Unfolding Predictions in Semantic Interpretation

From IMC wiki
Jump to: navigation, search

Insights from 400 Blindness to Thematic Role Reversals:

Wing-Yee Chow, University of Maryland, 16/01/2013 (supervisor Colin Phillips)

ERPs in processing sentences in real time (bloody psycholinguists)

Comprehension only e.g. like JP stuff...

How do we make predictions - algorithms (recipes) for

(nobody cares about why?)

  • Ingredients:
    • lexical items
    • word order and sentence structure
    • compositional meaning
    • world knowledge
  • Steps:
    • How do they unfold in real time?
    • parallel/serial
    • interactions between different sources of info?
    • changes in impact of a piece of info

N400 "blind" to thematic role-reversals

  • N400 amplitude sensitive to predictability (cloze probability) of word
    • more predictable = smaller
  • But: no N400 comparing:
    • John wondered which thief the cop arrested
    • John wondered which cop the thief arrested
  • is a P600 though
  • Claim: structural role information not being used at this point
  • NB: use of BA in Chinese version means you know unambiguously which is the subject before you get to that point
    • Orphan BA philanthropist adopt
    • Philanthropist BA orphan adopt

Experiment 1 - Chinese

  • Target verb may not be predictable in canonical sentence
    • Use cloze probabilities to differentiate low versus highly predictable verbs
    • ERPs from start of onset of critical verb and then for 1 second (Hmmm)
    • No N400 even with highly predictable

Experiment 2 - Chinese

  • Time for prediction
    • putting extra stuff between NPs and verb so structural role information has time to be processed
    • N400 effect now happens. Whoopdedoo
    • How good is the control comparison (uses time phrase either at start of phrase or between NP and verb)
      • Silence might be a confound so not done.

Interim Summary

  • Love the nods to time linear presentation.
  • Off-line stuff doesn't fully map to real-time predictive computations (whatever it is that cloze tasks typically measure, anyway)
  • Mechanisms?
    • Earlier predictive computations only have access to lexical semantic info
    • Later, structural role info also incorporated...
  • Link to Garnsey et al 1989 (study looking at filler-gap dendency)

Experiment 3 - English

  • Manipulates argument role:
    • The owner forgot...
      • which customer the waitress served
      • which waitress the customer served
  • and argument identity
    • The secretary confirmed...
      • which illustrator the author had hired
      • which readers the author had hired
  • No N400 in role reversal
  • Yes N400 in argument identity
  • So different types of info impact predictive computations at different timescales

New questions

  • How and when does XXX impact predictive computations
    • (case, gender, number negation etc)
  • Challenges to language learners/impaired speakers/bilingual speakers etc
    • Plans to extend paradigm to show how comprehenders update predictions in real time.
      • Using connectives (e.g. nonetheless)
  • How does this stuff link to eye-tracking experiments (seems to contradict it? e.g. Dambacher & Kliegl 2007)
    • Problems with timing comparisons
      • Cumulative progression analysis Scheepers et al (unpublished)
      • Is this like that Daniel Richardson thing (windows of matching?)

Conclusions

Why do I have a sense that this is not actually going to further the undertanding of language-in-use in any meaningful way? I think it's the same as the objection to P&G. All this predictability stuff, okay, so we can do it and we use slightly different information at slightly different times, but perhaps what you predict depends on why - and that must be to have a head start in producing something that is an apropriate response at an appropriate time, right? So how do N400s or lack thereof matter??

What would DS say? (are outstanding requirements resolved in some specific order? The lexical stuff has been accessed so is available; the structure bit is being built so isn't necessarily?)

Q: Richard Breheny says: isn't there a debate on the N400 effect about it being predicting the word form (parallels with studies with auditorily presented data?)

Q: (How) do these things work in production? (what is the evidence from those tasks where you have to do both a predictable word and a non-predictable one that people like to do on patients with schizophrenia etc)