Research
One question with several faces: how do people discover the structure of a language from the input they get, and what keeps that structure available once it is learned?
Lines of work
-
Segmentation and cue conflict
Fluent speech has no reliable pauses at word boundaries. Learners can use transitional probabilities between syllables, lexical stress, and language-specific restrictions on which sounds may end a word. In natural speech these cues frequently disagree. We build artificial languages in which two cues are deliberately pitted against each other, then ask which one wins, at what age, and why.
-
Memory, delay and consolidation
A word segmented in the lab can be gone twenty minutes later. We study what infants and adults retain after a delay, when isolated words rescue a representation that statistics alone could not sustain, and how segmentation connects to word learning.
-
Individual differences and multilingual learners
Learners are not interchangeable. Verbal short-term memory, working memory and skill in the native language each predict different parts of performance. We test adults learning from natural speech in a language they do not know, and we work with the bilingual Spanish–English communities of South Texas.
-
Language development in atypical populations
Statistical learning has been proposed as a mechanism behind language difficulties in developmental disorders. We review and test what that claim can and cannot explain.
-
Memory and metacognition with AI
People now get a share of their information from systems that sound authoritative and leave no trace of where the content came from. We study source monitoring, confidence, cognitive offloading, and whether access to these tools helps every student equally.
Current projects
-
When cues collide
Five- and eight-month-olds hearing streams in which transitional probabilities and native-language phonotactics point to different word boundaries. With Megha Sundara and Hironori Katsuda at UCLA.
-
Stress against statistics
A companion set of experiments pitting transitional probabilities against lexical stress, using trochaic and iambic versions of the same artificial language. Materials are synthesized in the lab so that stress is carried by duration alone.
-
Statistical word segmentation in natural speech
Turkish-speaking adults listening to fluent Italian. Verbal short-term and working memory predict learning of high-probability words; native-language comprehension predicts the low-probability ones — a double dissociation in who learns what.
-
MIRAGE
Metacognition and Information Reliance in AI-Generated Encounters. After reading answers that came from a chatbot, from a reference page, or from their own memory, what do people later recall — and do they remember where it came from? With Aylin Özdeş at Clarkson University.
-
Cognitive offloading and attention
Whether handing a task to an AI assistant lowers cognitive load or simply moves it, and whether the answer differs for students with and without ADHD.
How we work
We build our own speech materials. Artificial-language streams are synthesized with MBROLA and edited in Praat, which lets us control duration, pitch and amplitude precisely enough to isolate a single cue.
Adult and online studies run on Gorilla and Qualtrics. Infant testing is carried out with our collaborators’ labs using the head-turn preference procedure. Analyses are done in R and JASP, with a preference for open materials and preregistration where the design allows it.
Students in the lab learn this pipeline end to end: building stimuli, programming a task, running participants, cleaning data, and writing it up.
Interested in collaborating?
We are glad to hear from researchers working on segmentation, memory, bilingualism, or cognition with AI.