Transcript Ferreira et al 01 Presentation Kokhan

Kate Kokhan
Department of Linguistics
University of Illinois at Urbana-Champaign
 What
is a garden-path sentence?
 Motivation
for the research by Ferreira et al.
 Description
of the experiments and the findings
 Implications
(1998)
for reanalysis: Fodor and Inoue
A garden path sentence is a grammatically correct sentence
that starts in such a way that a reader's most likely interpretation
is an incorrect one, luring him initially into an improper parse
that then turns out to be a dead end.
The "garden path" is a reference to the saying "to be led
down the garden path", meaning "to be misled.“
 Previous
work on how garden-path sentences
are processed had barely addressed how
these sentences are actually understood.
 No
data on the question of whether people
understand these sentences the way we
presume they should.
 General
goal was “to discover how people
understand language”
 Many studies showed that people had problem
understanding some sentences (if they are
presented visually without punctuation):
While Anna dressed the baby spit up on the bed.
The baby is
the object of
dressed
When they
encounter spit
up => error
Goal of all models of reanalysis is to
describe and motivate the mechanisms
used by the parser to detect errors,
deduce useful information about the
nature of the necessary repair from those
errors, and ultimately to create successful
parse.
Is reanalysis always successful? – No
=>
Another goal of models of reanalysis is to
explain why it is that successful revision is
possible for some sentences but impossible
for others.
 Ferreira
& Henderson (1991, 1998) =>
While Anna dressed the baby that was small and
cute spit up on the bed.
the error signal
the head of the
misanalysed phrase
Reanalysis is more difficult if the head of the
misanalysed phrase is distant from the error signal!
However, the length of the sentence is not important.
3 experiments but the same paradigm:
1.
Participant were asked to read garden-path
sentences or corresponding control sentences
2.
They answered YES/NO questions
3.
Indicated their confidence in that answer
Primary dependent measures are questionanswering accuracy and confidence.
(1a) While Bill hunted the deer (that was brown and
graceful) ran into the woods.
G-P
(1b) While Bill hunted the deer (that was brown and
graceful) paced in the zoo.
G-P
(1c) While Bill hunted the pheasant the deer (that was
brown and graceful) ran into the woods.
Non-G-P
Difference between (a) and (b) => the plausibility of
misinterpretation
Note: each participant saw only one out of the 6 possible
versions but experience each of the 6 conditions. Read
the sentences at their own pace.
(1a) While Bill hunted the deer (that was brown and
graceful) ran into the woods.
(1b) While Bill hunted the deer (that was brown and
graceful) paced in the zoo.
(1c) While Bill hunted the pheasant the deer (that
was brown and graceful) ran into the woods.
Question: Did Bill hunt the deer?
If people answer YES to (a) and (b) => they did not
end up with the appropriate interpretation for the
sentences.
If they say YES if the material in the parenthesis is
included => tendency to say YES is attributable to
the initial syntactic misanalysis
 With
1.
2.
2 innovations:
Instead of using the control condition, they
created non-garden-path conditions by
reversing the order of the subordinate and
main clauses => The deer ran into the woods
while the man hunted
Asked people two questions: Did the man hunt
the deer? + Did the deer run into the woods?
Note: both phrases were long instead of short vs.
long in Experiment 1
 (2a)
While Bill hunted the brown and
graceful deer/ the deer that was brown and
graceful ran into the woods.
 (2b)
The brown and graceful deer/ the deer
that was brown and graceful ran into the
woods while Bill hunted.
 Questions:
Did Bill hunt the deer? or Did the
deer run into the woods?

