Please enable JavaScript.
Coggle requires JavaScript to display documents.
PARSING AND SYNTACTIC AMBIGUITY :PENCIL2: D+L2 LECTURE 3 - Coggle Diagram
PARSING AND SYNTACTIC AMBIGUITY
:PENCIL2:
D+L2 LECTURE 3
The Garden Path Theory
Frazier & Fodor (1978); Frazier & Rayner (1982)
Argues that parsing is...
Serial
The parser is only entertaining one possible hypothesis about what the tree structure is at any one time.
Minimal Attachment: go for the tree structure with the fewest nodes.
Late Closure: Attach low down in the tree structure.
Autonomous
Semantic context is irrelevant to the parser
Meaning does not influence the parsers decision on syntactic roles
Incremental
Word by word
Try and design a tree structure as the information comes in.
Innate
We all have the same underlying parser
Eye-tracking studies
Frazier & Rayner (1982)
Garden path (late closure fails):
(1) “Since Jay always jogs a mile and a half
seems
a very short distance to him.”
The sentence above contains a point where according to the garden path theory, late closure would’ve been applied and it fails.
The cognitive system (parsing system) needs to re-evaluate the information which might lead the eyes to go back.
No garden path (late closure works):
(2) “Since Jay always jogs a mile and a half
this
seems a very short distance to him.”
Sentence 1 = slower to read - longer fixations in disambiguating region (underlined)
Sentence 1 = more regressions from disambiguating region to ambiguous region.
Evidence for serial parser by late closure
One tree structure being hypothesised at any one time.
BUT… the eyes are serial processors - can only focus on one part of the text at a time.
More support: Minimal attachment
Rayner & Frazier (1987)
“The criminal confessed his sins | harmed many people.”
Reduced relative.
Assume ‘
sins
’ could be the end of the sentence.
Minimal attachment
- ‘sins’ needs to be attached somewhere.
Simplest attachment (one with the fewest nodes) is to assume that it is the object of the sentence.
“The criminal confessed that his sins harmed many people.”
Doesn’t lead you down the garden path in the same way.
Unreduced form.
Similar to 1982 paper, but used sentences designed to elicit incorrect parse due to minimal attachment.
Once again garden path sentences led to:
Longer fixation at a unambiguous point (e.g. ‘harmed’)
More regressive eye-movements.
The Garden Path Theory: Review
Agreed that the parser should be incremental.
We need to try and assign syntactic roles as soon as we encounter the words in a sentence, whether we hear them or we see them.
Evidence that fits with the idea that we use heuristics (minimal attachment and late closure) in the context of a serial parser.
Only ever entertains one tree structure at a time.
INCREMENTAL ✓
SERIAL ✓
AUTONOMOUS ?
INNATE ?
Garden path theory turned out to be wrong, but provided the framework for other research.
Referential Context
(1) “The burglar blew open the safe with the dynamite.”
(2) “The burglar blew open the safe with the new lock.”
We now know that 1 should be easier to parse than 2 (
minimal attachment
)
But is this because we are presenting the sentence in an impoverished context?
E.g. We haven’t been given any information about a safe with any other type of lock?
“With the new lock” - infelicitous
What if the referential context requires more specificity?
Maybe we have seen two safes, one with a new lock and one with an old lock.
In this context, this level of information is therefore perhaps required?
Would be interesting to see if in this context, this sentence was easier to process.
Altmann & Steedman (1988)
2 contexts:
A burglar broke into a bank carrying some dynamite. He planned to open a safe. Once inside, he saw that there was…
(a) ...a safe with a new lock and a strongbox with an old lock
(1 safe context)
(b) ...a safe with a new lock and a safe with an old lock
(2-safe context)
Target sentences:
Minimal attachment works
(MA):
“The burglar blew open the safe with the dynamite.”
In the absence of any context, this one should be easier to process.
Minimal attachment doesn’t work
(NMA):
“The burglar blew open the safe with the new lock.”
Garden Path Theory doesn’t care about the context - just cares about the parsing of the words in the sentences - first sentence should always be easy to parse, second sentence should always be relatively hard.
Weak evidence against the garden path theory.
Suggests that context matters and that it can reverse minimal attachment effects.
Referential Theory
Incremental parallel processing
Alternative parses are generated in parallel each time a new word is encountered.
Weak Interaction
Context can affect the selection between these alternatives.
No longer autonomous.
No such thing as a truly neutral context.
The less that is given, the more assumptions have to be made = harder to process, slows down processing - hypothesis generation (Crain & Steedman, 1985)
E.g. ‘The horse raced past the barn fell.”
Context can override garden path effects
How does language interact with our exploration of the world?
