Language Flashcards
Language in the brain
Commonly activated
regions include superior, middle and
inferior temporal gyri in both
hemispheres and the left inferior
frontal gyrus (Broca’s area). White
matter tracts, especially arcuate
fasciculus and the extreme capsule
also play a major role.
Language building blocks
- Words: The representations of words in the
mental lexicon (ML) contain information about
their spelling, pronunciation, meaning,
grammatical category etc. Some of these
features are explained in more detail below (this
list is not intended to be exhaustive - several
others have also been discussed in the
literature): - Phonemes: Phonemes are the smallest units of
speech sound that allow discrimination between words in a given language. Speech errors
like spoonerisms show us that words are not monolithic blocks of information, and that
phonemes are cognitively real. Phonological overlap between words has also been shown to
affect their identification, further suggesting that phonemes contribute to the structure of the
ML - Morphemes: Morphemes are the smallest units in the language that have meaning (e.g.
dog,–s,-ness). They combine to form more complex words (dog+s, brave+ly, dark+ness).
Errors such as ‘a geek for books’ instead of ‘a book for geeks’ point to the presence of
morphemes, since the phoneme –s and its position alter the meaning of this phrase (this
error is known as ‘stranding’). Morphological information also contributes to the structure of
the mental lexicon. - Syllables and stress: Information about syllables and stress is also likely to be represented
in the mental lexicon. At least two types of evidence support this conclusion: a) expletives
can only be inserted into words with an appropriate syllabic and stress pattern (McCarthy,
1982). b) stress can alter the meaning of a word, and some brain-damaged patients show
selective stress errors (e.g., Cappa et al, 1997 documented the case of CV, an Italian patient
who produced speech errors in which the stress fell on the wrong syllable; however, the
phonemes were properly selected).
The components of
written word recognition are
extracting information from text; letter identification; access to the orthographic lexicon; grapheme to phoneme conversion; retrieval of word meanings.
Extracting information from text
While we read, our eyes make a series of movements (saccades), which are separated by
fixations. The role of fixations is to bring text in the foveal vision. The average fixation is 200-
250ms; the average saccade is 8 letters. 10-15% of the time readers move their eyes
backwards. The span of effective vision is about 14-15 letters to the right of fixation, and 3-4
letters to the left.
Letter identification and the visual word form area (VWFA)
Letters are identified following the initial analysis in the visual cortex. Letter recognition
involves two stages: letters are first recognised based on their physical characteristics and
then identified irrespective of the shape (A=a).
Alexia results from lesions in the
left posterior temporo-occipital regions, particularly the so-called visual word form area
(VWFA) in the left fusiform gyrus.
Alexia
Alexia is a selective impairment in identifying written words and
letters.Alexic patients have problems in determining
letter identities but they are still able to make
discriminations based on the physical characteristics
of letters (e.g., whether the letters are normally
oriented or real). Their oral spelling is intact,
suggesting that they retain information about orthography. Alexia results from lesions in the
left posterior temporo-occipital regions, particularly the so-called visual word form area
(VWFA) in the left fusiform gyrus.
Orthographic lexicon
The orthographic lexicon stores representations of spelling for familiar words. Access to
information stored at this level is especially needed for words with irregular spelling (yacht,
colonel, aisle), these cannot be named reliably unless readers have previously learnt their
pronunciations and orthographies.
Grapheme to phoneme conversion
In addition to reading previously known words, we can also read out loud newly encountered
letter strings (novel words or pseudowords). They cannot be retrieved from the orthographic
lexicon, suggesting that we have a mechanism to transcode letters into sounds.
Dual-route cascaded model(DRC, Coltheart et al 2001)
DRC postulates two reading routes,
lexical (i.e., lookup route) and non-lexical (grapheme to phoneme conversion route). Lexical
route is faster for reading words, and is also faster
for reading regular than reading irregular words.
Non-lexical route is faster for reading pseudowords.
When the lexical route is damaged, the processing
must go through the grapheme to phoneme
conversion, which is inadequate for words with
irregular spelling. This is the pattern seen in
patients with surface dyslexia.
Triangle model (Seidenberg & McClelland; 1989; Harm &
Seidenberg, 2004).
