Stealing interpretations?�Failed investigations into structural reanalysis

​

Prof E. Matthew Husband

25 January 2023

@ Edinburgh University

Linguistics Society

​

While the man hunted the deer ran into the woods.

​

Did the man hunt the deer?

YES NO

2

Misinterpretation in Garden-path Sentences

Christianson, et al, 2001

3

Misinterpretation

The horse raced past the barn fell.

​

Paraphrase?

​

​

‘The horse raced past and the barn fell.’

Fodor & Inoue (1994)

4

Main Idea

• We can use the products of language comprehension to investigate intermediate stages of language processing.
• Systematic misinterpretation as evidence for certain misanalyses of garden-paths.

​

Road Map

• Garden-paths and misinterpretation
• Adopt the Diagnosis Model of Reanalysis (Fodor & Inoue 1994, 2000)
• Temporary surface constituents (subordinate object/main clause subject ambiguity)
• Temporarily derived constituents (reduced relative clause ambiguity)
• ‘Steal’ as a reanalysis operation
• Three sentence-question studies investigating the comprehension of reduced relative clauses
• Conclusions and future directions

​

5

Diagnosis Model of Reanalysis

• Assumes a Garden Path Model for initial (first-pass) parsing
• Incrementally builds and maintains a single structural parse.
• Ambiguity resolved by Late Closure & Minimal Attachment.

Fodor & Inoue 1994, 2000

6

Parsing Mechanisms

• Attach
• Connect incoming words and constituents to the current syntactic structure such that resulting syntactic structure is grammatically well-formed.
• Ambiguity? Resolve using the following principles:
• Minimal Attachment (assume a minimal amount of additional structure when making an attachment)
• Late Closure (attach to the more recently constructed phrase if grammatically possible)

​

7

Subordinate-Main Construction

8

S

VP

S

NP

S’

NP

the man

NP

the deer

V

hunted

VP

NP

Late Closure

Early Closure

C

While

Subordinate-Main Construction

9

VP

S

NP

S’

NP

the man

NP

the deer

V

hunted

VP

S

C

While

Interpret

∃e ∃x ∃y [hunt(e) & Ag(e,x) & man(x) & Th(e,y) & deer(y)]

Diagnosis Model of Reanalysis

• Assumes a Garden Path Model for initial (first-pass) parsing
• Incrementally builds and maintains a single structural parse.
• Ambiguity resolved by Late Closure & Minimal Attachment.

​

• A repair strategy for reanalysis
• Reanalysis is restructuring and repairing of the initial parse.

​

• The difficulty of reanalysis is related to how well the symptom (the error signal) is linked to the repair (what needs to be reanalyzed).
• Successful reanalysis guides the parser back to the source of the garden-path.

Fodor & Inoue 1994, 2000

10

Repair Mechanisms

• Attach (Anyway)
• When attachment of new word cannot be made in a grammatical way, attach it to least sever attachment site.

​

• Adjust
• When an attachment creates conflict between two nodes, eliminate the problem by minimally altering the one less recently acted on
• Steal
• Allows the parser to pull out constituents from the current structure.
• Reaccess
• Allows the parser to reaccess lexical items.

​

11

Subordinate-Main Construction

12

VP

S

NP

S’

NP

the man

NP

the deer

V

hunted

VP

S

V

ran

C

While

Attach (Anyway)

∃e ∃x ∃y [hunt(e) & Ag(e,x) & man(x) & Th(e,y) & deer(y)]

Subordinate-Main Construction

13

S

VP

ran

S

NP

?

S’

NP

the man

V

hunted

VP

Steal

C

While

NP

the deer

∃e ∃x ∃y [hunt(e) & Ag(e,x) & man(x) & Th(e,y) & deer(y)]

Subordinate-Main Construction

14

NP

?

