2016 Annual Meeting Till Poppels & Roger Levy (UCSD) - - PowerPoint PPT Presentation

▶

Oct 12, 2022 471 likes •719 views

2016 Annual Meeting Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12 The father bought his son for a bicycle. literal non-literal Was something bought for the son? 100% 0% No

SLIDE 1

2016 Annual Meeting

SLIDE 2

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

SLIDE 3

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

The father bought his son for a bicycle.

“No” “Yes” 0%

literal non-literal

Was something bought for the son?

100%

SLIDE 4

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

The father bought his son for a bicycle.

“No” “Yes” 33% 66%

non-literal literal

Was something bought for the son?

SLIDE 5

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

The father bought his son for a bicycle. Was something bought for the son? literal non-literal 33% 66% The cook baked Lucy for a cake. Was something baked for Lucy? literal non-literal 47% 53% The apprentice fetched a hammer the carpenter. Was something fetched for the carpenter? literal non-literal 33% 67% The bartender poured the customer for a drink. Was something poured for the customer? literal non-literal 21% 79% The man ordered his girlfriend for some champagne. Was something

rdered for the

champagne? literal non-literal 33% 67% The charity built a house the hurricane victim. Was something built for the hurricane victim? literal non-literal 25% 75%

SLIDE 6

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

𝑸 𝑵 𝑱

Speaker error Environmental noise Listener error Levy (2008) Anderson (1990);

SLIDE 7

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

𝑸 𝑵 𝑱

Speaker error Environmental noise Listener error

Plausibility

∝ 𝑸 𝑱 𝑵 𝑸(𝑵)

SLIDE 8

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

DO/PO benefactives

The father bought his son for a bicycle. The father bought his son a bicycle. The father bought a bicycle for his son. The father bought a bicycle his son.

𝑄 𝐽 𝑁 = 1 𝑄 𝐽 𝑁 = 1 𝑄 𝐽 𝑁 = 1 𝑄 𝐽 𝑁 = 1

1 insertion 𝑄(𝐽|𝑁) = 𝑄(𝑗𝑜𝑡: 𝑔𝑝𝑠) 1 deletion 𝑄(𝐽|𝑁) = 𝑄(𝑒𝑓𝑚: 𝑔𝑝𝑠)

Was something bought for the son? Yes No

Transitive/Intransitive

The t-shirt shrank the dryer. The t-shirt shrank inside the dryer. The dryer shrank the t-shirt. The dryer shrank inside the t-shirt.

𝑄 𝐽 𝑁 = 1 𝑄 𝐽 𝑁 = 1 𝑄 𝐽 𝑁 = 1 𝑄 𝐽 𝑁 = 1

1 deletion 𝑄(𝐽|𝑁) = 𝑄(𝑒𝑓𝑚: 𝑗𝑜𝑡𝑗𝑒𝑓) 1 insertion 𝑄(𝐽|𝑁) = 𝑄(𝑗𝑜𝑡: 𝑗𝑜𝑡𝑗𝑒𝑓)

Did the dryer shrink something? Yes No

Active/Passive

The girl was kicked by the ball. The girl kicked the ball The ball was kicked by the girl. The ball kicked the girl.

𝑄 𝐽 𝑁 = 1 𝑄 𝐽 𝑁 = 1 𝑄 𝐽 𝑁 = 1 𝑄 𝐽 𝑁 = 1

2 insertions 𝑄(𝐽|𝑁) = 𝑄 𝑗𝑜𝑡: 𝑥𝑏𝑡 𝑄(𝑗𝑜𝑡: 𝑐𝑧) 2 deletions 𝑄 𝐽 𝑁 = 𝑄 𝑒𝑓𝑚: 𝑥𝑏𝑡 𝑄(𝑒𝑓𝑚: 𝑐𝑧)

Did the girl kick something? Yes No

𝑸 𝑵 𝑱 ∝ 𝑸 𝑱 𝑵 𝑸(𝑵)

SLIDE 9

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

implausible plausible implausible plausible implausible plausible

% literal responses

0% 100%

Active/Passive Transitive/Intransitive DO/PO

SLIDE 10

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

implausible plausible implausible plausible

implausible

plausible

% literal responses

0% 100%

Active/Passive Transitive/Intransitive DO/PO

Insert/delete: 1 preposition Insert/delete: “for” Insert/delete: “by” and “was”

The package fell to the table from the floor.

Structure-sensitive noise inference: undoing exchange errors

The package fell to the table from the floor.

from to

SLIDE 11

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

The package fell to the table from the floor.

cf. “spoonerisms” (e.g. MacKay, 1970)

Waste the term  Taste the werm Fighting a liar  Lighting a fire Battle ships and cruisers  Cattle ships and bruisers Busy Dean  Dizzy bean

Structure-sensitive noise inference: undoing exchange errors

The package [VP fell [PP to the table] [PP from the floor]].

