Language and cognition

Lecture 2 Language and the Brain

Cognition - internal brain processes (Linguistics as part of cognitive science)

4 approaches:
Psycholinguistics - study of language processes through experiments
Neurolinguistics - study of language processes in patients
Neuroscience of language - study of language processes in brain

Computational (psycho)linguistics - creation of computer models to better

understand language processes

Sim to psycholinguistics:
• Cognitive task (stimuli & responses - measures behaviour in terms of
speed (reaction times) and accuracy (correctness))
Theory (although!)
• healthy subjects (but also patients)
• Experimental in nature

But also differences:

- Complex techniques to measure brain activity

- Complexe statistics to process images
- Structure
Neuroscience of language

Main limitation of single-unit recording:

Micro electrodes, 1mm/110.000, control of indv nerve cells

Invasive and rarely used

EEG / ERP

- image of brain work (measuring of brain waves, only simple processes, bad
spacial resolution (image distorted), excellent temporal resolution)

Within few milisec-> direct

ERP (Event-related potential technique) - language processing

Take average

PET - position emission tomography -> radioactive substance moves to brain,

collision with neurones-where?

+ Spacial resolution
-Temporal delay, indirect, invasive
fMRI - Magnetic resonance imaging (localisation of internal cognitive х
processes)

BOLD: blood-oxygen level-dependent contrast

Brain needs oxygenated blood with specific magnetic properties that scanner
can capture
Limits: bad temporal resolution, indirect, little room for individual variation,
stimuli limitations (visual and auditory)
Small temporal delay of activity (4-6 milesecs)

Why?

MEG - Magnetoencephalography

Good temporal and spatial resolution (worse than MRI)

-> direct

TMS - Transcranial magnetic stimulation

Coil administers short magnetic pulse- temporal lesion
Whether a part. region is necessary for a cognitive task?

Why?
+:

Causal relations, more flexible (injured/not), diff regions of lesion, info what
region is active

-:

Complex to understand brain (sometimes improving behaviour through

compensation), no exact knowledge of precise brain region because of
connections btw remote brain regions, limited location use, small risks

tDCS
Sends electricity to brain, few risks

Anodal tDCS Improves cognitive function (positive stimulus) > therapy

Brain – complex matter

~100 billion neurons

Huge connectivity of neurons 1 with 10 000 others

Brain = central nervous system

Of 3 parts:

• Cerebrum (‘big brain’) -> cerebral cortex (l and r), limbic system (l and
r), basal ganglia (l and r)
• Cerebellum -> left and right (symmetrical)
• Brain stem
Cerebellum ’s have left and right hemisphere, structure identical -> small
asymmetries can occur (facial, limb etc)

Individual differences in functioning often reflect individual differences in

connections
2 principles of organisation of complex system with relation to human brain:
- Principle of cost control -> maximum number of short-distance
connections
- Principle of efficiency -> maximum number of long-distance
connections
-
Trade of btw cost control and efficiency
- A number of modules or small areas with densely clustered (short-distance)
connections
- A number of hubs or areas with large numbers of (long-distance)
connections to other areas

*Airport

4 strengths of the cognitive neuroscientific approach of language:

• Empirical verification of theoretical controversies
• Complexity of cognitive systems mapped
• More than 10000 studies son fMRI published
• Re-analysis on previously obtained data based on new insights
9 limitations:

• Problem of ecological validity and task specificity -> lab is not a real
world!
• Correlation, not causation
• Overinterpretation of effects and reverse inference (involvement, not
necessity)
• Brain functioning more complex than language to describe it
• False positives due to many data points
• Replication problems
• Sometimes less activation, in cognitive processes (default mode
network)
 • Testing of theories is rare, often explanatory
• Neuroenchantment
Functional specialisation -> functional integration (activation of other
regions)
Inhibit or excite neurons

Lecture 3 (to be finished with audio lecture aw)

Computational linguistics – designing computer models to simulate

human cognition and see how it works.
What is the difference between computational modeling of cognition and
artificial intelligence?

Putting theories into a computer program

AI – constructing programs that create intelligence superior to human one.