This model postulates three components: semantic units that encode the
meaning of words; phonological units that specify word sounds (phonemes); orthographic
units that represent word orthography (letters). This is a completely interactive model: the
components are interconnected and contribute jointly to the recognition of words and
pseudowords. Words are represented as patterns of activated semantic, phonological, and
orthographic units. Despite this massive interactivity,
differences would naturally emerge among regular
words, irregular words and pseudowords since
different activation patterns correspond to each of
these classes. Phonological units and their connections
to orthographic units are particularly relevant for
pseudowords (pseudowords have no meanings, so the
contribution from semantic units is necessarily
limited). On the other hand, the orthographic and
semantic units and the connections between them are
critical in the recognition of words with irregular spelling.
Visual word processing in the brain
Reading activates regions in the occipital, temporal and frontal brain areas, dominantly on the
left. Information is passed
from the occipital regions onto the VWFA,
where the letter stings are identified. It is
then distributed to numerous brain regions
to encode word meaning, sound and
articulation. These more anterior regions
are ‘amodal’, as they support processing of
both written and spoken language
The major issues addressed in the spoken word
recognition literature are therefore
the mechanisms for segmentation of the spoken stream,
lexical selection, access to meaning, and the effects of context.
Word segmentation (metrical segmentation strategy (MSS))
The first problem that listeners encounter is the identification of
word boundaries (segmentation problem). This is visible on
spectrograms, where there are no pauses corresponding to
word boundaries. This problem is also apparent while listening
to foreign speech. Cutler and Norris (1988) proposed that
English speakers use a metrical segmentation strategy (MSS) to
segment speech. This is based on syllabic stress, reflecting the
fact that the rhythmic structure of English is stress-timed (i.e.,
some syllables are more stressed than others, e.g. the first
syllable car in carpet is more stressed that the second).
Stressed syllables are perceived as having greater loudness,
increased duration, and change in pitch. They contain full
vowels, while unstressed syllables have reduced vowels (usually a schwa, as in the initial syllables
of behind and averse). Many grammatical words (pronouns, articles, prepositions) are typically
unstressed in continuous speech (e.g., him, it, of). The MSS theory proposes that stressed
syllables are taken as likely onsets of content words (nouns, adjectives, verbs, adverbs) and that
continuous speech is segmented at stressed syllables.
Lexical selection
The goal
of this rapid searching process is to identify the stored representations that match the input
information. The lexicon contains word representations that encode various features of word
sounds.
For English words, these include phonemes and stress, in other languages, other features
may be encoded (e.g., tones in Chinese).
Words are represented in the lexicon in an abstract
format that does not encode low-level acoustic characteristics (e.g., differences between male
and female voices). Lexical search starts as soon as word onset has been identified. One of the
clearest demonstrations comes from Marslen-Wilson’s (1975) shadowing paradigm study.
Participants listened to spoken passages and repeated back what they were hearing. Words in the
input were occasionally incorrect but participants corrected the incorrect words, and this often
occurred before the incorrect word was presented in full. Further studies estimated that words in
context can be recognised within 175-200 ms of their onset, or when only half or less than half of
their acoustic content has been presented.
Lexical selection- cohort model of spoken word recognition
Many words can be uniquely identified even
when they are incomplete. For example,
alligator is the only English word
corresponding to the sequence allig, so we
do not need to listen to the whole word in
order to recognise it. The point at which
words can be reliably recognised is called
uniqueness point
Lexical selection-Allopenna et al. (1998) provided evidence that information coming at later points can also
activate lexical entries.
The participants saw a display containing several objects and were
instructed to grasp one of them (e.g., beaker). The distractor
objects had names that shared either the initial or the final
part of the target word (e.g., beetle, speaker). Eye
movements allowed detecting which words were fixated at
each point in time. The results showed that onset-related
words competed early on, whereas end-related words
competed only at a later point in time. This shows that the
word onset is not the only portion that triggers lexical
activation. Information coming at later points can also
activate lexical entries, such that the word speaker can
interfere with the recognition of the word beaker. This is
desirable since (a) word boundaries are not reliably detected;(b) noise may prevent listeners from hearing word onsets.
Access to meaning
Swinney et al (1979) provided data for understanding the access to meaning in speech comprehension.
