Subjective: Real-world text analysis with small language models

Xanda Schofield
Harvey Mudd College

Stevenson 1300
12:00 PM

Large language models are all the rage today - but what about small language models? There are lots of experts in specific text domains working with more text than they could possibly read, from scholars of literature and history to customer sales representatives processing complaints. In this talk, I'll share about a particular type of small language model that these folks have used for a couple of decades now: the statistical topic model. We'll explore the kinds of conclusions we can (and can't) draw from these simpler models, common pitfalls that arise in using them, and some lessons we might take away when working with these models' much bigger transformer-based cousins. Throughout, we'll revisit the theme of where subjectivity fits in computer-assisted text analysis.