You are on page 1of 16

PROCESSING NOTES  One a logogen has been triggered or activated,

its threshold for activation is temporarily


WEEK 14: P97-117
lowered
LEXICAL ACCESS: set of mental representations and  As a result, less evidence is needed in the
process that are involved in identifying which specific acoustic and visual input channels to reactivate
words we are hearing (during spoken word processing) the logogen.
or seeing (during visual word processing).  This mechanism can account for repetition
priming effects, it is easier to recognise a word
Principles of spoken word processing: (studies) the second time you see it than the first
1) William Marslen-Wilson (1973): shadowing because the activation threshold is lower the
task to estimate how much time it took people second time around
to identify words. Subjects listen to recorded Assumptions of Logogen Model:
speech and they try to repeat (shadow) as
quickly as possible the words that they hear. 2) Information flow is strictly bottom up; auditory
and visual processing units affect the
Result: fast shadowers were able to perform activation of logogens, but logogens do not
lexical access very fast and that higher order affect the activation levels of auditory and
aspects of the speech stream. visual processing units that feed into the
logogen.
2) Word monitoring (involves listening to 3) There are no direct connections between and
utterances and responding as quickly as among the logogens themselves. As a result,
possible when a specific target word appears in the activation level of one logogen does not
the input) and gating tasks (involves listening affect the activation of other logogens.
to short snippets of the beginnings/onsets of
How is it successful?
words)
1) Proves that exposure to high-frequency words
THEORIES: to explain how people take inputs from
lowers the threshold for activation in the
the auditory or the visual system and match those
logogens that represent those high-frequency
inputs to stored representations of word form.
words  less external evidence needed 
1) John Morton – logogen respond faster to high-frequency words
2) Explains why high-frequency words are shorter
Bottom-up driven system that takes spoken or visual than lower-frequency words  shorter words
input and use it to activate previously stored word form pack less phonological and/or orthographic
representations (Morton, 1969) information
Each word in a person’s vocabulary is represented by a 3) Explain high-frequency words are easier to
logogen  Words are therefore recognised when the recognise when they have been degraded by
activation levels of their corresponding logogen exceed noise. Noise in the signal decreases the quality
some threshold. of the bottom-up input, but high-frequency
words don’t need as much bottom-up input, so
Logogen could receive inputs from either they are recognised even in noisy environment
- Spoken words
- Written words
- Preceding context 4) Frequency ordered serial bin-search
 Word form representations were activated by
 The semantic input mechanism allows for bottom-up input from the auditory system 
context to influence the time it takes to people use auditory/visual cues to search their
recognise a word long-term memories for a matching stimulus
 Context words that are semantically relatd to  Lexical representations are organised into bins and
an individual logogen will raise the activation they are organised according to word frequency
of the logogen before the listener gets direct (high-frequency word = front bin, lower-frequency
perceptual evidence that the corresponding word = back bin)
word is actually present in the input
 Self-terminating – the searching process stops  Lateral inhibition – processing units within a layer
itself when it succeeds of units in the network try to reduce or inhibit each
 Words are organised in the bins according to other’s levels of activation – letters have excitatory
shared roots (dog, dogs, dogged, dogpile,..) – a or inhibitory feedback connections from other
word that has a low surface frequency will be letters
responded to quickly if its root frequency is high  Word superiority effect – a class of behaviours
indicating that we have an easier time recognising
5) COHORT letters and phonemes when they appear in the
 Views the process of lexical access as involving 3 context of a word than when they appear by
processes: themselves or alone

o Activation (contact)
DIFFERENCE BETWEEN COHORT AND TRACE
Multiple word form representations are activated in
response to the auditory stimulus TRACE COHORT
Views word form Views word form
o Selection
activation as resulting activation as reflecting a
Sorting through the activated word form from a process of massively parallel
representations to find the one that best matches the competition and mutual process without
auditory stimulus. It depends on bottom-up stimulus. inhibition competition until the
selection phase
o Integration More activated words Allows for unlimited
candidates are parallel activation of
Happens when the features of the selected word are associated with less word candidates, the
incorporated into the evolving representation of the activation being gained number of activated
entire utterance. Properties of the selected word by any one candidate, candidates does not
(grammatical class and meaning) are evaluated with and greater competition affect the speed with
respect to how well they fit with the preceding context. between candidates which the correct
candidate is identified
 As an autonomous process – it is affected by Relies on global Word onsets (beginning)
auditory stimulation. Word representations are similarity match to are critical because they
activated as soon as the initial sounds in the determine how active a determine which
acoustic stimulus have been perceived, and it is stored word form representations will
possible for people to identify individual words in becomes make it into the cohort
the speech stream before they hear the entire Activation of word nodes Words are recognised
word. will be a function of when the acoustic
 Word that are highly predictable in context may be similarity. Hence, words stimulus reaches the
recognised a bit faster than less predictable words that share offsets should recognition point –
prime each other’s people can recognize
meanings words and access their
meanings without
having to wait until the
6) TRACE very end of the word
 Highly interactive – have connections between
processing units that allow units within the same
level to affect one another
 Combining bottom-up as input and top-down for
feedback
 Assumes that activation is cascaded – units
receiving input begin to send output as soon as any
activation at all comes in from other units
 Letter representations start to become activated
as soon as any visual feature has been identified
WEEK 15: TRAXLER CHAPTER 10