Participants listened to a story containing words with ambiguous meanings (e.g., bug = insect vs. spying device). At the same time, a letter string (ant, spy or sew) appeared on the screen and they were asked to decide whether these strings corresponded to existing English words.
Early on, responses were faster for both ant and spy relative to sew, suggesting activation
for both of the meanings of the word bug. However, activation of the non-compatible meaning
(spy) faded after 200 ms – after that time, responses were faster only for words with the
compatible meaning (i.e, when the paragraph was about insects, priming only appeared with
ant).
These results suggest that: (a) different meanings are initially activated and contextual
information is not used to determine which words are considered for recognition, but that (b)
contextual information is critical for the selection of the appropriate word meaning from the
multiple activated alternatives.
Context effects
Data show that, when presented in
noise, word recognition rate is higher when words are presented in sentence context than in
isolation. (Warren, 2008).
So, context plays an important role in speech recognition. The demonstration gives an example
of Warren’s phoneme restoration effect (Warren, 2008).
The advantage for contextual presentation appears even if words are presented in a
good acoustic environment: when spoken words were sliced out from recorded conversations and
presented alone, word recognition dropped to about 50%. However, the inclusion of one or two
neighbour words from the original conversation was sufficient to increase word recognition
dramatically (Pollack & Pickett, 1964). Evidence from a gating study by Tyler (1984) further
shows that context does not affect the set of initially activated candidates, but rather the process
of selection and narrowing down the initial set of candidates activated by the sensory input
Speech processing in the brain
Marinkovic et al (2003) used magnetoencephalography (MEG) to investigate the precise timing of
spoken word recognition in the brain. Around 50 ms after a word is heard, the activation appears
in the early auditory regions in left and right temporal lobes. This is where the acoustic properties
of words are processed. The activation then spreads out
to middle and inferior temporal regions, as well as to
inferior frontal regions. These are the amodal language
regions, hypothesized to process the meaning and the
grammatical properties of words.
Marinkovic et al (2003) used magnetoencephalography (MEG) to investigate the precise timing of
spoken word recognition in the brain. Around 50 ms after a word is heard, the activation appears
in the early auditory regions in left and right temporal lobes. This is where the acoustic properties
of words are processed. The activation then spreads out
to middle and inferior temporal regions, as well as to
inferior frontal regions. These are the amodal language
regions, hypothesized to process the meaning and the
grammatical properties of words.
Sentence processing-syntactic rules
The rules that govern how words in the language can be combined a
The major theories of sentence parsing
Garden-path model (Frazier & Rayner, 1982) argues that the initial parsing is purely syntactic,
with meaning not involved in the selection of the initial syntactic structure.
Constraint-based
theories (MacDonald et al, 1994) suggest that the initial interpretation depends on all available
sources of information (syntactic, semantic, general world knowledge).
The unrestricted racemodel (Van Gompel et al, 2000) argues that all sources of information are used to identify the
initial syntactic structure,
while the theory of ‘Good-enough’ representations (Ferreira et al,
2002) argues that processing depth / type of information depends on the task.
Sentence parsing cues
Experimental evidence shows that people use a variety of information in order to understand
and interpret sentences: structural syntactic principles, statistical regularities, grammatical
categories, prosodic cues etc.
Syntactic principles:
Frazer and collaborators (1987) suggested that listeners initially primarily
rely on principles such as late closure and minimal attachment to organise words into preferred
sentence structures.
For instance, experiments using reading times and eye-tracking have
shown that English-speaking participants prefer the interpretation of the girl holding the book in
the sentence John hit the girl with the book.
Syntactic principles-the late closure principle
For instance, experiments using reading times and eye-tracking have
shown that English-speaking participants prefer the interpretation of the girl holding the book in
the sentence John hit the girl with the book.
The late closure principle states that this is because
new items are attached to the phrase or clause that was most recently processed, if
grammatically possible. This allows the listeners to overcome the limitations of the cognitive
system, which is under pressure to extract the sentence meanings quickly and reliably. The
cognitive system can hold a limited number of words in short-term memory. As words arrive,
memory traces of previous words are lost. Late closure and minimal attachment make it
possible to accommodate incoming words within partially formed syntactic structures and
minimize the loss of information.
