Lecture 3: Sentence Comprehesion Flashcards
(19 cards)
Two types of ambiguous sentences
Globally ambiguous: Ambiguous throughout the entire thing, by end still two interpretations. eg The criminal shot the cop with the pistol.
Temporarily ambiguous: ambiguous till the last word, disambiguated by the end of the sentence (some of these sentence can be hard to process) eg The criminal shot the dog with the collar
Why study ambiguous sentences?
- if syntax of a sentence is unambiguous then it is very easy for the processor to determin the structure
- therefore it is very difficult to investigate processing mechanism (eg whether information other than the grammar affects structural processing)
- syntactic (structural or attachment) ambiguity: what information is used when the grammar does not dictate a particular analysis
Garden Path Sentences
-sentences that lead you up a dead end when you adopt one analysis but it turns out to be wrong (have to reanalysis)
Garden Path Model
Frazier 1987 and Rayner et al 1983)
- Garden-Path theory
- attempts to explain parsing (syntactic analysis of a sentence) of both ambiguous and un ambiguous sentences (similar to sentence comprehension but focus on syntax
- the parser follows two very simple principles: minimal attachment and late closure.
- first minimal attachment is applied and if it doesn’t solve the ambiguity late closure is applied
- assumes that words are incorporated into previous structures immediately from left to right (incremental processing)
Minimal Attachment
If grammatically permissible attach an ambiguous phrase into the tree structure using the fewest number of additional nodes
-simplest: fewest nodes in the tree structure
eg The criminal shot the cop with the pistol (globally ambiguous), whether PP with the pistol is to be attached to VP headed by shot or NP the cop. (remember to explain in terms of syntactic structures not semantics)
-VP has the fewest nodes so simples (aka is preferred), minimal attachment in this case is plausible so no reanalysis necessary, -if minimal attachment is plausible then there is no difficulty but if not plausible have to reanalyze
*look notes
Minimal attachment
- The criminal shot the cop with the pistol
- The criminal shot the dog with the collar
- The criminal shot the dog with the pistol
- Globally ambiguous, VP preferred, plausible (no difficulty)
- Temporarily ambiguous, VP attachment preferred, not plausible (a collar cannot be used to shoot) difficult, need to reanalysis (NP attachment)
- temporarily ambiguous, VP attachment preferred, plausible (no difficulty), no reanalysis
Rayner, Carlson and Frazier 1983
- eye-movement experiment: they measured where people look and for how long when reading sentences
- two conditions: VP (the spy saw the cop with binoculars but the cop didn’t see him) and NP (the spy saw the cop with a revolver but the cop didn’t see him (should reanalysis when get revolver))
- 20 participants and 12 items
- Garden path model: difficulty should arrive with a revolver in NP condition
- reading times for the end of sentence VP
Minimal attachment examples
LOOK IN NOTES!!! several more
Late Closure
-if minimal attachment does not make a prediction regarding ambiguity resolution (equally many nodes), the garden path model claims that a second principle applies
-if two structures involve the same number of nodes, attach the incoming phrase to the phrase currently being processed (=as low in the tree structure as possible)
-in many cases this amounts to a recency preference: you attach an ambiguous phrase to the most recent phrase
eg The man realized we left yesterday (subordinate clause or low attachment)
*more examples in notes
Traxler, Pickering and Clifton 1988
-Eye movement experiment
1. The steak with the sauce that was tough didn’t win a prize (attach tough to steak, not what late closure prefers, violates it)
2. The steak with the sauce that was runny didn’t win a prize. (attach runny with sauce, what late closure prefers, consistent with it)
Result: reading times for that was tough/runny 1>2
Assumptions of Garden-Path Model
- when we process sentences we first use syntactic information
- in many cases the sentence structure in unambiguous so this results in the correct structural analysis
- in many other cases the structure is ambiguous, even in such cases we use syntactic information first
Properties of Garden-Path model
- Modular model: initial syntactic processing is not influenced by non-syntactic information eg plausibility), syntactic processing is autonomous, other info is used after initial syntactic analysis
- serial model: only a single analysis is adopted at a time