Word Frequency Effect: recognition times are faster 1) EYE MOVEMENT CONTROL AND READING
for words seen more frequently than for words seen
less frequently.
 Highly predictable words = skipped
Word Supremacy Effect: people have better
recognition of letters presented within words as  Very short function words = skipped
compared to isolated letters and to letters presented  Content words = fixated
within nonword strings

Context Effect: influence of environmental factors on Perceptual span: to find out from what region of text
one's perception of a stimulus. (top-down design). people can extract information on a given fixation
Distortion Effect: Study:

1) moving window paradigm: part of the text is


displayed in the normal way, and part of the
text is replaced with something else (XXX)
2) Boundary change paradigm: an invisible
boundary is located somewhere in the text –
the purpose is to deny readers accurate letter
preview information

Dual-route and Dual-route cascaded (DRC) models:

 two most intensively investigated and highly


developed accounts of visual word processing.
 There are two separate ways that people can
use visual input to access entries in the mental
lexicon
o 1) access the lexicon by ‘sounding out’
the word
o 2) access the lexicon directly without
activating phonological codes

DRC:

1) you can use a sequence of letters to access a


word’s lexical entry by converting that
sequence of letters into a sequence of
phonemes
2) apply grapheme-to-phoneme correspondence
(GPC) rules on a letter-by-letter basis
3) readers appear to access the lexicon by
activating the sounds that go with each letter
in a more-or-less left-to-right fashion
4) the activated phoneme then make contact
with an entry in the lexicon that has the same
sequence of phonemes

Phonological/Assembled Phonology Route:


 when exception words (with irregular phonological units that are connected
spelling-pronunciation) have run through to words in that neighbourhood
the assembled phonology system, it
compiles a set of phonemes that do not
match any of the entries in the mental
lexicon.
 Regularisation error: when phonological  Recent neuroimaging data show that
route stays active and can cause readers to overlapping brain areas are activated when
people read novel words compared to real-
mispronounce exception words
words  implies a unified system for reading
 Regular words are accessed via assembled
novel words and familiar words
phonology
 We can also use it to generate
pronunciations for words that we have
never seen before/assigned any meaning
to
DYSLEXIA – when an individual has a
problem reading, even though they are otherwise
Orthographic/Direct Route: intellectually and behaviourally normal
 DRC – proposing that some words can be 1) SINGLE-DEFICIT MODELS
accessed via a separate system that People who have trouble reading non-words
bypasses the assembled phonology route than exception words (
and contacts the lexicon directly
 Exception words accessed via visual code,
and once the matching lexical entry has
been activated by the visual input, the
lexicon makes the correct pronunciation
available.
 Irregular words are accessed via the direct
route

Single-route model

Based on a neural network model containing three


groups of processing units

 Orthographic units
 Hidden units
 Phonological unit

Regularity effect: regular words are easier to recognise

Frequency by regularity interaction:

 Orthographic neighbourhood – words often


come in groups that resemble one another
(book, took, nook, look)
o Letters that activate large
neighbourhoods will lead to more
units sending activation to the
WEEK 16: TRAXLER (54-70) listeners can accurately identify the missing vowel
sound.
Speech perception:

Liberman and colleagues (1952): phonological content


of speech could be described in terms of formants EVIDENCE FOR COARTICULATION:
(steady-state, stable patterns of vibrations) and
 Cross-spliced stimuli study
formant transitions (consists of short bursts of sounds
– often coincide with rapid increases or decreases in
Parts of one spoken word have been chopped
frequency)
off and bolted onto a different word.

