Stanford Humanities Center
Join us for the next CESTA Tuesday lunch seminar, titled "Metonymic Language Models: Literary Data after AI" by Benjamin Glaser, Associate Professor of English at Yale University. This hybrid talk+workshop contrasts the organization of literary data—and the challenge of theorizing it—with Large Language Models’ implicit, task-determined theory of language. For example, poetry’s sub- and supra-lexical prosodic features (stress, meter, intonation) are not available to LLMs as phonology, but rather correlate to concatenations of words within certain genres (“sonnet,” “poetry”). Metaphors and similes correlate to the concatenations “x is/is like a y” or compounds like “x is a y that z’s.” These approximations of the Jakobsonian “poetic function” reflect training data, instruction tuning, and prompting strategies, but not literary data or aspects of language beyond word selection. What datasets could force attention to different linguistic and literary features, and what impact would (re)training on such data have on the broader behavior and meta-linguistic capacity of generative AI? To reach that point, we also need to determine strong computational methods for asserting the divergence (and perhaps convergence) of human and AI textuality. After an opening discussion and presentation on poetic meter, participants will be invited to consider how their disciplinary objects get projected into LLM data / vector space. We may explore: pre-training data, synthetic data, fine-tuning, prompt engineering, and differences between models. While this talk focuses on text, we might also consider visual, audio, and multimodal models. RSVP for lunch or to receive the Zoom link here.
Note: a pre-circulated paper for the Stanford Workshop in Poetics (also January 14, 3:30) may be of interest (there is overlap).
About the Speaker
Benjamin Glaser's research focuses on the history and practice of prosody, particularly its relationship to modern poetry. Their book project, Modernism’s Metronome, examines debates over prosody from the late nineteenth century through the 1930s, highlighting both the rejection and the enduring influence of traditional meter and scansion practices. They are also the co-editor, alongside Jonathan Culler, of Critical Rhythm, a collection of essays that introduces new approaches to studying poetic rhythm. In addition, they contribute to the development of the Princeton Prosody Archive (1750–1923), an innovative digital collection containing over 10,000 texts related to historical poetics and prosody. The archive aims to enable scholars to pursue new research projects in historically significant but often overlooked discourses of prosody. Their interest in prosody extends to the poetics and politics of Hip-hop and other musical genres. More broadly, their teaching bridges literature and music, with a particular emphasis on sound and its historical mediations.
