Prediction Biases Sampling

Prediction, Biases, and Sampling Algorithms in Sentence Comprehension

LPP Graduate Seminar

Matt Husband

TT 2024

The past twenty years has seen a sea change in psycholinguistic theorizing about the mechanisms of sentence comprehension. Predictive mechanisms that anticipate the upcoming linguistic signal at a variety of different levels of representation are now well established for language comprehension. Much of the theorizing about these predictive mechanisms has taken place at Marr’s (1982) computational level of analysis, proposing that these predictions reflect incremental probabilistic updating of linguistic representations during the comprehension process (Kuperburg & Jaeger, 2016). How comprehenders execute this updating process algorithmically, however, is currently unclear. The goal of this seminar is start building bridges between our current computational accounts of the comprehension process and the algorithms that realize these processes.

We are not completely in the dark about how to proceed. Questions concerning the algorithmic nature of probabilistic computational processes have been developed and discussed in areas of cognitive psychology outside of the psycholinguistics literature. Perhaps the most well developed are in the domain of probability judgment and decision making. As has been argued for language comprehension, this literature has proposed that human probability judgments are probabilistic computational processes that are argued to be rational, coming close to the Bayesian ideal, especially when the hypothesis space is given and relatively small (Frank & Goodman, 2012; Griffiths & Tenenbaum, 2006, 2011; Oaksford & Chater, 2007; a.o.). However, as the hypothesis space grows or becomes more generally unknown, computation of the Bayesian ideal becomes intractable. Under these more typical everyday conditions, humans tend to depart from rational expectations in predictable ways, often generating only a subset of hypotheses which leading to systematic biases.

Explaining these biases has guided research to models that approximate Bayesian processes at the algorithmic-level of analysis. One class of processes that has proved productive and promising is sampling. Sampling algorithms provide methods for estimating complex distributions. In the limit, different sampling algorithms are indistinguishable as they all converge to the ideal response. However, within resource limitations and resulting small sample sizes, different algorithms display distinct behaviors and biases. This suggests that by discovering the biases that humans have, we can narrow down to different classes of algorithms that are known to share these biases.

While this approach has been fruitful in cognitive domains like probability judgment, surprisingly little research has investigated sampling algorithms in the context of sentence processing (cf. Levy, et al., 2008; Hoover, et al., 2023). We will wrap up the seminar with some initial ideas and directions for psycholinguistics, considering how an empirical exploration of biases in sentence comprehension might relate to different classes of sampling algorithms.

Week 1: Prediction in Sentence Comprehension

Week 1 sets the stage for our current understanding of predictive mechanisms in sentence comprehension as it is currently understood theoretically as incremental belief (Bayesian) updating. We will examine some evidence linking metrical like Bayesian surprise and information-theoretical surprisal to measures of processing time and neural activity, focusing on word predictability as a target domain.