Single-syllable words divided into onsets and


COARTICULATION EFFECTS ON SPEECH PERCEPTION
codas. /p/ is onset and /ress/ is coda.
 Coarticulation: Articulators are moving
simultaneously but for different phonemes, or Coarticulation means that the way the burst is
that phonemes overlap in time, which pronounced depends on the coda, and the way
explicitly implicate a belief in some sort of the coda is pronounced depends on the burst.
underlying “segment” that has its physical
expression in articulatory behaviour. Result: If two syllables are recorded, and the
 The brain coordinates these individual end of one is spliced onto the beginning of the
articulator movements in a very ingenious way, other, people are more likely to misperceive
such that movements needed for adjacent the coda as matching the original unspliced
vowels and consonants are produced version that the burst came from.
simultaneously.
Thus, eliminating information that comes from
THINGS THAT HAPPEN IN COARTICULATION: coarticulation makes the perceiver’s job
1) When you produce a phonemic sound (i.e. /d/), the harder.
way you make the sound and its physical form is
different when the /d/ sound is followed by THE MOTOR THEORY OF SPEECH
different vowels PERCEPTION
2) Despite the differences of your phonemic 1) Gestures represent the fundamental unit of mental
production and the actual physical properties of representation of speech (Cooper et al, 1952)
sound waves, you perceive the two signals as being  When you speak, you attempt to move your
the same phoneme (perceptual constancy – articulators to particular places in specific ways
different physical patterns are perceived as being  The motor part of the speech production
the same) system takes the sequence of words you want
to say and comes up with a gestural score that
tells your articulators how to move.
3) Speech sounds being spread out over time – not
 By knowing what the gestures are, you can tell
clean breaks between phonemes, acoustic parts
what the set of words was that produced that
of one phonemes may overlap partially or entirely
set of gesture
with other phonemes
4) Silent centre vowels are perceived when
 Alvin Liberman: we map acoustic signals to the
researchers edit a recording to remove part of the
gestures that produced them, because there is
acoustic signal for an utterance. For example,
a closer relationship between gestures and
when the middle part is erased, the word does not
phonemes than there is between acoustic
sound entirely normal, but people will correctly
signals and phonemes
identify which phoneme was present in the original
utterance
2) Speech perception is accomplished by a naturally
As long as the preceding and the following consonants selected module (Fodor, 1983).
carry information that results from coarticulation,
 This speech perception module monitors First stage: syntactic information only
incoming acoustic stimulation and reacts
Second stage: analysis of the semantic information
strongly when the signal contains the
characteristic complex pattern that make A sentence is globally ambiguous if it has two distinct
up speech. interpretations. = The cop chased the criminal with a
 While the speech module recognises an fast car
incoming stimulus as speech, it pre-empts
Local ambiguities persist only for a short amount of
other auditory processing systems,
time as an utterance is heard or written and are
preventing their output from entering
resolved during the course of the utterance so the
consciousness.
complete utterance has only one interpretation. = The
CHAPTER 4: SENTENCE PROCESSING critic wrote the book was enlightening

Globally ambiguous sentence: have sequences that can


be organised in more than one way, and those different MODULAR VS INTERACTIVE
ways of organising the sentence are all consistent with
the grammar of the language MODULAR INTERACTIVE
assumes that each factor assume that all available
Ambiguity leads to involved in sentence information is processed
processing is computed at the same time and
 longer reading times
in its own module, can immediately
 lower comprehension accuracy which has limited means influence the
 different patterns of brain activity in of communication with computation of the final
comprehenders than unambiguous sentences the other modules analysis.
that say the same thing

processing cost depends on a variety of factors:

 what information the listener has just


processed
 what contextual information is available

how to determine the effect of ambiguous sentence to


processing costs on the listener?

 Measure the time taken for someone to


understand the ambiguous sentence

Temporary ambiguous: contain a sequence of words


that can be configured in more than one way, but the
sentence as a whole has only a single grammatically
licensed or acceptable structure

The garden path model theory

 proposed by Frazier and Fodor in 1978


(Christianson et al. 2001).
 suggests that when encountering ambiguous
sentences, only one meaning is initially
processed.
 Then, upon reaching the end of, or a key point
within, the sentence, if the meaning ascribed
does not work the sentence is reparsed until a
satisfactory meaning can be ascribed.
The longer it takes people to understand part of a
sentence, the greater the processing load that part of
Sentence Processing: Traxler Chapter 4
the sentence imposes

Syntactic parsing: mental process or set of processes


1. Describe the relationship between sentence that takes sequences of words and organises them into
structure and sentence meaning. How does hierarchical structures
the way that we organise words in sentences
Syntactic parser: a mechanism that carries out
influence the meanings we assign to those
processes that identify relationships between words in
sentences?
sentences

