Lecture 4: Sentence comprehension 2 Flashcards
(21 cards)
Context effects of Garden-Path model and constraint-based models
Garden-Path Model: Context is initially ignored, but has an effect on later stages of processing.
Constraint-based Models: Context has an immediate effect (but is not the only factor).
There are different types of contexts, but here we deal with ‘discourse’ contexts (particularly ‘referential’ contexts) that are provided OUTSIDE the target sentence.
How can context affect syntactic processing?
(pictures of all police men)
Suppose I want to tell you which picture I like. What would be wrong if I said?
I like the picture of the policeman.
The fact that I don’t mention ‘with the pistol’ (or anything to distinguish him from the other policemen) would presuppose that there is only one policeman in the context.
That is, the presupposition wouldn’t be satisfied. (not enough information)
(pictures of one policeman and other professions)
Suppose I want to tell you which picture I like. What would be wrong if I said?
I like the picture of the policeman with the pistol.
The fact that I mention ‘with the pistol’ presupposes that there is more than one policeman in the context. (even though actually just one
That is, the presupposition is not satisfied. (redundant)
How can presuppositions affect syntactic processing?
Presuppositions about modifiers and the number of objects of the same kind in the discourse:
‘the policeman’ (without a modifier) – there is only one policeman within the relevant discourse.
‘the policeman with the pistol’ ; ‘the fat policeman’; ‘the policeman who is wearing a green uniform’ etc (with a modifier) – there are more than one policeman.
The criminal shot the cop with the pistol. NP attachment (cop with the pistol): Presupposes that there are multiple cops. VP attachment (shot with the pistol): Presupposes that there is only one cop.
Referential Theory (eg Altmann and Steedman)
The Referential Theory explains how discourse context may affect syntactic ambiguity resolution.
The ideas of this theory have been later incorporated into the Constraint-based Models.
Prediction: The interpretation that contains fewest unsatisfied presuppositions is the preferred one.
Referential Theory (eg Altmann and Steedman) Example
The criminal shot the cop with the pistol.
Suppose we are reading this sentence without any extra-sentential context.
NP attachment: ‘with the pistol’ attached to NP ‘the cop’
In NP attachment, ‘with the pistol’ indicates which cop we are talking about. It singles out one cop (the one with the pistol) from other cops.
NP attachment presupposes that there were other cops in the context. Otherwise, why would we specify which cop we are talking about?
However, in our case (when reading the sentence in isolation), there is no context that indicates that there is more than one cop.
Therefore, NP attachment is odd without a context that mentions other cops.
NP attachment contains a presupposition that is not satisfied.
→ People should prefer to attach with the pistol to VP (shot with the pistol). (when no other discourse context)
VP attachment: ‘with the pistol’ attached to VP ‘shot…’.
VP attachment assumes that there is only one cop that is being talked about, because we do not specify which cop is talked about.
Therefore, it does not contain an unsatisfied presupposition, and this interpretation should be preferred.
Referential and Garden-Path Theory
They would make same predictions for sentences in isolation:
The Referential Theory assumes the same parsing principles as the Garden-Path Model (Minimal Attachment & Late Closure) when there is no context outside the sentence.
The two theories could lead to different predictions for sentences in a context.
Two-referent context
-referential and garden-path model
There was a cop who had a pistol, while another cop was
holding a stick.
The criminal shot the cop with the pistol.
Referential Theory:
Two cops in the context. Therefore, one cop needs to be selected, and the presupposition of the NP attachment interpretation is satisfied. NP attachment should be preferred. (NP need to satisfy presuppositions)
Garden-Path Model:
Contextual information is initially ignored. Minimal attachment always applies. VP attachment is preferred. Context could has an effect on later processing.
One-referent context
-referential and garden-path model
There was a cop who had a pistol, while an old lady was
holding a stick.
The criminal shot the cop with the pistol.
Referential Theory:
One cop in the context. Therefore, there is no need to select one cop in the ambiguous sentence, and the presupposition of the NP attachment interpretation is not satisfied. That is, VP attachment should be preferred.
Garden-Path Model:
Same as before: Contextual information is initially ignored. Minimal attachment always applies. VP attachment preferred. Context has an effect on later processing.
Summary of referential theory vs garden path model
The criminal shot the cop with the pistol.
No context: The Referential Theory and the Garden-Path Model make the same predictions: VP attachment preferred.
Two-referent (or multi-referent) context: The Referential Theory predicts a NP attachment preference, whereas the Garden-Path Model predicts an initial VP attachment preference.
One-referent context: Both theories make the same predictions: VP attachment preferred.
