Colloquium Featuring Dr. Suhas Arehalli

On the Forking of Garden Paths: A case study in large-scale psycholinguistics

Modern developments have allowed for psycholinguists to develop precise broad-coverage models of how processing difficulty surfaces during language processing. As a result, there is a need for precise, model-friendly data through which the predictions of these models can be empirically tested. We investigate one case --- Surprisal Theory (Hale 2001, Levy 2008) --- which posits that a word's processing difficulty can be derived solely from its predictability in context. Advances in natural language processing have allowed researchers to estimate word predictability and thus make predictions about processing difficulty on nearly any set of experimental materials. However, the scale of most experimental data has limited evaluation mostly to just the direction of predicted effects.

In this talk, I'll be exploring a large-scale dataset --- the Syntactic Ambiguity Processing Benchmark --- designed to test surprisal theoretic models on a particular syntactic phenomenon: Garden Path sentences. I'll also discuss two additional projects that demonstrate how this data we collected can be recruited to (1) evaluate potential repairs to surprisal models, as well as (2) facilitate comparisons across different subpopulations. Along the way, we'll explore 3 major questions about Garden Path processing: Is the difficulty associated with Garden Paths reducible to (un)predictability broadly? If not, to what extent can considering structural (over and above lexical) predictability help account for this difficulty? And, finally, to what extent is this garden path difficulty shared between native and non-native processing?

Suhas Arehalli is currently an Assistant Professor of Computer Science at Macalester College. Prior to this, he completed his PhD in Cognitive Science at Johns Hopkins University in 2023. His research interests span the intersection of Natural Language Processing and Sentence Processing, with a focus on how computational models of language can help us understand the mechanisms underlying processing difficulty during online language comprehension.

Subscribe to Colloquium

This is a 2025-2026 Colloquium event. Registration for Colloquium events is not required. There are two ways to receive details and virtual meeting information:

Sign up for the Colloquium listserv: you will receive an email about every event, including virtual meeting information when the event is remote or hybrid

Request information for the next upcoming Colloquium event only