MODELS OF PARSING: 1) TWO-STAGE


2. What do experiments in which reading times MODELS
are measured tell us about the process of
interpreting a sentence? What do these 1) Listeners build the wrong structure for some
experiments have to say about incrementality temporarily ambiguous sentence while they are
and immediacy? processing the ambiguous part
3. Describe a prominent two-stage of sentence 2) Discover their error when they get to the
processing. What experimental evidence disambiguating information
supports such an account? 3) Revise their initial syntactic and semantic
4. What kinds of information can influence the commitments
process of building a syntactic structure for a
sentence? Give examples of each kind of Garden Path Theory (Two-stage model)
information.
 First stage
5. Explain how constraint-based models differ
o the incoming sequence of words is analysed
from two-stage models. Describe
to determine what categories the words
experiments that support constraint-based
belong to (Noun, Verb, Preposition etc)
models of sentence processing.
o once categories have been identified, the
6. Describe the argument structure hypothesis.
parser can build a syntactic structure for the
How does it compare to two-stage and
sequence
constraint-based accounts. Why might a
 Second stage
person believe in the argument structure
o Standard meaning is computed by applying
hypothesis?
semantic rules to the structured input
7. Describe two alternatives to both the two-
stage and constraint-based account of Thematic interpreter: assign roles to each of the
sentence processing elements in the syntactic three based on their position
8. Describe how long-distance dependencies in the tree and how they are connected to other words
differ from local dependencies. Describe two
 If the thematic interpreter produces a meaning
accounts of long-distance dependency
that lacks one or more of these qualities, one
processing.
remedy is to send signal to the syntactic parser
that prompts the parser to try to find an
alternative structure for the sequence of
Globally ambiguous: have sequences of words that can
words
be organized in more than one way, and those different
ways of organising the sentence are all consistent with GARDEN PATH THEORY: assumes that the parser
the grammar of the language. begins to build a syntactic structure as soon as the
Temporarily ambiguous: contain a sequence of words lexical processor begins to deliver information
that can be configured in more than one way, but the about word categories.
sentence as a whole has only a single grammatically
 Thematic processor work on a word-by-
licensed or acceptable structure.
word basis (semantic interpretation does
not wait until the end of a phrase or a
clause – but the process of interpretation  Disadvantage: heuristics do not always lead to
will slow down or stop as listeners detect the correct solution – resulting in delay in
either syntactic or semantic problems with getting to the correct interpretation as it is
the input. outweighed by the overall time savings that
the heuristics provide
 Rush to interpretation that leads people
astray
 Because decisions are made on a word-by-
Heuristics:
word basis, the parser is forced to choose
between alternative structures when more 1) LATE CLOSURE
than one structure is compatible with the  Do not postulate unnecessary structure. If
input possible, continue to work on the same
phrase or clause as long as possible
Assumptions of Garden Path Theory: 2) MINIMAL ATTACHMENT
 When more than one structure is licensed
 Assumes that people can only build one and consistent with the input, build the
syntactic structure at a time (represents structure with the fewest nodes
serial processing system) 3) MAIN ASSERTION
 Assumes that the overarching principle  Given a choice between two structures,
that the parser relies on is simplicity build the structure where the new
(parser seeks to build the least complicated elements relate to the main assertion of
structure it can) the sentence
o Simple structure take less time to
MODELS OF PARSING: 2) CONSTRAINT-
build than more complicated
structures BASED MODELS
o Simple structure place lower Difference from Garden Path Theory:
demands on cognitive resources
and reduced demands on working  CB parsers are capable of pursuing
memory also translates into multiple structural possibilities
greater speed simultaneously (parallel processing)
 Garden path parser relies solely on
Consequences following the assumptions: word category information for its
 For garden path theory to work as a inputs, but constraint-based persers
general theory of parsing, it needs to can draw on a much wider variety of
explain how people make choices when cues to decide what structures to
more than one syntactic structure is build and the relative emphasis to
possible place on each alternative structure
 For garden path theory to work as a theory  Referred to as ‘one-stage’ models
about how people parse sentences, there because lexical, syntactic, and
should be evidence that people have semantic processes are all viewed as
problems at just those points in sentences taking place simultaneously
where garden path theory says structural Story context effects:
reanalysis is taking place
1) Problems for classic garden path theory:
 If the parser only pays attention to word
Two heuristics (basic rules that can be applied quickly category information to build syntactic
and consistently to make decisions about which structure for a sentence, then the information
structure to build at any given point) that appears in preceding sentences should
have no effect on the initial processing of a
 Advantage: decisions can be made very quickly
on the basis of incomplete information given sentence
 According to referential context account, the information, and the semantic properties of
parser can use contextual information to specific words.
decide which syntactic structure it will favour
ARGUMENT STRUCTURE HYPOTHESIS
at a given point in time (create a story)
 Contrary to what the garden path theory  Argument frames and their corresponding
predicts, the parser does seem to pay syntactic structures are important because
attention to information that context makes they determine how some elements of
available at least some of the time to make sentences are interpreted
decisions about which syntactic structure to  Contends that the subcategory properties of
build for a new sentence the verb determine how the prepositional
phrase is interpreted
LIMITATIONS, CRITICISMS, AND SOME
Subcategory frequency effects
ALTERNATIVE PARSING THEORIES
 Some words are in the category Verb, but they
(for constraint-based theory)
serve different purpose (transitive,
intransitive) – they belong different  The parser may not always favour likely
subcategories structures over less likely, but simple
 Constraint-based theory says that structural structures
information is associated with individual  Absence of evidence that sentences with
words in the lexicon and this information simple syntactic structures are ever hard to
influences which structural hypotheses will be process
pursued as sentences are being processed - it
will use subcategory information to determine HARLEY – CHAPTER 10
which structural analysis to favour when more
than one structure is consistent with the input  Recording eye movement – we spend no
 Constraint-based parser takes information longer reading the ambiguous regions of
about the past and uses it to predict the future sentences than the unambiguous regions
(predicts direct object after seeing ditransitive of control sentences, but we spend longer
verb) in reading the disambiguation region

