Fifth year PhD student Suhas Arehalli, received a Distinguished Paper Award for his paper at the Conference on Computational Natural Language Learning (CoNLL) 2022. The paper, Syntactic Surprisal From Neural Models Predicts, But Underestimates, Human Processing Difficulty from Syntactic Ambiguities, is a collaboration with Tal Linzen (New York University) and Brian Dillon (University of Massachusetts).

Congrats, Suhas!

Prominent theories of human language processing suggest that the we can determine how much difficulty we’ll have reading a word in a sentence by determining how (un)predictable that word might is, but experimental evidence has found certain kinds of sentences where predictability underestimates difficulty. This work investigates one hypothesis for why this could be: the way we typically measure predictability to get that experimental evidence — using neural network models trained to predict words — underestimates the importance of the structure of language.