Connectionist models – distributed models (other: localized)

Simulating neural networks

Consist of units that are connected to each other, they can excite or inhibit
They form different layers: input/intermediate/output
No explicit rules are programmed
Back-propagation = association between input and output is learned through
comparison actual responses versus correct ones

Limitations:

• Simultaneous presentation of multiple stimuli leads to ‘superposition

catastrophe’
 • Unable to learn general rules, only discover associations/patterns
• Brain is more complex than those models
• Back-propagation is a slow form of learning >< humans also exhibit
one-trial learning

Production systems vs Connectionist models

On the basis of ‘production rules’ – rule-based learning

A working memory and conflict resolution strategy

Advantages of computational approach

• Contributes to accurate theorizing
• Contributes to broad theory building (more than one domain)
• Integrates not only behavioural data but also neuroimaging data > check
whether hypotheses are plausible in reality
• New models build on existing ones (theory from psycholinguistics)

Limitations of computational approach:

• Complexity of human cognition
• Some models become too complex and incomprehensible
• Difficult to falsify(?)
• Overfitting of models > no predictive power
• Each programmer has own way of programming, not always transparent
and easily evaluated

Evaluation of approaches:
• Increasingly more research requires a combination of multiple
approaches.
• Each approach has unique contribution
• Each approach has pros and cons. The value of a specific theory
increases when multiple different approaches converge to the same
conclusion. (= convergent operations)
How to get out of replication crisis
• Meta-analyses: integrating different studies
• Pre-registration/Registered reports
Lecture 4 Listening (to be finished with audio lecture aw)

5 sub-components essential to listening as a skill

Models of listening comprehension:

1.

Basic unit of listening – phoneme or syllable

Phoneme – sound with meaning-distinctive function

Allophone – sound without meaning-distinctive function
2.

5 problems that complicate listening process:

Problems for the brain in decoding

1. Problem of segmentation
2. Problem of coarticulation
3. Problem of individual differences
4. Problem of speed of spoken language
5. Problem of speech errors

Coarticulation-> predict next morpheme

Fluidity by articulators
Mcgurk effect -> top-down processing

Our brain hears smth that is not there

Importance of lip-reading to support listening

Problem of speaking errors is overcome by top-down proccesses that correct

what was said incorrectly

Problem of segmentation
Overcome by:

1. Lexical clues (word knowledge, sentence context)

2. Segmental clues (phonotaxis, coarticulation)
3. Word accent

TRACE model
3 tears : word units, phoneme units, feature units

In between units – facilitation

Word superiority effect: phonemes better recognized in words than non-
words > top-down effect

Phonemic restoration effect: incomplete word -> brain fills this gap and
produces sounds that are not there based on the context

e.g It was found that the *eel was on the shoe

Lexical identification shift or Ganong effect: ambiguous sounds are assigned
to part. phoneme when they lead to a word
> Processing of ph. and words occur integrated (parallel) rather than nicely
demarcated from each other (serial)
Predictions

Context very important in word interpretation

Interaction model = context influences comprehension of words early in

listening process
Autonomous model = context influences comprehension of words very late in
the listening process

Interaction model most in line with empirical evidence

BUT 2 caveats:
- Less context-word interactions when speech is clear and not ambiguous
- Interaction context-word especially in predictable contexts
4 strengths of TRACE model:

1. High impact, many citations

2. Explains many effects such as phonemic restoration effect, Ganong
effect, categorical perception and word superiority effect
3. Combination of bottom-up and top-down processes > interaction model
4. Performs very well if it sees a lot of noise in the imput (degraded speech)

6 limitations of TRACE model:

1. Limited scope because concerns only word recognition and says nothing
about understanding
2. Lacks a level of conceptual meaning that can be triggered by top-down
processes
 3. Too much emphasis on top-down processes, our brain does not correct
everything (e.g. mispronunciation and ambiguous sounds)
4. Consists of many theoretical assumptions (immune to criticism?)
5. Model is incomplete (e.g. Orthographic influences and effect of
cognitive load)
6. Studies show that phoneme level can be activated as early as 50ms after
stimulus-onset)
Characteristics of Cohort model:

According to Cohort model word recognition is a competitive process – diff.

lexical items are activated - and after a dropout race, the correct word is finally
recognized
The cohort = all words corresponding to the phonemes of the word already
pronounced (or similar phonemes according to the new model)