 Cross-linguistic frequency data Serial autonomous model: construct parse


 Semantic effects trees on purely syntactic grounds, and then
 Prosody decide using semantic information whether it
 Visual Context effects makes sense or not

SUMMARY (PART 1) Parallel autonomous model: construct all


possible syntactic representations in parallel
 A constraint-based parser can activate (using solely syntactic information), and then
multiple syntactic structures simultaneously use semantic or other information to choose
 It ranks different structures based on how most appropriate one (Mitchell, 1994)
much evidence is available for each in the
Disambiguated = resolved
input
 Evidence for a given structure and its Reduced relative clause = one that modifies
accompanying semantic interpretation can the main noun, and it is “reduced” because it
come from multiple sources, including story lacks the relative pronoun “which” or “that”
context, visual context, subcategory
 Deletion of punctuation, complementiser within the clause currently being processed,
(that) can also produce misleading results independent of the number of words in the
 Disfluencies (prosodic cues) influence the clause (Caplan, 1972)
way in which people interpret garden path  The processing load is highest at the end of
sentences the clause, and eye fixations are longer on
o When interruption (saying “uh”) the final word of a clause (Just & Carpenter,
comes before an unambiguous NP, 1980)
listeners are more likely to think
Techniques to explore the size of syntactic unit:
that the NP is the subject of a new
clause rather than the object of an 1) Click displacement technique (Fodor & Bever,
old one 1965; Garret, Bever, & Fodor, 1966)

Factors for Garden Path Theory: The major processing units resist interruption:

1) Serial autonomy model:  We finish what we are doing, then process


 The single syntactic representation we are other material at the first suitable
constructing on syntactic grounds turns out opportunity
to be incorrect
2) Parallel autonomy model:
 One representation is much more active than Methodology  Participants heard speech over
the others because of the strength of the headphones in one ear, and at certain points in the
syntactic cues (but this turns out to be sentence, extraneous clicks were presented in the
wrong) other ear
3) Interactive model:
 Various sources of information support the
analysis more than its alternative That he was* happy was evident from the way
he smiled
Earlier theories:
Conclusion  the clause is a major unit of
 Transformation: Savin and Perchonock perceptual and syntactic processing
(1965) found that sentences with more
transformations in them took up more Problems  click migration occurs in the response
memory space stage; reflect the operations of memory rather
than the operations of syntactic processing
Slobin (1966)  examined the processing of…
2) Anticipatory Eye Movements
1) Reversible passive: where subject and
object of the sentence still makes People anticipate information based on semantic
pragmatic sense associations to the syntactic structure
2) Irreversible passive: one that does not Conclusion  language processor operates
make sense after this reversal incrementally; (Pickering, 1999)
Size of parsing: i) it rapidly constructs a syntactical analysis
 Jarvella (19721): listeners only begin to purge for a sentence fragment
memory of the details of syntactic ii) assigns it a semantic interpretation
constituents after a sentence boundary has iii) relates this interpretation to world
been passed knowledge
 There is a clause boundary effect in recalling
words: it is easiest to recall words from
3) Surface Structure Cues readers only consider all relevant
initially one syntactic information is used
Parsing strategies that uses only syntactic cues
structure for any given and several syntactical
(the, a before NP) and S-V-O word order. sentence and meaning meanings are initially
Kimball (1973): 7 principles to sentence parsing is not involved in the considered before the
(SAUSAGE MACHINE) selection of most appropriate is
preliminary syntactical selected.
i) Parsing top-down (except the presence meaning.
of ‘and’) (start from the sentence node Fundamental Processor uses
and predict constituents) principles of parsing: multiple sources of
ii) Right association (new words are information, including
preferentially attached to the lowest  Minimal syntactic, semantic,
possible node in the structure attachment – discourse, and
incoming frequency-based
constructed so far)
material should (called constraints)
iii) New nodes (functional words signal a
be attached to
new phrase)
the phrase ^^ Garden path occurs
iv) The processor can only cope with marker being when the correct
nodes associated with two sentences constructed analysis of a local
nodes at any one time using the fewest ambiguity receives
v) Closure (the processor prefers to close nodes possible little activation
a phrase as soon as possible)
vi) Fixed structure (having a closed path, it  Late closure –
is computationally costly to reopen it incoming
and reorganised the previously close material should
constituents) be incorporated
vii) Processing (when a phrase is closed it into the clause or
exits from short-term memory and is phrase currently
being processed
passed on to a second stage of deeper,
semantic processing)
SAUSAGE MACHINE – it divides the language input
into something that looks like a link of sausages.
i) First stage: preliminary phrase  Ferreira & Clifton argued that semantic
packager (PPP) information does not prevent or cause
ii) Second stage: Sentence structure garden-pathing but can hasten recovery from
supervisor (SSS) it

Problems with Sausage Model  Wanner (1980)


 First stage of parsing is short-sighted and
does not use semantic or thematic
information
..........................................

Garden Path VS Constraint-based Model

Garden Path Constraint-based


Autonomous Interactive
Fraizer & Rayner MacDonald (1994)
(1982)
SUMMARY  Some aphasics show difficulties in parsing
when they cannot rely on semantic
o Two ideas from Chomsky’s original work
information
that were picked up by early
psycholinguists were the derivational
theory of complexity and the autonomy of
syntax
o The earliest experiments supported the Syntactic Processing
idea that the more transformationally A. syntactic parser: (together with the grammar)
complex a sentence, the longer it took to guides the order of which elements of a sentence are
process; however, experiments using processed and the manner in which syntactic
psychologically more realistic tasks failed structure is built up.
to replicate these findings
B. Some grammatically complex sentences are easy to
o Although linguistic theory influenced early
parse and some grammatically easy sentences are
accounts of parsing, linguistics and
hard to parse.
psycholinguistics soon parted ways
o The clause is an important unit of syntactic Ex: (complex sentence but easy parsing) Sarah
processing saw the goblin who displeased Jareth the
other day.
o In autonomous structures, only syntactic
information is used to construct and select Ex: (easy sentence but hard parsing) The
among alternative syntactic structures, in horse raced past the barn fell.
interactive models non-syntactic
C. garden-path sentences: sentences which are easy
information is used in the selection process for the grammar to produce, but hard for the parser.
o Psycholinguists have particularly studied The parser is “led down the garden path” to the
how we understand ambiguous sentences, wrong structure.
such as garden path constructions
Ex: The horse raced past the barn fell.
o One of the most studied types of garden
path sentences is the reduced relative (The Parser’s garden path: “The horse raced past
horse raced past the barn fell) the barn” = simple sentence.
o Early models of parsing focused on parsing “The horse raced past the barn fell” = Ack!
strategies sing syntactic cues
o Kimball proposed 7 surface structure
parsing strategies Correct Parse: “The horse raced past the
barn” = “The horse which was raced past the

barn” “raced past the barn”


 Late closure: we prefer to attach incoming
modifies “the horse”.
material to the clause or phrase currently
being processed “The horse raced past the barn fell” = The
 Minimal attachment: we prefer the horse (modifier) fell.
simplest construction, where simple
means the structure that creates the
B. Garden path sentences reveal the preferences of
minimum number of syntactic nodes
the parser.