To avoid the inference: when someone is
intransitively hunting, he or she may be hunting a
deer => Reflexive Absolute Transitive (RAT) verbs,
e.g. dress and bathe
Two sub-experiments:
1. The non-GP condition was created by reversing the
order of the subordinate clause. 2x2 design= a
sentence was either G-Pathing or not and the verb
was either normal or RAT
2. They varied whether or not a comma separated the
subordinate and the main clause: While Anna dressed,
the baby spit up on the bed. The same 2x2 design
 The
general reasoning argument:
High numbers of incorrect answers to the
comprehension questions stem from pragmatic
inference: the man is hunting, and the deer,
which is a common quarry for hunters, is running
away => the man is possibly hunting the deer.
1. The strong head position effect in Experiment 2=> no obvious
reason why “the deer that was brown and furry” should be more
likely to be inferred to the quarry of the hunt than “the brown and
furry deer”, yet the subjects were more likely to derive an
incorrect interpretation from the former construction than from
the latter.
2. Elimination of syntactic ambiguity without changing the
information load: reversing the clause order and separating the
clauses with a comma => significantly fewer incorrect YES
responses than the GP sentences
3. Use of RAT verbs =>participants had trouble reaching the
correct interpretation of RAT verbs in the GP conditions
=> Pragmatic inferencing cannot account for the
misinterpretation effect in the GP conditions.
 Some
garden-path sentences were completely
reanalyzed but for others, no reanalysis
occurred. The significant proportion of
incorrect YES responses in the GP conditions
could have been attributed to guessing on the
trials on which the reanalysis did not take place
=
> evidence against this possibility. Some
degree of reanalysis must have taken place for
the NP initially parsed as the object of the
subordinate clause to be recruited as the
subject of the matrix clause.
Two alternative explanations:
 Radical (that reanalysis terminates with and
the parser settles on a structure that is not
licensed by the principles of human syntax)
 The NP that is necessary for the main clause
is stolen from the subordinate clause, yet not
erased from the original position
=>this explanation is consistent with a model of
parsing and reanalysis by Fodor and Inoue (1998)
“Attach Anyway” => directs the parser simply to
attach an incoming element even if it does not fit
into the current phrase marker, temporarily
ignoring the ungrammaticality that results
 “Adjust” =>resolves the grammatical conflict
minimally, so that the adjustment may result in yet
another conflict elsewhere in the tree. Adjust then
moves on to correct that conflict and so on.
Note: Adjust is constrained by Grammatical
Dependency Principle (GDP)=>Adjust operations apply
between nodes that are in some sort of a
grammatical relationship: a head and its argument or
a verb and its subject


Reanalysis of the sentences used in Ferreira et al.’s
experiments does not involve elements in such a
relationship => in such circumstances, Fodor and Inoue
suggest that the parser engages in “theft”, i.e. the
lexical string the deer is stolen from the subordinate
clause and assigned a position in a matrix clause, in
violation of GDP.
Theft could result in the dual thematic role assignment
of Theme and Agent to deer because when GDP is
violated the backward repair operations (Adjust) that are
normally set in motion by Attach Anyway do not take
place.
 as a result, the parser would not end up being able to
return to the subordinate verb, re-access the lexicon,
and locate its intransitive argument structure.
 The subordinate verb would, therefore, remain
transitive, yielding a structure that could support the
dual Theme/Agent (mis)interpretation.

 For
the plausible sentences (e.g. While the man
hunted the deer ran into the woods, the
overall interpretation is checked according to
the Minimal Revisions Principle (MRP)
 MRP:
any revisions deemed necessary and
consistent with the error signal should maintain
as much of the initially assigned structure and
interpretation as possible.
 The MRP predicts that the proposition stating
that the man hunted the deer would not
necessarily be erased, because that
interpretations should be maintained to the
extent possible.
RAT verbs=> reanalysis differs
The correct analysis of the sentence While Anna
dressed the baby spit up on the bed had the parser
assigning the role of Theme to an empty category
after dressed, resulting in a specific theme for the
verb dressing (Anna herself).
 contrasts with While Bill hunted the deer ran
into the woods: hunt’s Theme role cannot be
assigned to any syntactic constituent at all.
Therefore, the interpretation that must be
created is generic: Bill hunted something or
other.
 The participants’ level of confidence suggests
that they are relatively insensitive to any
violations that might result from the early
termination of Adjust operations

 “It
appears that people work on sentences
until they reach a point where it subjectively
makes sense to them and then processing may
cease… these garden-path sentences somehow
produce an illusion of comprehension in our
participant...They will be misunderstood
despite the best attempts of the
comprehender to come up with a correct
analysis”