Brian Dillon
University of Massachusetts, MA·Psychology
How do people pull meaning out of a sentence as they read it, and when does that process break down? Sentences are often temporarily ambiguous: the opening words fit more than one interpretation, and only later does it become clear which one was intended.
- Fields
Psychology
Computer Science
Neuroscience
- Topics
psycholinguistics
sentence processing
semantic ambiguity
- Grants
- 1 known grant, through Aug 2029
What they work on
- Explain why a language model's perplexity, its skill at predicting the next word, correlates positively with worse fit to human reading times, by comparing model predictability estimates against reading-time data.
- Test whether semantic training objectives, which reward building meaning rather than only predicting words, bring language models' processing closer to that of human readers.
- Use causal intervention techniques that manipulate a model's internal representations to focus its semantic context on a small number of candidate interpretations of an ambiguous sentence.
- Run human reading experiments on novel psycholinguistic stimuli to test whether the modified models correctly predict where and how much comprehension difficulty people experience.
More about the research
Human readers can struggle badly when a later word forces them to revise their reading. Large AI language models, the technology behind systems like ChatGPT, generally do not show this difficulty. That mismatch matters. If these models processed language more like people do, they would be better tools for studying comprehension and for addressing language processing difficulties, and the gap also connects to a known oddity: models need far more language data than a human learner does.
Why it matters
Better alignment between AI language models and human sentence processing could make these models more useful as scientific models of comprehension and could inform work on language processing deficits. It may also point toward architectures that learn language from far less data than current models require, addressing a long-standing limitation of the technology.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Tal Linzen
New York University, NY·Computer Science·Early-career
Explain why language models that predict the next word better match human reading times worse, by comparing model predictability estimates with human reading data.
Satoshi Tomioka
University of Delaware, DE·Social Science
Test whether the difficulty of alternative word orders comes from filler-gap dependency formation, the effort of linking a displaced phrase back to the position it moved from, using self-paced reading across several sentence types.
Eszter Ronai
Northwestern University, IL·Psychology
Identify the real-time processing signals of focus, the marking of new or emphasized information, in languages with fixed versus flexible word order, using eye-gaze data and statistical modeling.
The school
University of Massachusetts: 28 labs on PI Pages, $282M a year on research, 337 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.