• ProjectsSyntacticAdaptationSelfPacedReadingExperiment_EpisodicStatistical

Synopsis and Design

The two rapid adaptation experiments reported in Fine, Jaeger, Farmer, & Qian (submitted) provide evidence that humans rapidly adjust their subjective representation of the statistics of the input in order to more efficiently process language. The study leaves open several questions. One question is, assuming adaptation is implemented via a learning mechanism of some kind, what kind? There are numerous possibilities, but in this experiment we focus on distinguishing two specific proposals. The first proposal is the one outlined in our paper, namely that the type of learning involved in adaptation is at least computationally expressible as Bayesian belief update--each sentence that comprehenders process counts as a piece of evidence that is used to update a (subjective representation of a) probability distribution over syntactic structures. This predicts that if subjects process many main verb (MV) structures such as (1), this should take probability mass away from a competing structure such as a reduced relative clause (RC), like (2).

(1) MV: The experienced soldiers warned about the dangers before the midnight raid.

(2) RC: The experienced soldiers warned about the dangers conducted the midnight raid.

(3) MV (unambiguous): The experienced soldiers spoke about the dangers before the midnight raid.

(4) RC (unambiguous): The experienced soldiers who were told about the dangers conducted the midnight raid.

Presumably this is why garden path effects are observed with sentences like (2) in the first place (subjects have far more experience with sentences like (1)). We have extremely preliminary evidence for this in Experiment 2 of Fine et al. (submitted). In that experiment, subjects see far more RCs than they expect a priori, and relatively few MV structures (few, relative to the number of RCs). By the end of the experiment, the ambiguity effect for RCs (RTs in the bold region in 2 minus RTs in the bold region of 4) had completely disappeared, whereas the ambiguity effect for MVs (RTs in the bold region of 1 minus RTs in the bold region of 3) had increased, numerically though non-significantly.

This pattern of results is predicted by a belief-update model of learning.

By contrast, Kaschak and Glenberg (2004) offer an episodic account of syntactic adaptation which predicts a different pattern of results. Specifically, under their proposal, reading a temporarily ambiguous string like The experienced soldiers warned about the dangers... causes readers to retrieve both structures, the MV and the RC. This leaves an episodic trace for both structures, which in turn facilitates processing of both. Thus, if subjects see a bunch of sentences like (2), sentences like (1) should actually get easier to process, not harder.

This experiment provides a more direct attempt to adjudicate between these two proposals.

MV structures are FAR more common than RCs, so a balanced design in which MVs and RCs are interspersed (like Experiment 2 in Fine et al.) may not provide subjects with sufficient evidence against MVs to detect an actual increase in the processing cost of that structure. In order to address this, we employ a block design (below I refer to blocks as phases), which is outlined in the table below.