- reanalysis model: difficulty occurs when the initial analysis is inconsistent with the information that is used later
Constraint-based Models
eg MacDonal et al 1994, trueswell et al 1993 and 1994
-they claim that one of the most fundamental assumptions of the garden-path model is incorrect
-syntactic processing is NOT modular
-non-structural information is used immediately and at the same time as structural information
-interactive model: all potentially useful sources of information interact from the very earliest stages of sentence processing
AKA constraint-satisfaction models (group of models not a single definitive one)
Properties of constraint-based models
- interactive models: all sources of information (eg context, meaning, frequency) are used immediately and interact with each other
- parallel models: all analyses of an ambiguous structure are activated in parallel, the stringer the support for an analysis the higher its activation
- competition models: difficulty occurs when two or more analyses are about equally activated by the different sources of information
Garden path and constraint based models on plausibility/semantic information
- Garden-path: the use of plausibility/semantic info is delayed relative to structural strategies
- constraint-based: comprehenders use plausibility information immediately
Reduced relative/main clause ambiguity
a)1. the defendant examined the suspect. (main clause analysis, verb is part of main clause, defendant doing the examining)
2. The defendant examined by the lawyer was unreliable (reduced relative analysis, verb is part of a reduced relative clause, defendant is being examined)
b)1.The defendant examined by the lawyer turned out to be unreliable (temporarily ambiguous; reduced relative)
2.The defendant that was examined by the lawyer turned out to be unreliable (unambiguous)
Results: reading times for ‘by the lawyer are longer in 1 than 2 (ferreira and clifton 1986)
->consistent with garden path model: minimal attachment
->but could ambiguity resolution be affected by plausibility as claimed by constraint-based models in other cases?
c)The evidence examined by the lawyer was unreliable.
- Reduced relative analysis:
-verb ‘examined’ is part of a (reduced) relative clause, as in ‘The evidence that was examined by the lawyer
-‘evidence is being examined -> this is a plausible event.
??? The evidence examined the suspect.
- Main clause analysis is implausible:
Evidence can’t examine anything!
Constraint-based Models: Comprehenders should immediately use plausibility information and adopt the reduced relative analysis.
Garden-path Model: Comprehenders should initially ignore plausibility and therefore initially adopt the main clause analysis. They should use plausibility later.
Trueswell, Tanenhaus, & Garnsey (1994)
Eye-movement experiment:
2x2 (Animacy (animate/inanimate) x Ambiguity (ambiguous/unambiguous))
1. The defendant examined by the lawyer turned out to be unreliable. (animate - ambiguous)
2. The defendant that was examined by the lawyer turned out to be unreliable. (animate - unambiguous)
3. The evidence examined by the lawyer turned out to be unreliable. (inanimate - ambiguous)
4. The evidence that was examined by the lawyer turned out to be unreliable. (inanimate - unambiguous)
(1): ‘the defendant examined…’: main clause analysis - plausible.
(3): ‘the evidence examined…’: main clause analysis - implausible.
Predictions:
Garden-path model:
Reading times for ‘by the lawyer’: 1 > 2, 3 > 4
Constraint-based models:
Reading times for ‘by the lawyer’: 1 > 2, 3 = 4
Results:
Reading times for ‘by the lawyer’: 1 > 2; 3 = 4
Compatible with Constraint-based Models
Plausibility information affects the structural preference. (at early stage)
Clifton et al. (2003) redid previous experiment
It is difficult to interpret the absence of a difference between the 2 implausible conditions (3) and (4). This is a null effect.
Trueswell et al. (1994) may not have had enough experimental materials to find a difference. Statistically, it is not possible to prove the absence of a difference.
Clifton et al. (2003) used more materials & more subjects (plus other changes).
Results:
Reading times for ‘by the lawyer’: 1 > 2; 3 > 4
Compatible with the Garden-Path Model
Different from Trueswell et al. (1994)
Summary for plausibility effects (based on past two experiments)
Trueswell et al. and Clifton et al. have produced conflicting results.
Overall, the literature has shown both patterns of results.
However, it seems that we have slightly more evidence on plausibility effects in reanalysis, not in the initial analysis.
