacher-provided feedback when writing in Latin. Based on recent experience teaching a graduate-level Latin prose composition course, this paper addresses two areas in which a data-driven approach to providing better feedback on and more consistent assessment of student writing is most immediately practical: lexical choice and stylistic imitation. With respect to lexical choice, when students are learning to write in Latin, how do they know which words are likely to follow immediately upon each other? Or whether certain words tend to collocate? Or how to choose between near-synonyms (like facio and ago)? These are all text-analysis tasks that are made more tractable with the use of semantically oriented language models (Bamman and Crane 2008; Sprugnoli, Passarotti, and Moretti 2019; Bamman and Burns 2020). These models can be used in coordination with traditional resources like Dumesnil’s Synonymes Latins or Döderlein and Arnold’s Hand-book of Latin Synonyms for a more evidence-based word selection in composition. With respect to stylistic matters, a common prose composition assignment is to write something “in the style” of Cicero or Livy or Tacitus, among others. Yet the question of how to evaluate such mimetic composition remains non-trivial. Stylometric methods used in research on Latin literature (Forsyth, Holmes, and Tse 1999; Marley 2018: 22–60; Zhang, Cohen, and McGill 2018; Chaudhuri et al. 2019; Ramminger 2021) provide the means for measuring how Ciceronian students’ writing is (or is not) or evaluating whether their work is more Livian or Tacitean, doing so in a way informed by features extracted from the texts. In sum, I argue that Latin prose composition courses should avail themselves of methods and resources from digital humanities (and adjacent fields such as natural language processing and corpus linguistics) in order to provide students with instantaneous, data-driven feedback on lexical and stylistic choices while they are writing as well as to provide them with more consistent, rigorous evaluation of what they have written.