Word recognition when point of uniqueness is reached = only one word

matches the phonemes heard

e.g Crocodile - crockery

Cohort model: different stages

1. Access phase to word cohort

2. Selection phase: one word from the word cohort is activated
3. Integration phase: integration within sentence
Only in integration phase > more emphasis in this model on bottom-up effects
5 strengths of the Cohort model:
1. Word recognition is indeed a competitive process
2. Word recognition is sequential and changes throughout the recognition
of individual phonemes
3. Importance of point of uniqueness
4. Context effects sometimes (but not always) during integration phase
5. The new version of the model is better than the previous version
(similar phonemes may also belong to the cohort)
3 limitations of the Cohort model:
1. Context sometimes has an earlier effect on word recognition, esp. if the
word is very predictable within the sentence context or if the input is
of poor quality
2. Importance of word meaning underemphasised (concrete vs. abstract
words)
3. Top-down processes such as prediction of next segments receive too
little attention

Effect of spelling on listening comprehension

Automatic activation of orthographic code of a word during the listening

process about 300-350 ms. after presentation
Spoken word recognition slower for words both with many orthographic
neighbours and many phonological neighbours
e.g. Cat – lat Cat - chat

Motor theory
• Listeners imitate speakers’ articulation movements > McGurk effect
• During listening activity in motor cortex
• Mirror neurons
Finding that support the Motor theory:
1. Neuroscientific evidence showing that brain regions involved in
listening and speaking overlap
2. ERP studies showing rapid activation of motor regions in listening
3. Neurolinguistics evidence showing that listening is impaired following
damage to speech regions
 4. TMS evidence showing that listening is negatively affected upon
stimulation of speech regions

Limitations of the Motor theory:

1. Motor regions help predict the meaning of what is going to be said but
not the word itself
2. We do not understand how motor regions work together to select
correct lexical items
3. Certain listening regions not involved in speaking
4. Motor regions especially important when spoken input is not clear

Auditory analysis system impaired > pure word deafness

=> dissociation between perception of speech affected and other sounds

=> dissociation between listening on the one hand and speaking, reading and
writing on the other hand

=> problems with rapid changes in auditory input

Routes distinguished in neurolinguistics 3-route theoretical framework of the

listening process:
1. Semantic
2. Lexical
3. Sublexical

Strengths and limitations:

1. Patients have diff. problems with speech perceptions, evidence for the
3 pathways
2. But, sometimes difficult to make symptoms fit model (deep dysphasia)
3. But, explains components but not how they work (we need theories or
that)

Word meaning deafness

=> dissociation recognise familiar words and access their meaning (affected)
=> dissociation repeating words and non-words (affected)

=> dissociation understanding >< written language (not affected)

Lexical route still intact

Transcortical sensory aphasia

=> dissociation repetition words and non-words on the one hand and
comprehension on the other (impaired)
Sublexical route still intact
Deep dysphasia
=> unable to repeat words and non-words, but repeat semantically similar
words (e.g. air instead of cloud)
=> underlying phonological problems

Semantic route still intact

Lecture 5 – Reading

Sub-processes of reading
Sources of knowledge present to understand a written text:

• Word meaning (lexicon)

• Spelling
• Knowledge about the world
• Linguistic knowledge
The sequence of these cognitive processes is flexible

Phonological processe

We are taught reading on the basis of speech

Evidence:
• Inner voice
• Homophones
• Phonological neighborhoood

neurolinguistics shows that dissociation between phonological processing

and reading comprehension is possible
➔ Supportive, not essential/ crucial
Word Recognition.

Strengths of Interactive Actvation model:

• We use all the info that comes in to recognize words
• Bottom-up and top-down processes
• Explains word superiority effect and effect ortographic neighbors

Limitations:

• No attention for semantics & context

• No attention for phonological processes
• Too much attention to position of a grapheme in a word
• especially good for predicting processing of short words (max 4 lettrs)

Semantic priming

• A word is recognized faster if preceded by a semantically related word

• Within a sentence, several words are semantically related so that
facilitates word recognition

Lexical predictability (exact word) – only if highly predictable >< gradual

predictability (approximate meaning or word)
Eye-tracking is one of 3 major methodologies in reading research

Studies the reading of sentences or long stretches of text

2 others are:

lexical decision task – looks at the relationship between orthography &

semantics

naming task – looks at the relationship between orthography and

phonological

Basic processes involved in reading texts:

• Saccades = 20-30ms
• Fixations 200-250ms, on 80% content words – 20% function words
• Spillover effect – longer fixes if the previous word was difficult
• Perceptual span

E-Z reader model – is a serial model. Only one word is processed before
moving to the next >< SWIFT is a parallel model. Parallel processing of
current, previous & next word.