VP

ran

S

NP

the deer

S’

NP

the man

C

While

V

hunted

VP

Reaccess

S

∃e ∃x ∃y [hunt(e) & Ag(e,x) & man(x) & Th(e,y) & deer(y)]

Subordinate-Main Construction

15

VP

ran

S

NP

the deer

NP

the man

S

S'

C

While

VP

hunted

Interpret

∃e ∃x ∃y [hunt(e) & Ag(e,x) & man(x) & Th(e,y) & deer(y)]

∃e₁ ∃x [hunt(e₁) & Ag(e₁,x) & man(x)]

∃e₂ ∃y [run(e₂) & Ag(e₂,y) & deer(y)]

​

Interpret

Subordinate-Main Construction

• There is a temporary surface state where [the man hunted the deer] is a constituent.
• This appears to be the source of misinterpretation.

Christianson, et al, 2001

16

Reduced Relative Clauses

17

S

NP

The horse

V

raced

VP

S

VP

NP

Op

NP

Minimal Attachment

Reduced Relative Clauses

18

S

NP

The horse

V

raced

VP

PP

past the barn

Attach

Reduced Relative Clauses

19

S

NP

The horse

V

raced

VP

P

past

V

fell

NP

the barn

PP

S

NP

?

Steal

Reduced Relative Clauses

20

S

NP

The horse

V

raced

VP

P

past

V

fell

NP

?

PP

S

NP

the barn

Reaccess

Steal

Interpret

∃e ∃x [fall(e) & Th(e,x) & barn(x)]

Reduced Relative Clauses

21

S

NP

The horse

V

raced

VP

P

past

V

fell

NP

the barn

PP

S

NP

?

∃e ∃x [fall(e) & Th(e,x) & barn(x)]

Reduced Relative Clause

• There is not a temporary surface state where [the barn fell] is a constituent.
• But there may be a temporary stage of processing where [the barn fell] is a constituent!
• Question: Is this a source of misinterpretation?

​

• If yes, this would be strong evidence for Steal as an operation, and for repair models in general.

22

Sources of Misinterpretation

Temporary Surface Interpretations

• Misinterpretation arises from the availability of a temporary surface constitutent.
• Although this constituent is undone during reanalysis, the interpretation it temporarily licensed lingers in memory.

Temporarily Derived Interpretations

• Misinterpretation arises from a constituent that is temporarily entertained during reanalysis.
• Although this constituent is undone during reanalysis, the interpretation it temporarily licensed lingers in memory.

23

Studies

• Pilot Study:
• 2 (Garden-path vs Control) x 2 (Licensed Q [expect Yes] vs Unlicensed Q [expect No])

​

• Follow-up Study 1: Verb/preposition transitivity
• 2 (GP vs Ctrl) x 2 (Licensed Q vs Unlicensed Q) x 2 (Obligatory vs Optional transitivity)

​

• Follow-up Study 2: Semantic association
• 2 (GP vs Ctrl) x 2 (Licensed Q vs Unlicensed Q) x LSA (continuous)

​

24

25

Press spacebar to continue

The wine spilled from the decanter was expensive.

Was the decanter expensive?

Sentence Reading Time

Question Response

and

Question Response Time

Pilot Study

26

• 14 Subjects, 24 Items
• Measured
• Sentence reading time
• Question response and response time

Pilot: Sentence Reading Times

27

Pilot: Question Accuracy

28

Pilot: Question Response Time

29

Summary of Pilot Study

• No evidence for Misinterpretation
• No interaction between Structure and Question on question accuracy (or question response time).

​

• Other potential issues:
• No evidence of a garden-path effect in sentence reading times.
• Could be a trade-off between reanalysis cost (for garden-paths) and sentence length (for controls).
• Any garden-path effect could be swamped by reading time variability (~3.5 ± 2.5 sec).
• Perhaps other factors obscured the expected misinterpretation effect.
• Transitivity of the (relative) clause
• Semantic association between the NP and the main clause verb

30

Follow-up Study 1: Verb/prep transitivity

31

• 25 Subjects, 24 Items
• Measured
• Sentence reading time
• Question response and response time

Follow-up 1 (VPT): Sentence Reading Times

32

Follow-up 1 (VPT): Question Accuracy

33

Follow-up 1 (VPT): Question Response Times

34

Summary of Follow-up 1 (VPT)

• No evidence for Misinterpretation
• No interaction between Structure and Question on question accuracy.

