At this week’s MCQLL meeting, Jacob Hoover will be presenting: “When unpredictable doesn’t mean difficult”. Abstract below.
We will be meeting this Tuesday November 21st at 3:00PM. Meetings are held both in person in room 117 of the McGill Linguistics department and on zoom.
Abstract:
When humans process linguistic input, we do so faster and with less effort when it better matches what we expect about the intended meaning. The relationship between a comprehender’s expectations about what is going to be said and their processing difficulty has been extensively studied in the computational psycholinguistics literature, documenting a robust correlation between a word’s difficulty (measured by behavioural psychometrics such as reading time) and its surprisal (defined as log[1 / Pr(word | context)], which can be estimated from statistics of language use). One important justification for this relationship is that, under certain simplifying assumptions, surprisal is equivalent of the comprehender’s Bayesian belief update D_{KL}(posterior || prior) about meanings, incurred upon observing the word. Less predictable words tend to incur larger updates.
However, the equivalence between surprisal and belief update size doesn’t necessarily hold in general. In particular, processing of production errors (such as typos or spelling errors in written language) provide an intuitive source of counterexamples. I propose that such items, despite their being extremely unpredictable, often do not cause commensurately large processing effort.
In this presentation I will be discussing a pilot study I am designing to explore these predictions, and welcome feedback on this preliminary work.
I will be presenting remotely.