Predictive processing in naturalistic second-language comprehension

EEG and computational modeling of how bilinguals predict and retrieve information while listening to The Little Prince in their first and second language (NSF BCS-2518248).

2026– · with Jude T. W. Singleton, Dr. Edith Kaan, Dr. Jonathan R. Brennan, and Dr. Teresa Bajo

The project

How well can a listener anticipate what comes next, and does that ability depend on whether they are listening in their first or second language? This NSF-funded collaboration studies predictive processing in naturalistic comprehension: instead of short, controlled sentences, bilingual participants listen to whole chapters of The Little Prince in English and in Spanish while we record EEG. The story is then annotated with linguistic structure and with word-by-word measures from Transformer language models, so that brain responses can be linked to model-based estimates of expectation and memory demand.

At UF we are collecting data from about 60 L1 English / L2 Spanish listeners. Our collaborators run parallel studies with the same materials: Jonathan Brennan’s lab at the University of Michigan with L1 Mandarin / L2 English speakers, and Teresa Bajo’s lab at the University of Granada with L1 Spanish / L2 English speakers. Comparing within participants across their two languages, and across groups, lets us ask how language experience (exposure, use, proficiency) shapes prediction.

Current status

Data collection at UF is underway.

My role

  • Designed and troubleshot the E-Prime paradigm for bilingual naturalistic listening during EEG.
  • Built the Qualtrics instruments for proficiency testing (LexTALE, MELICET) and language history.
  • Run EEG sessions with participants and wrote the R scripts for initial analysis of comprehension accuracy.
  • Computed language-model surprisal for the Spanish and English stimuli and checked the forced alignment of the audio (TextGrids).

Connection to my dissertation

Prediction is only half of comprehension: listeners also have to retrieve earlier material from memory, for example to work out who a pronoun or a dropped subject refers to. My dissertation asks why L1 and L2 speakers differ in this kind of dependency resolution. Are syntactic representations less robust in a second language, or do L2 speakers weight retrieval cues (such as discourse prominence or gender) differently? I implement each account as a targeted intervention in a neural language model and test which one better predicts L1 and L2 reading times. This project’s naturalistic EEG data let me take the same question from controlled reading experiments to continuous listening. For example, at a Spanish verb with a null subject (Luego, Ø durmió…, “Then [it] slept…”), does a competing discourse referent make retrieval harder, and is that cost larger for L2 listeners?

Presentations

  • Zhang, D., & Kaan, E. (2026, June). Who did it? Referential retrieval in L1 and L2 Spanish naturalistic listening. Poster, Multilingual Minds and Machines Meeting (MMMM 2026), Max Planck Institute for Psycholinguistics, Nijmegen, Netherlands.

Funding

NSF BCS-2518248.

Illustration: Antoine de Saint-Exupéry, The Little Prince.