Altmann & Steedman (1988)
Self-paced reading experiment: 2x2 (Referential Context x Attachment)
->press button when done differnt parts of reading, less indirect method of tracking reading
- The burglar saw that there was a safe with a new lock and a safe with an old lock. He blew open the safe with the new lock and made off with the loot. (2 referents; NP attachment) -> match
- The burglar saw that there was a safe with a new lock and a strongbox with an old lock. He blew open the safe with the new lock and made off with the loot. (1 referent; NP attachment) -> mismatch (context sentence diff)
- The burglar saw that there was a safe with a new lock and a safe with an old lock. He blew open the safe with the dynamite and made off with the loot. (2 referents; VP attachment) -> mismatch (diff target sentence)
- The burglar saw that there was a safe with a new lock and a strongbox with an old lock. He blew open the safe with the dynamite and made off with the loot. (1 referent; VP attachment) -> match
Which of the preceding conditions are predicted to be hard according to:
Referential Theory: ‘mismatch’ ones (2&3)
Garden-Path Model: NP attachment ones (1&2)
Results: Reading times for ‘with the new lock/dynamite’:
1 4 (referential context effect)
Compatible with Referential Theory
Clifton and Ferreira (1989)
Argue that Altmann & Steedman’s results could be explained by the Garden-Path Model as well.
Problem:
A&S used a self-paced reading task, which would not distinguish first analysis and reanalysis.
cf. eye-tracking
The Garden-Path Model argues the parser has access to the referential information at a later stage of processing.
Summary of context effects
It is hard to use context to distinguish between the Referential Theory and the Garden-Path Model.
Context effects do occur.
However, they are often not as strong or early as claimed by the Referential Theory: Context makes the non-minimally attached analysis easier, but not always easier than the minimally attached analysis.
Constraint-based Models: Context is just one of the many sources of information that play a role, so it does not always have a strong effect.
Frequency information (different models views)
Garden-Path Model: Frequency information is initially ignored during syntactic analysis.
Constraint-based Models: Frequency information is used immediately.
Relative frequencies of alternative usages could affect ambiguity resolution.
Types of frequencies
- Frequencies of ‘argument structures’: Some verbs can take more than one structure:
-(Suppose) know + NP object > know + Sentential Complement
-‘know the cat’ > ‘know the cat is pretty’
-“The boy knows the cat…”: people would interpret ‘the cat’ as the direct object of ‘know’, instead of the subject of the sentential complement.
(use information and use pattern to determine cat NP object?) - Frequencies of ‘tense forms’: Some verbs can take the same form for different tenses:
- (Suppose) examined: past > past participle
- ‘The defendant examined…’: ‘examined’ as a past verb, not a past-participle modifier in a reduced relative - Frequencies of ‘syntactic categories’: Some words can be of more than one syntactic category:
- (Suppose) mushroom: noun > verb
- ‘Lung Cancer in Women Mushrooms’: ‘mushrooms’ as a noun, not a verb
Frequency and attachment preferences
How would you complete these sentences?
After the child had visited …
After the child had sneezed …
After the child had visited …
… the medical practice the doctor prescribed injections.
After the child had sneezed …
… the doctor prescribed injections.
‘visit’ is usually followed by a direct object.
but it COULD omit its object.
‘sneeze’ usually does not have a direct object.
but it COULD take an object: “The child sneezed the napkin off the table.”
Both verbs (visit, sneeze) can be used transitively (with an object) or intransitively (without an object). But importantly, for ‘visit’, transitive is more frequent, and for ‘sneeze’, intransitive is. Constraint-based Models: Frequency of the structures may affect how we process temporary ambiguities.
What do you predict if frequency information (and nothing else) is used to analyse ‘the doctor’?:
After the child had visited the doctor prescribed injections.
After the child had sneezed the doctor prescribed injections.
If frequency information (and nothing else) is immediately used:
After the child had visited the doctor prescribed injections.
EASY HARD
-Direct object analysis should be adopted at ‘doctor’, but turn out to be wrong at ‘prescribed’.
After the child had sneezed the doctor prescribed injections.
EASY EASY
-Subject analysis should be adopted at ‘doctor’, and turn out to be correct at ‘prescribed’.
Mitchell (1987)
Self-paced reading experiment:
- After the child had visited the doctor prescribed injections.
- After the child had sneezed the doctor prescribed injections.
Results:
Reading times for ‘the doctors’: 1 2
In (1), ‘prescribed’ indicated a lack of subject, therefore reanalysis took place for ‘the doctor’.
- Frequency information does not provide a strong enough constraint to prevent the processor from considering the direct object analysis after ‘sneeze’.
- Inconsistent with the Constraint-based Models
- > unless one assumes that frequency provides only a weak constraint (or weaker than other constraints that prefer the direct object analysis).
- Consistent with the Garden-Path Model
- > Late Closure would prefer to close the current clause as late as possible.
Tests on architectures
-garden-path and constraint-based
Garden-Path Model:
Serialism - The processor adopts only a single analysis at any time.
Constraint-based Models:
- Parallelism - The processor activates multiple analyses in case of syntactic ambiguity.
- The stronger the various sources of information (semantics, context, frequency) support an option, the stronger the activation is.
Hard to test serialism/parallelism (i.e., how many options are activated at a given time in processing) so very little research on this.