Often language is viewed as modular (e.g. Pinker)
Unaffected by other aspects of cognitive processing.
We have seen that key components of language processing are not autonomous (e.g. syntactic processing)
Can we determine whether language comprehension is at least more broadly autonomous?
For example, is comprehension influenced by our visual environment?
Interaction of language and vision
Tanenhaus et al. (1995)
One referent condition:
“Put the apple on the towel into the box.”
Should lead to a garden path response.
Now we need to specify which apple (less ambiguity)
Visual context of the sentence influences the likelihood of being garden-pathed.
Again evidence against autonomy/modularity of parser (or even modularity of language system)
Language system allows interaction with the visual world.
Supports referential models.
Apple = subject
On the towel = object
Put = verb
Influence of Animacy
Trueswell, Tanenhaus & Garnsey (1994)
Animacy = the extent to which they do stuff.
“The defendant examined…” (
animate
)
“The evidence examined…” (
inanimate
)
Animacy is a good example of a semantic variable that might influence parsing.
“Examined” requires an agent - someone to do the examining.
A defendant can examine, but evidence can’t.
Does “evidence” stop us getting garden-pathed?
Eye-tracking study
To what extent do readers get garden-pathed by…
“The defendant examined by the lawyer turned out to be unreliable.” (
animate
)
“The evidence examined by the lawyer turned out to be unreliable.” (
inanimate
)
Control conditions are unreduced (that was = inserted between defendant/evidence and examined)
Garden path theory suggests that we should be equally garden-pathed in both of these sentences because we don’t care about animacy.
Whereas a theory that included semantic information as part of the parsing process, as maybe a bias or a constraint in the parsing process, would say that the “defendant” condition would show a proper garden path effect, but maybe the “evidence” condition wouldn’t so much, as you might expect that evidence is going to be examined by someone else.
Animacy does matter and does make a difference to the size of the garden path effect.
Evidence against a fully autonomous par (the garden-path theory).
INCREMENTAL ✓
SERIAL ❌
AUTONOMOUS ❌
INNATE ?
Constraint-based frameworks
MacDonald et al. (1994)
Parser considers alternative options in parallel
Evidence in favour of different parses accumulates over time.
Multiple sources of constraints provide evidence for various competing parses.
Taking into account semantics, broader information, context.
No sharp distinction between initial and second stage of parse.
--> Regression?
--> Eyes are forcing a linear, serial process, because the eyes can’t be in more than one place at any one time.
--> May just be focusing on the most favourable option?
No priority to syntactic information.
Ambiguity resolution viewed as a constraint satisfaction process.
Lots of constraints (
semantic, contextual, syntactic
)
How can we put this information together to get the most likely option considering all these constraints?
Cross-linguistic evidence
Cuetos & Mitchell (1988)
Innateness
Looked at Spanish and English
Who had the accident?
English
: “The journalist interviewed the daughter of the colonel who had the accident.”
Did the daughter have the accident or did the colonel have the accident?
Late closure example
Attach it to the most recent construction.
In this case, you favour the interpretation in which the
colonel
had the accident.
Spanish
: “El periodista entrevisto a la hija del coronel que tuvo el accidente.”
Word for word translation.
Results
For
English
participants, about
40%
said that it was the
daughter
who had the accident.
Against the idea of late closure.
60% said it was the colonel who had the accident. (late closure)
This kind of individual variability is not explainable by the garden path theory.
For Spanish listeners, the bias is reversed.
60% said daughter
40% said colonel
Conclusions...
Language matters.
Spanish seems to prefer early closure
Similar results for eye-tracking.
INCREMENTAL ✓
SERIAL ❌
AUTONOMOUS ❌
INNATE ❌
More constraints
Gibson (1988)
Some syntactic preferences/problems can be reinterpreted in terms of limited memory resources.
Heuristics that we use could be a result of limited memory resources.
(1) "The senator who spotted the reporter shouted."
(2) "The senator who the reporter spotted shouted."
(3) 'The senator who the reporter next to the president spotted shouted."
(4) "The senator who the reporter who the president detested spotted shouted."
(5) "The president detested the reporter who spotted the senator who shouted."
The last sentence is easier to parse than the penultimate one, because the first of the two leaves all of the information needed until the very end - too much ambiguity at the start of the sentence to be easily comprehensible.
Can be thought of as a resource constraint (locality)
Adds further weight to a constraint-based approach.
Conclusions...
Garden path model provided rigorous theoretical structure to test
Relatively simple model (Occam’s razor)
We now have evidence for a more complex parsing process
Makes use of theoretical constructs of original model
More interactive, parallel, making use of all kinds of knowledge.
Parsing as constraint satisfaction, harnessing parallel massively interconnected power of brain (but reliant on its limited resources)