 In constraint-based models, lexical and Ex: Given “The horse raced past the barn”, the
syntactic ambiguity are considered to be parser can choose a simple Subj V PP
fundamentally the same thing, and structure, or a Subj + modifying phrase
resolved by similar mechanisms structure. The parsers chooses a simple Subj V
PP structure because it prefers this analysis.
https://books.google.co.uk/books?id=jYzDAAAAQBAJ
&pg=PA46&lpg=PA46&dq=literature+review+sentenc
e+processing&source=bl&ots=3gUga0E6a2&sig=yH9t7
XG2lplY9eTJt2La4ak5kZQ&hl=en&sa=X&ved=0ahUKE
wiRhuWTyZDSAhViCMAKHZwhAy8Q6AEIXjAJ#v=onep
age&q&f=false

Garden-Path Non-Garden Non-Garden


Path Yes Path No
While the While the  When there is a temporary structural
skipper sailed skipper sailed ambiguity, the parser does not wait until
the boat the ship the
the end of the sentence before analysing it,
veered off boat veered off
course course but bets on one analysis
While Tom While Tom While Tm grilled
grilled the grilled the the hamburger
hotdog hamburger the the hotdog
burned hotdog began began to burn
to burn
CHRISTIANSON EXPERIMENT
In one such study, Christianson et al. (2001)
*if there’s consecutive nouns/np/dp it is easier for the examined the traditional assumption that full
subject to determine that it is non-garden path resolution of a local syntactic ambiguity is
sentence necessarily performed as part of the process of
*SVO – participants will assume ‘the hotdog’ as the deriving a robust interpretation of a sentence.
object following the SVO pattern of English Christianson et al. asked participantsto read
 The minimal attachment principle states that “for sentences containing ambiguities as in (1) and
any sequence NP-V-NP (noun phrase-verb-noun then answer comprehension questions as in (2).
phrase), the second NP will be interpreted as a direct
(2) Did Anna dress the baby?
object” (Ferreira, Henderson 1990).

