Through this work, I've built up a lot of zeal for (py)ACT-R as a still-useful tool for algorithmic-level modeling (not just for cue-based retrieval!)---if you want to chat, or take a look at some tutorial assignments I've put together, don't hesitate to get in touch! (or say hi at HSP!)
Of course, the model shows some faults when it comes to fitting all the human data, but the part that most excites me here is the possibility of taking a hypothesis ("people use general skill learning to get situated in pragmatic tasks") and generating an explainable model with testable predictions.
Line plot titled "Empirical relationship with persistence", showing responses changing for individuals with higher persistence values. On the x-axis is "Participant persistence sextile", from 1 to 6, and on the y-axis is "Proportion of target selections", from a dotted line at 50% up to 100%. Three lines are plotted. A line for "Unambiguous" trials stays at 100% for all sextiles, a line for "Simple" trials grows from 60% to 80% across sextiles, and a line for "Complex" trials grows from 50% to 65%.
We model a default-literal reasoner that learns task-specific utility of implicatures on-task—a simple way to explain changes in responses as an experiment progresses.
Cool part: correctly predicts that params tracking individual variation in on-task learning also track diffs in pragmatic tasks!!
This paper with Sasha Mayn and Vera Demberg was the main project of my post-doc at @lst.uni-saarland.de! Really proud of it, and happy to have it out in @openmindjournal.bsky.social.
Headline: A time-conserving ACT-R agent does a nice job explaining low implicature rates in communication games.
Bonus example: In my diss, I found a Maze slowdown aligned with the classic "subordinate access" effect on homonyms. I interpreted it then as just a target processing cost... but was it? Interestingly, diffusion models agree: cases of slowdowns without extra errors are well-fit as slower t0. Cool.
Beyond just proof-of-concept, this is a cool way to see how Maze data comes from various processes, all affected by context. If your theory cares whether your Maze slowdowns come from target processing vs. foil processing vs decision-making, consider this modeling approach to get some insight!
Heading says "faster non-decision time and slower v: many more errors and noise, but just a subtle increase in time". There are two plots below. On top is an unrelated-foil condition, where diffusion trajectories start late but progress fairly quickly towards the top boundary for a target response. Below is a related-foil condition, where diffusion trajectories start earlier but are very noisy and slow, many of them end at the bottom boundary for a foil response.
Some simple max-lik fitting (nice tutorial: cran.r-project.org/web/packages...), and we can see the best fit to Laura's data indeed comes from models where context-related foils lead to lower drift rate---but ALSO faster non-decision processing! These foils seem harder to ignore AND faster to read.
Two plots are compared, featuring collections of many diffusion-model trajectories overlayed on top of each other to create distributions of expected RTs and responses. The top plot is labeled "Harder decisions: lower drift rate" and has more errors and slower RTs. The bottom plot is labeled "Easier decisions: higher drift rate" and has fewer errors and faster RTs.
Enter Ratcliff's diffusion model of decision-making, tried-and-true and easy to work with in R (thanks to @singmann.bsky.social and the rtdists package). Harder decisions can be modeled as lower drift rates, which have a joint effect, widening the distribution of expected RTs and increasing errors.
Seems like both effects come from the foil making the decision harder. But we can go beyond "seems like", cognitive psych has a long tradition of modeling how decision difficulty affects both RTs and accuracy! We wanted to see if these patterns work with those models.
Graphic demonstrating two trial types where sentences are seen in word by word self-paced reading, including the phrase "is an early bird". On top, there is a later Maze decision between the words "attend" or "fly". This decision had an average RT of 1377ms and an error rate of 18%. On the bottom, the decision is between "attend" and "cry", with average RT of 1269ms and error rate of 3%.
We focus on how Maze measures can be affected by the difficulty of deciding between target and foil. In Laura's dissertation, she found a cool pattern: when you use a foil associated with the literal meaning of a prior metaphor ("early bird" → "fly"), you get slower RTs, and more errors.
Title slide from a talk at AMLAP 2025 in Prague. Title is "Jointly modeling maze RT and accuracy using diffusion models: A first case study", by authors Jack Duff (UCLA) and Laura Pissani (Saarland). There is a graphic of a bird facing a series of diffusion model trajectories deciding between the words "attend" and "fly".
Sad to be missing AMLaP, home with some surprise COVID, but very thankful to my co-author Laura Pissani for going solo on our Saturday talk on diffusion models for the Maze task! For anyone else stuck at home, let me tell you here about what we've learned by modeling Maze RTs as decision-making.
awww hey now
... and that's not to mention the prospect of joining a department packed with thoughtful, curious, and imaginative linguists. I couldn't be more excited.
Meanwhile, recommendations are fully open for favorite LA spots, haunts, and locales!
This year, as the opportunities and support to do meaningful science in the US shrink day after day, I feel *beyond* fortunate to have this news. Especially since it brings a chance to double down on community-oriented work, + especially work to help serve 200k+!!! local speakers of Oaxacan langs.
Overjoyed to share that this coming fall, I'll be joining UCLA Linguistics as an assistant professor of psycholinguistics!
a plea to think carefully about surprisal + what it means to understand how we understand >> link.springer.com/article/10.1...
brand new paper in Computational Brain and Behaviour with @andreaeyleen.bsky.social at @mpi-nl.bsky.social
NYTimes caption reading: "Top photo of Jansen at the plate for the Red Sox in a game he was the listed hitter for the Blue Jays before first pitch: Paul Rutherford / Getty Images"
Gorgeous example of a preposition pruning error in what would have already been a confusing and wonderful RC. (Context: www.nytimes.com/athletic/572...)
And yet it's so oddly productive! Even "Are we any of us..."
If you'd be interested in attending the public portion of the defense to hear more, do reach out, there will be a remote option! In the meantime: many thanks to my committee, and everyone that's helped me along the way. (Yes, I am frantically writing my acknowledgments.)
I (begin to) argue for a picture where comprehenders select a single candidate interpretation on varying timelines, depending on the utility/risk of a rapid decision, given the type of ambiguity and their task/goals. Thankfully, lots more work to do to develop this further!
The dissertation—"On the timing of decisions about meaning during incremental comprehension"—is about, er, that. It aims to set up a wide view of interpretive garden paths, and when they do/don't show up, incl. with lexical ambiguity, scalar implicature, and causal inferences.
∃x & news(x) & |x| = 2:
(1) Next Friday, Dec 1, I'm defending my dissertation (!)
(2) In Jan, I will be moving to Germany (!) to join Vera Demberg's lab at Saarland as a post-doc, studying (& modeling) individual differences in pragmatic processing Couldn't be more excited!