Research

Words vastly underdetermine meaning, yet communication mostly succeeds. My research explains how, with general principles of cognition and communication: listeners infer what a speaker meant through rational, Bayesian inference over a noisy channel, and speakers shape their utterances so that this inference succeeds. Because such principles should hold for all people, I test them across typologically diverse languages, with behavioral, computational, and neural methods.

Lines of work

How listeners infer what was meant

Listeners combine the signal they perceive with prior expectations, and those expectations concern structure as well as meaning. In Russian and Hindi, where several word orders are grammatical, people reinterpret a plausible sentence when its word order is rare. The same inference explains misheard song lyrics: “laid him on the green” becomes “Lady Mondegreen” partly because green is rarely a noun. With Ev Fedorenko’s lab, I am finding that the brain’s language network responds to degraded input according to how much well-formed input surrounds it.

JEP:LMC 2026 · JEP:LMC 2024 · CogSci 2024 · Review in preparation

How speakers make informative words stand out

Speakers say more informative words longer, louder, and at a higher pitch. In conversations recorded over Zoom, a word’s surprisal predicted all three, in every speaker, and the same holds in phone conversations in 22 languages. In production experiments across 17 languages from 11 families, corrected words (“No, ANYA lifted Joey”) are enhanced in every language, including languages said not to use prosody this way. Undergraduate researchers ran these experiments in their native languages.

Cognitive Science 2025 · HSP 2023

How speakers order words for the listener

Early in a sentence the listener has little context, so an informative word placed there is easy to misunderstand. Speakers should therefore order words from least to most informative. In English and Italian, “The woman gave the dog a bone” sounds most natural when the dog is familiar and the bone is new.

CogSci 2026 · SLE 2025

Old puzzles, communicative answers

Acceptability ratings, a core method from fieldwork to evaluating language models, reflect communication: an error matters less the more of the sentence is well formed, a prediction borne out across 8 languages. The same view explains syntactic “islands” such as “What did John murmur that Mary bought?”, long cited as evidence that grammar is innate. Their acceptability tracks how frequent the verb is and the discourse, not a shared syntactic structure.

Cognition 2025 · Language, under review · Cognition, under review · RAILS 2025

Surveyor

An experiment platform I built in 2023 that has hosted more than 330 studies, including students’ own experiments in 9.59. Its structure teaches design by organizing materials into items and conditions. With an NIH-funded STIRR assistantship (2026–27), I am extending it to select appropriate analyses and generate complete methods reports, in support of rigorous, reproducible research.

mit-surveyor.com

In my own lab, I will ask how communicative principles shape why we say things the way we do, in any language: how listeners learn which corruptions of the signal are likely, from non-native accents to speech disorders and autism; how speakers arrange information within a sentence; how communication adapts to speech, writing, texting, and sign language; and how these principles compare across neurotypical and neurodiverse people.

Collaborations

Work on language and the brain with colleagues in Ev Fedorenko’s and Nancy Kanwisher’s labs: how the language network lateralizes by age four, its cerebellar components, three components of pragmatic language use, and what drives apparent food selectivity in the brain. With Noga Zaslavsky and Roger Levy: how discrete, systematic communication emerges.