05 - Zukunftsphilologie: The Rewards (and Perils) of Machine-Human Collaboration

This paper illustrates how artificial intelligence supports the editing of premodern Greek texts, outlining some rewards and perils of extreme interdisciplinarity. We use the term Zukunftsphilologie (Wilamowitz-Möllendorff 1872) to argue that AI can contribute to the "philology of the future" - and, simultaneously, that traditional philology can suggest some novel approaches in AI.

I. We begin with lacunae and scribal errors. For the latter, we present the results of an algorithm we designed to support the task of detecting and emending corrupted portions of text. We make the case for the usefulness of AI-generated "diagnostic conjectures" (Maas 1958: 53-54; West 1973: 58; see also Crane et al. 2007) and illustrate what AI-assisted philological workflows may, in the future, productively involve.

II. Next, we ask what counts as "ground truth", "proof", and "success" in philology/AI. For our tool, we define "success" as adoption by philologists in their regular workflow. To that end, in addition to statistically evaluating machine vs. human vs. machine+human performance (cf. Assael et al. 2022), we focus on presenting results which, according to peer review, substantially improve current understanding of premodern texts (cf. Pollock et al. 2015 for philology as 'making sense of texts'). Proof that a tool is useful for lacuna-filling is relatively easy to set up, because lacunae can be simulated by masking random portions of transmitted text (i.e. ground truth). Detecting and emending scribal errors, in a long tradition of hand-copying from earlier exemplars, is more challenging, because scribal errors are not random and there is doubt about textual authenticity (i.e. no ground truth). Our corruption-detection algorithm weighs the low chance that a portion of text should appear where it does against high machine confidence in the top emendation. It suggests to us a way to generate probabilities for spans of text using BERT, a current desideratum in machine learning. For AI, a scribal intervention that simplifies the text (lectio facilior) represents the greatest challenge, since "there is an important difference between a more difficult reading and a more unlikely reading" (West 1973: 51).

III. The difficulties posed by the lectio facilior are productive for AI and the Humanities. Strict protocols have been developed in classical philology, to protect difficult, wonderful, creative, non-conforming uses of language on the part of ancient authors. The AI approach we adopt for dealing with scribal errors in Greek can be used also beyond Classics, e.g. to assist writers of English in a manner that minimizes the dangers of linguistic conformism. As ChatGPT has shown, AI can take over the task of writing; our point is that it can also mimic the practices of classical philologists, allowing the writer to write first, then offering editorial support by focusing on unlikely portions of text and suggesting various possibilities for better comprehension, including emendation - with final adjudication on variants not reliant on statistics alone. This AI/philological approach supports conscious and intentional human diversity of expression.


Presenters

Barbara Graziosi, Princeton University; Charlie Cowen-Breen, Princeton University; Creston Brooks, Princeton University; Johannes Haubold, Princeton University



  SCS-9