1. Readers check the familiarity of the fixed word (word n)

2. Completion of familiarity check is the first step of lexical access
3. Then comes the second stage in which the semantics and phonology of
the word are retrieved
4. Both phases are faster with "easy" words (short, high-frequency,
predictable)
5. Upon completion of both phases, attention shifts to word n+1

Strengths:
1. Eye movements offer a good insight into the process of information
processing during reading
2. Model emphasizes importance of factors such as word frequency and
predictability
 3. Evidence that lexical processing consists of two phases
4. Processing next word during fixation on previous word facilitates
processing
5. Strong connection between eye movements and brain
6. No effect of next word on previous word -> serial rather than parallel

Limitations:
1. Mainly focus on when and what regarding eye movements, but little
focus on cognitive functions.
2. Little attention to higher-order processes of reading comprehension,
such as inference, integration of information and use of knowledge
sources during text comprehension
3. Unclear how the familiarity check happens (bottom-up or top-down)
4. Still more parallel processing than thought? (see OB1 reader)

Route 1 : grapheme- phoneme conversion. Uses

regular spelling

Route 2: lexicon+semantics. lexical

representations are activated
Route 3: Lexicon. Same as route 2 but without
semantics

Main reading disorders:

Surface dyslexia
• Problems with irregularly spelled words
 • No problems with regularly spelled words or non-words
• Route 1
Phonological dyslexia

• Problems with unfamiliar words and non-words

• No problems with familiar words, regardless of spelling.
• Route 2&3

Deep dyslexia
• Problems with unfamiliar words and non-words
• Semantic reading errors with high-frequency words: e.g. ship instead
of boat
• Does not fit well within the dual-route model
Strengths of dual route cascaded model:

1. first theory explaining brain disorders with reading

2. been very influential and supported by various types of dyslexia

3. existence of two routes with both parallel and serial processes is
supported by evidence from non-patients

Limitations:

1. little attention to semantics (part of lexical route)

2. focusing primarily on reading words rather than reading sentences
3. makes no statements about acquisition of reading skills, and
grapheme-phoneme rules
5. little attention to phonological processes (fast and automatic, see
earlier)
6. little attention to individual differences
7. 30 parameters, so complex model
8. applicability to languages other than English (e.g., French with many
"silent" graphemes)
9. is mainly focused on reading of one syllable words
Lecture 6 – Speaking (to be finished with audio lecture
aw)

Speaking and writing – productive skills

Language activity that offers a mixture of both skills – chatting/messaging

Info from neurolinguistic approaches about those skills:

• Dissociation between speaking and writing skills
• no total overlap in the brain
• only higher-order processes show similarities (planning, use of
knowledge,)

Fundamental aspects:
• Speaking use processes of listening skills
• Much more difficult that they first appear
• Consists of series of sub-skills
Speaking:

• Syntactic priming: reuse of syntactic structures used by other speaker

• Preformulation: ~70% fixed formulas
• Subspecification: use of simple expressions to fill speech

Sub-processes in language production process:

• Conceptualising - planning - conceptual/semantic level

• Formulating (internally!) - grammatical/syntactic - morphological/
morphemes – phonological/phonemes
• Articulate
• Monitoring
Flexibility of planning process -> circumstances

Incremental planning: planning while speaking

Factors that affect flexibility of PP:

- Speaking slowly
- Simple vs. Complex sentences
- Low cognitive (or time) pressure vs. High cognitive pressure
Cognitive effort as low as possible (little planning) vs. Accuracy (lots)

-> Planning is a higher-order process

Conflict-based monitoring theory

Conflict-based account predicts that speakers detect own errors by speaking
skills

Perceptual loop theory

Perceptual loop predicts speakers’ own error detected by listening skills

Some evidence for perceptual loop, but still mainly for conflict-based theory:

1. Neurolinguistic evidence: success of aphasia patients in detecting own

errors depends on speaking skills
2. Psycholinguistic evidence: error detection happens very quickly so
probably in early phase of fluency
3. Neuroscientific evidence: regions involved in error detection are more
likely to be conflict-based (cognitive control) than perceptual loop
(listening skills)
Lexical network

- Different levels need to work together

- Activation precedes selection (competitive)

Spreading-activation theory (Dell, 1986; 2013)