*all the verbs are transitive – relate with SVO pattern Christianson et al. reasoned that, if readers
construct the interpretation that is compatible
 Frazir (1978) suggests that the structure associated with the global structure of the sentence, the
with the preferred interpretation is initially chosen
answer to this question must be ‘‘No’’ as the baby
because is guided by formal principles like Minimal
is the subject of the matrix clause but not the
Attachment, which states that the parse using the
object of the embedded clause, whereas the
fewest nodes consistent with the rules of the
language should be preferred. answer to a related question (Christianson et al.,
2006) exemplified in (3) must be ‘‘Yes.’’
(3) Did Anna dress herself?
Problem:
Christianson et al. (2001, 2006) demonstrated in a
1) If we ask the participant to read out loud,
numberof experiments, however, that syntactic
they’ll read with intonation and that might
manipulations of the garden path (like below)
give them a hint to determine the sentence
2) If we ask them to read in silence, we don’t affected accuracy rates on follow-up
know whether they’ll read more than once comprehension questions.
o clause order Two ways to parse ambiguous sentence:
o disambiguation
1) In the first, the ambiguous choice point in the
o length of ambiguous region
parse is detected (either as a misparse, or as an
None of these manipulations altered the lexical alternative disfavoured parse), and reanalysis
content of the sentences, nor the inferences likely occurs, bringing the entire structure into
to be drawn from those collections of lexical items. compliance with the grammar and generating the
correct semantic interpretation for the string.
Moreover, Christianson et al. (2006) found that
the likelihood of answering questions such as (3) 2) The second possibility is that the ambiguity is
correctly was related to readers’ working not noticed at all. This could happen if, at the
memory capacity. ambiguous choice point, the parser initially settled
on the eventually correct interpretation and
The authors therefore argued that full reanalysis
thereby avoided the need for reprocessing.
required more cognitive effort than readers with
lower working memories were willing or able to
expend, and so these readers settled for ‘‘Good
Enough’’ reanalysis and interpretations (Ferreira,
Bailey, & Ferraro, 2002; Ferreira & Patson, 2007;
Ferreira et al., 2001).
Several recent studies have corroborated
Christianson et al.’s observation that misanalyses
appear to linger (Kaschak & Glenberg, 2004;
Staub, 2007a; Sturt, 2007; van Gompel, Pickering,
Pearson, & Jacob, 2006), though researchers differ
on why this might be.
1) Christianson et al. (2001; see also
Christianson et al., 2006) attributed the
lingering effect to a failure to fully
reanalyze the partial, and ultimately
incorrect, syntactic structure constructed
during the initial parse of the garden path
sentence.
2) However, Kaschak and Glenberg (2004)
suggested that the persistence of previous
syntactic structure stems from episodic
memory traces of the initial structure,
3) whereas Sturt (2007) suggested that the
effects are due to semantic persistence.
4) Van Gompel et al. (2006) observed that the
structure of initial misanalyses of the sort
of garden-path sentences used in
Christianson et al. (2001) primed
subsequent production, but remained
agnostic (doubtful) as to whether the
source of this structural priming was
episodic memory or incomplete syntactic
reanalysis
o Constraint-based model: phrase the deer is
initially treated as the object of the hunted
THEMATIC ROLES ASSIGNED ALONG THE GARDEN
because the verb hunt may be frequently
PATH LINGER
transitive than intransitive and the
1) Both the length of the ambiguous region sentence lacks comma
and the plausibility of the ultimate
interpretation affected the likelihood that
sentences would be fully reanalysed # If analysis is successful = all steps described
2) Compared garden path sentences with above are completed and an interpretation of the
non-garden path sentences and compared sentence licensed by the syntax and consistent
performance on two different types of with the input string is produced
comprehension questions
3) Use small class of syntactically unique verb
(to provide evidence against the position # if not, sentence-level interpretation cannot be
that people employ pragmatic inference derived, or at least cannot be derived
when faces with syntactically difficult systematically, consistently, and confidently
garden paths)
 Reanalysis of such sentences is not always
complete, so that comprehenders often  If a sentence is an “easy” garden path, the
derive an interpretation for the full assumption is that people fully reanalyse
sentence in which part of the initial the sentence
misanalysis persists.  If a sentence is a “difficult” garden path, it
is generally assumed that, although some
Assumption #1: sentence meaning is derived from
sort of initial parsing mistake is
a complete representation built up from the
recognised, recovery is hampered or
component parts of an utterance into a fully
prevented due either to syntactic
specified syntactic structure
limitations on reanalysis or to activation
#gardenpath: levels of alternate structures falling close
to zero
o Garden path sentences result in some sort
of misanalysis in the initial syntactic
analysis or parse
METHODOLOGY:
o By investigating how garden path
sentences are initially parsed, reanalysed, 1) Present a garden-path sentence followed
and ultimately comprehended, we can gain by a question probing the interpretation of
insight into the underlying mechanisms that sentence
that’s usually operate accurately 2) Manipulated a number of properties of
garden path and non-garden path
Late closure: the incoming material is attached
sentences to determine which factors
inside the clause or phrase currently being
facilitated full reanalysis resulting in
processed
syntactically licensed interpretations of
Early closure: closing off the subordinate verb’s specific problematic constituent phrases
argument structure despite the presence of an and which blocked or partially blocked
apparently compatible object in the input string such interpretations
RESULTS (PREVIEW):
1) Initial thematic role assignments were
surprisingly resistant to revision, this was
While the man hunted the deer ran into the woods.
reflected in participants’ interpretations of  misanalysis in addition to pragmatic
the sentences plausibility
a. Interpretations based on the initial,
incorrect parse were quire
persistent and confidently held, GENERAL DISCUSSION
even though such interpretations
 much can be learned about the
reflected only partial reanalysis and
architecture of the language processing
depended globally unlicensed
system from comprehension measures
syntactic structures
2) Initial interpretations preserved was
influenced by syntactic factors that
Result:
created, exaggerated, or eliminated the
garden path  participants were accurate at answering
a. Ambiguous region length questions about the subjects and action of the
b. Head position within the main clause of a garden path sentence
ambiguous phrase o reveals that the parser and
c. Clause order comprehension system engaged in
d. Comma use successful reanalysis to some extent (it
e. Verb argument structure could not have been answered
f. Plausibility (non-syntactic factor) accurately with only initial analysis and
located an NP for the second clause
 comprehension system has some tendency to
Ferreira and Henderson (1991, 1998): the longer allow the interpretation based on the initial
one is committed to (ultimately) incorrect syntactic analysis of the subordinate clause to
thematic assignments, the more difficult full linger
reanalysis is.  misinterpretation of garden path sentences
 If yes responses reflect an interpretation
based on the perseverance of the initial
thematic assignments, the proportion of
incorrect ‘yes’ responses should be higher
when the ambiguous region is longer
 Length of the ambiguous region should
have a greater effect on responses to
garden path compared to non-garden
path sentences
o Because in in garden path sentences

Ferreira and Henderson (1991, 1998): if the head


of a phrase occurs early within it, reanalysis is
much harder than if the head occurs at the end of
the phrase

 Reverse the order of the subordinate main


clause – to assess further the extent to
which the comprehenders’ tendency to
give an answer due to the syntactic

You might also like