4 thoughts on “It takes two the tango: Predictability and detectability affect processing of phrase structure errors

  1. Hi Anthony, thanks for an interesting talk! Two questions: 1) I may have missed your saying this, but I assume that ERPs were timelocked to the onset of the violation (of/the)? In normal speech, these words would be very short, so the ERP is probably overlapping with the onset of the next word – were the post-critical regions matched? 2) The way you described the paradigm made it sound like you were comparing e.g. “the” in the expected condition with “for” in the violation condition. Assuming that once people encounter the violation, they adjust their expectations for further upcoming sentence material, then a preposition might allow/predict a wider range of expectations/repairs (i.e. a PP with various nodes/layers) than a determiner (i.e. noun or adj+noun). Could this have affected the ERP at the violation? Apologies if I’ve misunderstood the design!

    1. Hi Kate,

      Thanks for the questions!

      1) Yes, the ERPs were time-locked to the onset of the violation or correct word for each sentence. We spliced our target words into the larger recording, so the context on both sides of the target words were identical. I definitely agree that these words are extremely short—we constantly go back and forth between thinking about using nouns/verbs for this reason. However, we are also interested in the fact that these words *are* so short and possibly more difficult to detect when processing a discourse.

      2) Interesting point! One way to test this would be to ask participants online to complete the sentences with the target word and the violations. I suspect that you are correct and that people would be able to recover the sentence more readily after encountering an unexpected preposition than vice versa. Then, using these data we could see how the ERP responses differ between sentences that are recoverable and those that are not. I would suspect that in sentences that are recoverable, there would be another error signal when the sentence continued in the way that it was meant to—as the sentence did not adjust after hearing the violation.

      I hope these responses answer some of your questions! I am happy to continue chatting about this—thanks so much!

      1. They do, thanks! Re 2), you could even perhaps break down your current ERP data into preposition violations vs. determiner violations… I was thinking maybe the ERP would be larger in magnitude for the preposition ones, as being more readily recovered would trigger more processing and associated cost (vs. when you have fewer options to recover). Just a thought. Good luck with it!

  2. Hi Anthony, very cool stuff.
    Your second possible interpretation about why there was no P600 in the detectable, predictable trials is interesting. If I remember correctly, you said it is because the errors were too small that listeners didn’t bother correcting them. A recent preprint paper by Rachel Ryskin and colleagues however showed that word neighbor errors (e.g., anecdote & antidote, which I think can be regarded as small error as well) elicited a P600 effect. They also didn’t explicitly probe detection, although they used less naturalistic, sentence-by-sentence paradigm. I wonder what you think about the difference between yours and theirs, because I really would like to know what P600 means in these literature. Does it reflect detection of syntactic categorical error, attempt to correct, or successful correction?

    By the way the last read-out question about why the null effect of detectability in the predictable trials was from me. I think the next step you mentioned is a good way to go. Our lab has been analyzing eye movements contingent on trial-based detection, and found the same results as yours; when participants behaviorally did not detect the error, they showed no disruption on the eye movements.

    Finally, do you think noisy-channel inference played a role in your experiment?

Leave a Comment or Question Below