Phonological (un)certainty weights lexical activation Laura - PowerPoint PPT Presentation

Phonological (un)certainty weights lexical activation Laura Gwilliams , David Poeppel, Alec Marantz & Tal Linzen 7th January 2018 1

big ballet blind based on the Cohort model of spoken word recognition: Marslen-Wilson, 1978 bath baptist bond band ballot book b break band black bind boast balance back

ballet balance band bath b a band baptist ballot back based on the Cohort model of spoken word recognition: Marslen-Wilson, 1978

ballet balance b a l ballot based on the Cohort model of spoken word recognition: Marslen-Wilson, 1978

balance b a l ə ballot based on the Cohort model of spoken word recognition: Marslen-Wilson, 1978

balance b a l ə n based on the Cohort model of spoken word recognition: Marslen-Wilson, 1978

But what about ambiguity? Real world speech is noisy and ambiguous ; there is not a • direct mapping between speech and phonemes b p p b p b b p Laura Gwilliams | CMCL | January 2018 7

pin ballet prove bath pacify bond palate beef book p b pants balance paddle bind boast poke panda

pin pin ballet ballet prove prove bath bath pacify pacify bond bond palate palate beef beef book book p b b p pants pants balance balance paddle paddle bind bind boast boast poke poke panda panda

pin pin ballet ballet prove prove bath bath pacify pacify bond bond palate palate beef beef book book p b p b pants pants balance balance paddle paddle bind bind boast boast poke poke panda panda

pin ballet prove bath pacify bond palate beef book p b b p pants balance paddle bind boast poke panda

ballet bath pacify palate p b a pants balance paddle panda

ballet palate p b a l balance

palate p b a l ə balance

p b a l ə n balance

Two Computational Models = phoneme a = acoustic input SWITCH-BASED ACOUSTIC WEIGHTED 1 cohort of words 1+ cohort of words • • binary acoustic term continuous acoustic term • • Laura Gwilliams | CMCL | January 2018 16

Research Question Does acoustic-phonetic uncertainty weight activation at the lexical level? Laura Gwilliams | CMCL | January 2018 17

Prediction aids speech comprehension The brain predicts future linguistic content in terms of • phonemes, morphemes, words and syntactic structures When input is predictable , it is easier to process; reflected • as a relative reduction in neural amplitude x brain response x x x x x x x x x predictability Laura Gwilliams | CMCL | January 2018 18

Quantifying predictability • Surprisal : Probability of an outcome • Entropy: Uncertainty over future input Laura Gwilliams | CMCL | January 2018 19

Critical Variables • Surprisal : Switch-based Acoustic-weighted • Entropy: Switch-based Acoustic-weighted Laura Gwilliams | CMCL | January 2018 20

Stimuli = .75 = .25 Acoustic weighted: = 1 = 0 Switch-based: b b p p b p p b parricade barricade barricade parricade -40 10 .95 Power/frequency (dB/Hz) .75 Frequency (kHz) -60 8 -80 6 -100 4 -120 2 -140 0 Laura Gwilliams | CMCL | January 2018 21

Protocol + palate + + ∞ ms 500 ms < 2000 ms Laura Gwilliams | CMCL | January 2018 22

Procedure & Analysis averaged 200:250 ms (1) x 25 208 sensors (3) time (ms) (2) Laura Gwilliams | CMCL | January 2018 23

Procedure & Analysis Laura Gwilliams | CMCL | January 2018 24

Model Setup • Control variables : • Critical variables : phoneme latency (ms) phoneme latency (number of phonemes) trial number acoustic-weighted entropy block number acoustic-weighted surprisal stimulus amplitude switch-based entropy phoneme pair ambiguity switch-based surprisal Laura Gwilliams | CMCL | January 2018 25

Results Acoustic − Weighted 8 Switch − Based 6 Chi − Squared * * 4 ‘ ‘ n.s n.s n.s n.s n.s 2 n.s 0 2 3 4 5 6 Phoneme Position Laura Gwilliams | CMCL | January 2018 26

Discussion Fine-grained acoustic information does weight lexical candidates • There is a dynamic interaction between different levels of linguistic • description: phonological <-> lexical Not a single heuristic applied in all situations: perhaps reflects that • the brain commits to an interpretation of the phonological category after a certain period of time Acoustic − Weighted 8 Switch − Based 6 Chi − Squared * * 4 ‘ ‘ n.s n.s n.s n.s n.s 2 n.s 0 2 3 4 5 6 Phoneme Position Laura Gwilliams | CMCL | January 2018 27

Research Answer Acoustic-phonetic uncertainty can weight activation at the lexical level Laura Gwilliams | CMCL | January 2018 28

laura.gwilliams@nyu.edu @GwilliamsL With big thanks to: My supervisors, Alec Marantz and • David Poeppel , as well as everyone in the Neuroscience of Language Lab and Poeppel Lab ! Funding: G1001 Abu Dhabi Institute