Garden-Path Model:
Reanalysis causes processing difficulty. It occurs when an initially adopted structure turns out to be incorrect (e.g., implausible).
Constraint-based Models:
Competition causes processing difficulty. It occurs when two syntactic analyses are about equally activated.
Also, the Models assume reanalysis causes processing difficulty as with the Garden-Path Model above.
Competition vs reanalysis
experiment
Van Gompel, Pickering, & Traxler (2001):
- The hunter killed only the poacher with the rifle not long after sunset. (globally ambiguous)
- The hunter killed only the leopard with the rifle not long after sunset. (VP attachment)
- The hunter killed only the leopard with the scars not long after sunset. (NP attachment)
-‘Balanced’ global ambiguity: People don’t have a strong preference for VP or NP attachment, due to the presence of ‘only’. [without only, VP attachment would be preferred]
Constraint-based Models:
Ambiguous Cond: Competition should occur, as both analyses are equally plausible.
VP & NP Conds: No competition or reanalysis should occur, as one analysis is much more plausible than the other.
Garden-Path Model:
All: The emphasis put by ‘only’ is initially ignored, and VP attachment is preferred at ‘with the’.
NP Cond: Reanalysis should occur at ‘scars’, as the initially adopted VP attachment is implausible.
Ambiguous & VP Conds: No reanalysis should occur, as the preferred VP attachment is plausible.
Reading Times: Constraint-based Models: Amb (most difficult) > VP = NP Garden-Path Model: Amb = VP Incompatible with both Constraint-based (Amb > VP = NP) and Garden-Path (Amb = VP How can we explain these results?
Race model (look next slide)
Competition vs Reanalysis Race model (experiment based on last one)
Race Model (Van Gompel et al., 2001)
- The two analyses are involved in a race (competition).
- The one that is supported by most information is constructed first.
- Because there is no strong preference in this case, on half the trials, readers initially adopt VP attachment, and on the other half NP attachment.
- When the initial analysis is implausible, readers reanalyse and difficulty occurs. Therefore, difficulty occurs on half the trials in the VP and NP attachment conditions.
- When the initial analysis is plausible, no reanalysis occurs. In the ambiguous condition, both VP and NP attachment are plausible, so no difficult occurs.
“The hunter killed only the poacher / leopard with the…”
AMB: killed only the poacher with the…
- syntax->no bias
- semantics ->no bias
- VP analysis 50%, NP analysis 50%
VP: killed only the leopard with the..
- syntax->no bias
- semantics ->no bias
- VP analysis 50%, NP analysis 50%
NP: killed only the leopard with the…
- syntax->no bias
- semantics ->no bias
- VP analysis 50%, NP analysis 50%
“The hunter killed only the poacher / leopard with the rifle / scars…”
AMB: killed only the poacher with the rifle…
- syntax->no bias
- semantics ->no bias
- VP analysis 50%, NP analysis 50%
- no reanalysis necessary so both easy
VP: killed only the leopard with the rifle…
- syntax->no bias
- semantics -> VP
- VP analysis 50%, NP analysis 50%
- VP no reanalysis necessary so easy but if choice NP need reanalysis to VP so it is hard
NP: killed only the leopard with the scars…
- syntax->no bias
- semantics -> NP
- VP analysis 50%, NP analysis 50%
- if choice VP have to reanalysis to NP so hard, if choice NP no reanalysis necessary so easy
Why: AMB
Do people always get the correct analysis?
While Anna dressed the baby that was small and cute spit up on the bed.
Did Anna dress the baby?
Christianson, Hollingworth, Halliwell, Ferreira (2001):
While Anna dressed the baby that was small and cute spit up on the bed. Did the baby spit up on the bed? - Subjects answered yes. (correct answer) Did the Anna dress the baby? - Subjects also often answered yes (~ 60%)! Control condition: While Anna dressed, the baby that was small and cute spit up on the bed. (comma disambiguates it) Did the Anna dress the baby? - YES answer: 12%
Good enough representations theory
and experiment
Ferreira (2003):
People often do not process the syntactic structure of sentences fully. Instead, they construct a “good enough”, incomplete representation. (not complete)
In the case of syntactic ambiguity, people often fail to reanalyse completely:
- While Anna dressed the baby that was small and cute spit up on the bed.
- They adopt the appropriate analysis (the baby spit up on the bed).
- But they fail to abandon the inappropriate analysis (Anna dressed the baby)
Is the following sentence plausible?
The dog was bitten by the man.
Ferreira (2003): Subjects quite often (26%) fail to identify correctly who carried out the action (saying ‘the dog’ instead of ‘the man’) and who was affected by it (15%).
This happens much less often (1% - 9%) with active sentences (‘The man bit the dog’).
The findings from Christianson et al. (2001) and Ferreira (2003) are inconsistent with the Garden-Path Model and Constraint-based Models, because the two theories assume that people construct complete syntactic representations.