• Based on study of speech errors

• Emphasise on parallel and cascading processes, interactive and flexible
• Bottom-up feedback

Characteristics:
 • Spreading activation between and within levels: conceptual/semantic -
syntactic - morphological - phonological; and also within a sentence
• Categorical rules at all levels (e.g. sentence order)
• Lexicon (connectionist) with nodes for all levels
(concepts, words, morphemes, phonemes)
• Insertion rules: highest activated item is
chosen/selected
=> BUT due to spreading activation, several neurons (nodes) are active
together, and sometimes a wrong word is selected too early

5 findings that support spreading-activation theory:

1. Mixed-error effect: incorrect word both semantically and phonemically

related to correct word
2. Lexical bias effect: pronunciation errors are mostly existing words
3. Anticipatory errors: sound pronounced too early (e.g. a Tanadees from
Toronto)
4. Switching errors: (e.g ‘I need to send a woman to my e-mail’)
5. Words are often exchanged a short distance apart

5 strengths of spreading-activation theory:

1. Mixed error and lexical bias can be explained by interactivity
2. Oher speaking errors also explained
3. Link with other cognitive activities, such as listening skills (cf. TRACE
model)
4. Our ability to make new sentences is explained by this
5. Theory predicts too many errors and thus needs to be complimented by
conflict-based monitoring theory (speaking skills error detection)

4 limitations:
 1. Little attention to meaning construction and intended effect speaking
skill (semantics and pragmatics)
2. Nothing about timing of speech processes
3. Applies mainly to speech errors and less to normal speech
4. Unclear how interactive speech processes are, and unclear which
contexts determine degree of interaction

WEAVER ++ (Levelt et al., 1999)

• Based on lab research on image naming
• Emphasis on serial processes, feedforward and structured

Characteristics:
• Also diffusion of activation, but feedforward
• 3 levels: lexical concepts > lemmas (abstract representations already
syntactically and semantically but not yet phonologically completed) >
word forms (morphemes and phonemes)
• Lexical selection = competitive process
• Serial processing
• Speech errors are avoided by monitoring

4 strengths of WEAVER ++ theory:

1. Speaking skills consist of a number of sub-processes

2. Shift from studying errors to normal fluency (precise timing)
3. Simple and elegant model with clear predictions
4. Fluency indeed embraces a competitive process

4 limitations:
1. Focus on word level and little explanation of sentence level
2. More interaction between levels than thought (neuroscientific
evidence)
3. More interaction between levels than thought (speech errors)
 4. Lemma level is controversial

Compromise:

• Limited interaction
• Too much interaction (spreading activation) > many speaking errors
• Too little interaction > no production of new sentences and creativity
• Minimal distraction and interference > WEAVER ++
Traditional distinctions made in the study of aphasia and symptoms of each of
these forms:

• Broca’s aphasia
=> slow, non-fluent speech

=> dissociation between production and comprehension sentences

=> BA44 and BA45 in the inferior frontal gyros

• Wernicke’s aphasia

=> fluent and apparently grammatical speech but without meaning

=> problems with speech comprehension

=> posterior part BA22 – superior temporal gyros

5 points of nuance made to this traditional distinction:

1. No consensus on exact locations of both regions

2. Most aphasics have extensive brain lesions
3. Language brain regions much more extensive than Broca and Wernicke
alone
4. Differences between languages
5. Aphasics may also have more general problems than just with language
=> maybe more important to look at which language level is affected instead
of distinguishing between comprehension/production

Anomia = inability to name objects

Semantic, phonological or both

BUT semantic processing not always complete when phonological processing

begins > interaction

Agrammatism = no word-finding problem but word order problem

Short sentences, only content words, no function words, no inflectional

morphology

Problems with processing grammatically complex sentences

Problem with verb tenses

Situational context (more problems outside than at home)

General problems with learning sequences

Relationship with syntactic level of speech processing

Possible domain-general problem of sequencing

Great diversity of symptoms

Aphasics can sometimes do more than expected

More research into relationship domain-specific and domain-general

Jargon aphasia = intact syntactic processing but impaired word finding

(inverse of agrammatism)

=> neologisms
Causes:

- Problem with selection of phonemes

- Problem with cognitive control
- Problem with monitoring own speech

Lecture 7 - Writing

Writing of extended texts > psycholinguistics

Spelling > neurolinguistics

Processes of writing in Hayes’ writing model:

• Planning (retrieval and structuring of ideas)
• Sentence generation | translation (from ideas to language)
• Transcription
• Revision (reread and edit)

Writing is similar to speaking – productive processes

But different in time!
In speaking second phase of sentence generation is called formulation, third
phase – articulation, revision – speech monitoring

Time spent on each of Hayes’ writing processes

=> directed retrospection

Constant switching btw processes (6-8 switches/minute)

Monitor making switches possible (central executive working memory) >

related to cognitive control/executive functions

- Time on planning decreases

- Time on revision increases
- All 3 processes engaged during each phase of writing
- Revision most cognitively demanding, then planning and translation

Strengths of model of Hayes:

• 3-4 processes crucial during writing

• Not strictly separated because of many switches
• Monitor present to enable smooth switching
Limitations:

• Unclear which factors determine when writers shift processes

• Social aspect of writing de-emphasized
• No precise specification of interaction between writing processes

Predictor of writing skills in children – reading -> learn about structure and
style of good writing, enhances vocabulary and knowledge
Better writing:

- More time on planning and revision

- Feedback from reader = knowledge effect
- Use meta-cognitive strategy (‘would readers understand this’)
- Adaptation to type of text and profile of readers
Working memory

= active part of memory (limited capacity), temporary storage of information

while other information is being processed (s4 processes simultaneously)

3 components:
Central executive system (CES) => coordination of all cognitive activities in
the brain

Visuo-spatial sketchpad => visual and spatial processing

Phonological loop => memorizing linguistic (esp. auditory) information
Central executive
Individuals with more effective CES functions -> better writers

The quality and complexity of writing are determined by working memory

capacity
Dysexecutive agraphia -> impairment of working memory/central executive
> less effective writing in terms of planning and organizing
Phonological loop

Tested through ‘articulatory suppression’ (psycholinguistics) -> shorter

sentences and slower transcription, inability to produce words with multiple
syllables in correct order
But patients with impairment to phonological loop can still write
(neurolinguistics)

Visuospatial sketchpad
Plays a role, esp. during planning and revision, but possibly also during writing
concrete words
Writing impairments based on the brain condition – dysgraphias
Diagnosed by neurolinguist by means of writing test
Word categories used: regularly spelled, irregularly spelled, non-existing
words

(a) model for understanding dyslexia

(b) model for understanding dysgraphia
*2 routes are not strictly separated, and lexical (direct) route impacts the non-
lexical route
-> dual-route model is an oversimplification of the writing process

People with Alzheimer use both routes e.g ‘bouket’ for bouquet

Surface dysgraphia

=> lexical route impaired

Affected spelling of irregularly spelled words

Phonological dysgraphia
=> non-lexical route (sublexical) / phonological route is impaired

Affected spelling of non-words

Dyslexia and dysgraphia often co-occur but some patients have greater
difficulties with reading than writing or vice versa (link with motor skills in
writing, link with visual skills in reading)

• Sometimes dissociation between both skills

• There is a significant overlap in involved brain regions
• There is likely one orthographic lexicon for reading and writing, which
sometimes shows dissociation because access from spoken or written
language is disrupted (links with motor or visual cortex)

Brain parts involved in writing:

• Intraparietal sulcus and superior parietal lobule in the left hemisphere

=> selection and/or representation of letter shapes (graphemes)
• Superior frontal sulcus in the left hemisphere => interface between
abstract letter combinations and the generation of motor commands
Frontal brain part – motor commands connected with

• Posterior cerebellum right hemisphere => motor activity

Lecture 8 Morphosyntax

Morphosyntaxis – set of rules determining the relationship between

(sentence) structure and function

Examples of structures:
• Clause level (main/subordinate clause)
• Phrase level
▫ Noun phrase (NP)
▫ Verb phrase (VP)
▫ Adjective phrase (Adj P)
▫ Adverb phrase (Adv P)
▫ Prepositional phrase (PP)

Examples of functions:
• Subject
• Object
• Verb
• Complement
Morphosyntax: structure-function relations
Lexicosemantics: structure-meaning relations

Phonology: structure-meaning differentiation relations

Same processes for reading and listening

Important question: relation between (morpho)syntactic and semantic

analysis
Parsing in the brain
The importance of syntactic analysis (parsing) , and the relationship between
morpho/lexical and morphosyntactic analysis, is highly language-dependent.
Synthetic vs. Analytical languages

Synthetic languages => structure-function relationships are determined by

structure of the word and combination of morphemes (inflectional
morphology), sentence order plays a limited role

Analytical languages => structure-function relationships determined by

sentence order, and inflectional morphology plays a limited role.