SLIDE 12

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

The package [VP fell …]

[PP to the table] [PP from the floor] [PP to the floor] [PP from the table] [PP from the floor] [PP to the table] [PP from the table] [PP to the floor]

implausible plausible non-canonical canonical 97% 3% 5% 95% Plausibility Norming Canonicality Norming

[PP from …] [PP to …] 95% - 5% [PP to …] [PP from …] [PP with …] [PP about …] 80% - 20% [PP about …] [PP with …] [PP to …] [PP about …] 81% - 19% [PP about …] [PP to …] [PP from …] [PP about …] 67% - 33% [PP about …] [PP from …] [PP for …] [PP in …] 51% - 49% [PP in …] [PP for …] [PP in …] [PP at …] 58% - 42% [PP at …] [PP in …] [PP to …] [PP for …] 97% - 3% [PP for …] [PP to …]

response ~ plausibility + canonicality + (1 + plausibility + canonicality || item) + (1 + plausibility + canonicality || subject)

SLIDE 13

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

𝑸 𝑵 𝑱

Speaker error Environmental noise Listener error

Plausibility

∝ 𝑸 𝑱 𝑵 𝑸(𝑵)

Canonicality

Predictions

1. Noise inference whenever

prior probabilities permit

2. Additive effects of

plausibility and canonicality

SLIDE 14

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

𝑸 𝑵 𝑱

Speaker error Environmental noise Listener error

Plausibility

∝ 𝑸 𝑱 𝑵 𝑸(𝑵)

Canonicality

Predictions

1. Noise inference whenever

prior probabilities permit

2. Additive effects of

plausibility and canonicality

SLIDE 15

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

𝑸 𝑵 𝑱

Speaker error Environmental noise Listener error

Plausibility

∝ 𝑸 𝑱 𝑵 𝑸(𝑵)

Canonicality

implausible non-canonical implausible canonical % literal responses

0% 100%

plausible non-canonical plausible canonical * * * Predictions

1. Noise inference whenever

prior probabilities permit

2. Additive effects of

plausibility and canonicality

SLIDE 16

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

[PP to the table] [PP from the floor] [PP to the floor] [PP from the table] [PP from the floor] [PP to the table] [PP from the table] [PP to the floor]

implausible plausible implausible plausible

implausible

plausible

% literal responses

0% 100%

Active/Passive Transitive/Intransitive DO/PO

SLIDE 17

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

[PP to the table] [PP from the floor] [PP to the floor] [PP from the table] [PP from the floor] [PP to the table] [PP from the table] [PP to the floor]

implausible plausible implausible plausible

implausible

plausible

% literal responses

0% 100%

Active/Passive Transitive/Intransitive DO/PO implausible non-canonical implausible canonical plausible non-canonical plausible canonical Exchanges * * Swapping nouns in active/passive? The ball was kicked by the girl. What’s the difference?

Function vs. content words?

Opposite pattern in spoonerisms. (MacKay, 1987)

Adjuncts vs. Complements?

Possible, but speculative. Interim Summary

We know that prepositions can be

exchanged.

We don’t know that nouns can’t

be exchanged.

Why exchanges don’t occur in

active/passive sentences is an

pen question.

SLIDE 18

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

“That’s not the right kind of process, intuitively.”

Structure-sensitive noise inference: undoing exchange errors

1. Do people REALLY consider

all conceivable interpretations during language comprehension?

“That’s not a computationally feasible mechanism.”

SLIDE 19

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

“That’s not the right kind of process, intuitively.”

Structure-sensitive noise inference: undoing exchange errors

1. Do people REALLY consider

all conceivable interpretations during language comprehension?

“That’s not a computationally feasible mechanism.”

SLIDE 20

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

“That’s not the right kind of process, intuitively.” “That’s not a computationally feasible mechanism.”

Structure-sensitive noise inference: undoing exchange errors

1. Do people REALLY consider

all conceivable interpretations during language comprehension?

2. If we open the door to non-literal

interpretations, does that mean that anything goes? What about: “The cat is on the mat.”

Marr (1982) “In order to understand bird flight, we have to understand aerodynamics; only then do the structure of feathers and the different shapes of birds’ wings make sense.

SLIDE 21

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

Noise inference occurs whenever (and to the

extent that) literal interpretations are unlikely

Replicated results with active/passive,

transitive/intransitive, and DO/PO materials

Comprehenders undo exchange errors
Utterance priors driven by content and form

Structure sensitivity!

Comprehenders’ noise model exhibits

SLIDE 22

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

SLIDE 23

Till Poppels & Roger Levy (UCSD) Structure-sensitive noise inference: undoing exchange errors / 12

[- plausible] [- canonical] [- plausible] [+ canonical] [+ plausible] [- canonical] [+ plausible] [+ canonical] FILLERS