​

• Instead, question accuracy affected by Transitivity.
• Obligatory transitivity increased question accuracy in the garden-path condition.

35

Follow-up Study 2: Semantic association

36

• 24 Subjects, 24 Items (all with Optionally Transitive verbs/prepositions)
• Measured
• Sentence reading time
• Question response and response time

Follow-up 2 (LSA): Sentence Reading Times

37

Follow-up 2 (LSA): Question Accuracy

38

Follow-up 2 (LSA): Question Response Times

39

Summary of Follow-up 2 (LSA)

• No evidence for Misinterpretation
• No interaction between Structure and Question on question accuracy.
• No effect of LSA on question accuracy either.

40

Review of question accuracy

41

• Pilot: Did we find evidence for misinterpretation?
• Nope!

​

• Follow-up 1 (VPT): What if we took transitivity into account?
• Nope!

​

• Follow-up 2 (LSA): What if we measure semantic association between NP and predicate?
• Nope!

​

What does this mean for theories of reanalysis?

• Option 1: Reanalysis is not repair of an existing parse.
• Steal does not form a temporary constituent.

​

• Option 2: Temporary constituents entertained during reanalysis are not interpreted.
• Steal forms the relevant constituent, but this constituent is abandoned before it can be interpreted.
• Such constituents are only interpreted when the parser gives up.
• Perhaps this is what happens in the particular case of “The horse raced past the barn fell.”

42

Other avenues?

• Assuming Option 2 is on the right track, we need to think about other effects to reveal the intermediate stages of reanalysis.

​

• Evidence from memory retrieval?
• Some preliminary evidence that reanalysis may use ‘steal’ from memory retrieval paradigms (Suckow & van Gompel, 2017; Martin & McElree 2018).
• Ongoing research on whether retrieval cues influence garden-path reanalysis (e.g. Fang & Husband 2022 HSP Short Talk).

​

43

Some lessons

• Sometimes really clear theoretical ideas fail to bare empirical fruit.

​

• When to call it quits?
• Probably not when the first study fails.
• But when several studies all point against the idea, it may be worth reevaluating.

​

• Reevaluation may then lead to better outcomes.
• Theorizing about ‘steal’ as memory retrieval may yet prove useful for theories of reanalysis.

44

Many thanks to several groups of Final Honors Schools students from 2018, 2019, and 2020

Thank you!

45

46

Reduced Relative Clauses

47

S

NP

The horse

V

raced

VP

P

past

V

were

NP

the barns

PP

S

NP

?

Cue-based Retrieval

[+N]

[+pl]

Partial match

Cue-based Retrieval

Number Attraction and RRCs

48

• 72 Subjects, 40 Items
• Measured
• Self-paced reading time

Number Attraction and RRCs

49

31 msec

t=1.909; p=.060

6 msec

t=0.514; p=.609

Str: t=5.749; p<.001

Agr: t=1.832; p=.070

Str*Agr: t=1.248; p=.214

Subordinate-Main Construction

50

S

VP

were

S

NP

?

S’

NP

the men

V

hunted

VP

C

While

NP

the deer

[+N]

[+pl]

Cue-based Retrieval

Partial match

Cue-based Retrieval

Number Attraction and Sub/Main

51

• 72 Subjects, 40 Items
• Measured
• Self-paced reading time

Number Attraction and Sub/Main

52

18 msec

t=2.503; p=.012

3 msec

t=0.452; p=.651

Str: t=2.040; p<.051

Agr: t=1.454; p=.146

Str*Agr: t=2.082; p=.037