Parsing occurs rapidly => investigated by ambiguous sentences

Prosody (stress, intonation, rhythm, word duration, pauses) => help resolve
ambiguity on syntactic level

Advantages:
- Immediate integration into parsing
- Pauses at phrase boundaries have more influence than other prosodic
cues
- Primarily during speech comprehension, but also during reading (‘inner
voice’) pauses > commas
- Evidence of EEG/ERPs > implicit prosody hypothesis

2 implications of the implicit prosody hypothesis:

1. Presence of commas can help improve understanding of ambiguous

sentences
2. Prosody training can enhance the reading skills and comprehension of
children
3 sidenotes about role of prosody in parsing in the brain:
 1. Prosodic analysis does not occur at just one particular point in the
sentence but at multiple points. It is the overall pattern of prosodic
phrasing that influences comprehension
2. There is significant variability among individuals regarding production
of prosody
3. The implicit prosody hypothesis plays a larger role in reading aloud
(when the sentence was not skimmed in advance) than in silent reading

Theories about parsing

Garden-path model

“to be led up the garden path” (misled)

e.g. The horse raced past the barn fell

1. One initial syntactic structure is selected for sentence parsing

2. Semantics initially do not play a role > syntax – first, serial model
3. Minimal attachment: the simplest syntactic structure is selected
(preference for main clauses, SVO structure, and attachment at the top
level)
4. Late closure: if possible, new words are attached to the previous phrase
or constituent
5. Minimal attachment is a more significant principle than late closure
6. If the initial syntactic structure conflicts with additional information
(e.g semantics), a revision (reanalysis) takes place

− This model has primarily been tested through the study of eye
movements during sentence reading (eye-tracking)
− Also been tested through the neurolinguistic approach

E.g. Minimal attachment => ‘The girl knew the answer by heart’
NOT ‘The girl knew the answer was wrong’
E.g. Late closure => ‘Since Jay always jogs a mile it seems like a short distance
to him’
NOT ‘Since jay always jogs a mile seems like a short distance to him’

In patients with semantic dementia (where word meanings are no longer

available), a dissociation was observed between semantic and syntactic
analysis. Patients could still accurately judge the grammaticality of sentences,
but they could no longer retrieve word meanings

Limitations of the Garden-path model:

1. It is not accurate that readers/listeners select an initial syntactic
structure and stick to it until revision is necessary. Other sources of
information (semantics, etc.) can also be utilized in this initial phase to
perform sentence parsing. This model lacks flexibility compared to
constraint-based model.
2. Other sources of information that are demonstrably utilized include
prosodic information and visual context, and this information can
override the principles of minimal attachment and late closure
3. Revision / reanalysis of sentences that were incorrectly parsed does not
always occur, whereas this model sees it as a necessary processing step
4. The model is difficult to falsify. If proponents of the model are pointed
to the significant role that is played by prosody and visual context, they
argue that these factors are already part of the second processing phase,
which begins shortly after the initial phase
5. This model is particularly applicable to English, where word order plays
a crucial role in sentence parsing. In synthetic languages with flexible
word order, certain cues (such as inflectional morphemes) quickly
reveal the sentence structure

Constraint-based model
 1. Sentence parsing relies on various sources of information (syntactic,
semantic, world knowledge), imposing constraints on the possible
interpretations of a sentence. >< garden-path
2. Different syntactic constructions are simultaneously activated (in
parallel) >< garden-path
3. The syntactic structure most frequently associated with a particular
verb is more strongly activated (verb bias). ><garden-path
4. Syntactic structures that were recently activated are more likely to be
reactivated (parsing flexibility). >< garden-path
E.g. Different syntactic constructions simultaneously:

Because it was John that Ralph threatened the neighbour recorded their
conversation
> subject analysis of ‘neighbour’

Because it was John that Ralph threatened the neighbour recorded their
conversation

> object analysis of ‘neighbour’

>< garden-path: only subject analysis is possible, constraint-based: 2 analyses
are in conflict

E.g. Verb bias:

Verb ‘read’ can occur in combination with direct object or with subordinate
clause. DO is more frequent > this construction will be selected

• The professor read the newspaper had been destroyed.

• The professor read the newspaper during the break (DO)!

>< Garden-path: because of principle of minimal attachment

E.g. Recency assumption:

1) The experienced soldiers warned about the dangers before the midnight
raid (SVO)
=> Garden-path: Minimal attachment
 => Constraint-based: Verb bias

> Same analysis: (1) easier than (2) -> (1) Main verb; (2) verb in (relative)
subordinate clause

2) The experienced soldiers warned about the dangers conducted the

midnight raid.

But if recently exposed to (2), then switch to (2) rather than (1) > flexibility of
parsing >< garden-path.

Strengths of Constraint-based model:

• Simultaneity of information deployment
• Flexibility, syntactic expectations adapt to frequency and recency, and
are therefore not based on universally applicable basic principles

Limitations:

• Lack of precision in predictions

• Influence of non-syntactic factors on parsing within the garden-path
can also be explained by assuming that the second phase quickly follows
the first phase

Lecture 9

Unrestricted race model:

This is a blend of the 2 previous models

Like garden-path model – posits that we initially select a sentence structure
that is only abandoned when additions info becomes available > leading to
revision of the sentence

This model suggests that this selection is no based strictly on syntactic grounds
(><garden-path model) but on basis of all available info (similar w constraint-
based model)

Strength

• Best capable of explaining ambiguity advantage

Limitation
• Readers do not always select correct syntactic structure in the end.
• Task requirement can impact the ambiguity advantage: if the task
requires all syntactic constructions to be active, an ambiguity
disadvantage is more likely than an ambiguity advantage

“The second wife will claim the entire family inheritance for herself.”
3 models make diff predictions about the reading speed of this sentence:

1.garden-path? ‘bold’ (the entire family inheritance) as quickly as underlined

(for herself ) because correct interpretation in line w minimal attachment.
2.constraint-based? ‘Bold’ slower than ‘underlined’ bc conflict between both
interpretations only resolved when ‘underlined’ is parsed
3. unrestricted-race? ‘bold’ quicker than ‘underlined’ because sometimes
selection of incorrect construction which only becomes revealed at
‘underlined’

Good-enough representations
All previous models assume that processes of sentence parsing always occur in
sentence comprehension. But there are conceivable reasons why this is not
the case:
 1. There are differences among individuals regarding
grammatical knowledge. These differences are
associated with IQ, education, linguistic prof.
2. Parsing requires a lot of cognitive effort
Strengths:

1. Language processing sometimes occurs superficially

2. More thorough processing in case of tough comprehension

questions or in individuals with high working memory capacity

Limitations:
1. Unclear rules are involved in the heuristic route
2. heuristic route can comprise limited processing as well as top-down
processing (considered more difficult)

3. reanalysis (algorithmic route) does not always yield better results (cf.
unrestricted race model)
(4) misinterpretation can occur after correct syntactic analysis (cf. because of
poor integration syntax and semantics, the mouse was eaten by the cheese)

Hypothesis of 2 routes for parsing:

1.heuristic: superficial and good-enough

2.algoritmic: more profound and accurate

Attention = selective allocation of cognitive resources

3 states (Tomlin & Villa)

Orientation

Alertness. => controlling input taken in for processing

Detection
Input=> attention => intake => internal system => output

What guides attention and awareness?

 Learning intentions/goals
 Learning context
o Task demands
o Teacher instruction
Salience: the extent to which a feature stands out from its environment

Noticing Hypothesis
Attention & awareness necessary to language acquisition
Noticing: state lying between attention & awareness

• (Conscious?) registration of a feature within L2 input that it becomes

intake
• Potential to be processed by internal system => acquired

Generative Theories of SLA

Universal Grammar

SLA reflects intuitions from unconsciously understood grammatical

constrains

Minimalist Program & Bayesian models

• Noticing of formal differences basis for constructing unique
representation (Ladier; Bybee)
o Noticing of tokens => Exemplars => mental representations
▪ Important role for attention/awareness in unconscious
generalisation processes
Input Hypothesis
Distinguishes learning from acquisition

• Acquisition: unconscious, represents native-like L2 competence

• Learning: conscious, of little consequence to real-life language use
o Attention & awareness only important for metalinguistic
knowledge
Output hypothesis

Awareness of not only input but also what is missing from output
“noticing the gap”

Explicit knowledge/learning = conscious

Implicit = unconscious