Language and Communication

1
For more information on Alan Garnhams part of
the Language and Communication course see the
following web page, and the links from
it http//www.biols.susx.ac.uk/Home/Alan_Garnham
/Teaching/Lang_and_Comm
2
Language and Communication - 2002

• Word Identification - Introduction and Written
  Word Identification

3
General Issue

• The listener or reader hears or sees a perceptual
  pattern that is meaningless in itself (think of
  hearing or seeing a word in a language you dont
  know)
• The pattern conveys meaning because in learning a
  language a person learns to associate sounds (and
  visual patterns in reading) with meanings (by
  creating a store of knowledge about the words of
  the language - the mental lexicon)
• How does a particular occurrence of a perceptual
  pattern (e.g this instance of the word pattern)
  get associated with the right meaning?
• Once the words have been identified, more complex
  meanings can be built up from them (see later
  lectures)

4
FROM A GEORGIAN BOOK OF PRAYERS
5
Spoken Word Identification (Listening)

• Actually rather complicated because
• Word boundaries may be clearly perceived but are
  not usually marked by breaks in the speech stream
• Although we dont usually notice it, sounds
  change with context
• E.g. what sounds do you actually make when you
  say green man
• Something like greeman

6
Written Word Identification (Reading)

• In principle easier, because
• In print
• Word boundaries are clear (spaces/new lines)
• Letter forms are clear
• And even in writing
• Word boundaries are usually clear
• Letter forms may or may not be clear

7
Studying Word Identification

• Generally people ask what makes word
  identification easy or difficult?
• Time spent identifying a word can be a measure of
  difficulty
• Measures of identification time are usually
  indirect

8
Some Identification Time Measures

• Measure how long people actually spend looking at
  a word when READING (using eye movement
  monitoring techniques)
• Measure how long people take to start saying a
  word (naming or pronunciation time)
• Measure how long people take to say a string of
  letters is (or is not) a word (lexical decision)
• Measure how long people take to categorise a word
  (apple is a fruit)

9
Some Other Measures

• Measure how accurately people identify a briefly
  presented word (tachistoscopic recognition)
• Measure how much of a (spoken) word people need
  to hear to recognise it (gating)

10
Factors Affecting Identification

• Perceptual Clarity
• Length
• Frequency (how common the word is)
• Familiarity? Age of Acquisition?
• Priming (by prior, related, material)
• Repetition
• Form-based
• Associative
• Semantic (words, sentences, discourse)
• Attentional (Neely)

11
Factors Affecting Identification

• Neighbourhood Effects
• Are there other similar words in the language?
• If there are, identification can be speeded,
  especially for uncommon words
• Regularity and Consistency
• Imageability/Concreteness
• Part of Speech (Noun vs. Verb)
• Morphological Complexity

12
Models of Written Word Identification

• Serial (search) models
• Forsters autonomous serial search model
• Parallel (detector) models
• Mortons Logogen model - various versions
• Connectionist Models
• Interactive Activation
• Mixed Models
• Beckers Verification Model
• Norriss Checking Model

13
Serial versus Parallel Models

• Serial search model - analogy of looking through
  a dictionary
• Parallel/direct access model - perceptual
  properties of a word can directly access a
  lexical item (by activating its detector), and
  multiple lexical entries are activated
  simultaneously, i.e. in parallel.

14
Accessing a normal dictionary

• speculum(sp?kj?l?m) n., pl. -la (-l?) or -lums.
  1. a mirror, esp. one made of polished
  metal......
• sped(sp?d) vb. a past tense and past participle
  of speed.
• speech(spit?) n. 1. the act or faculty of
  speaking.....
• speechless(spit?lIs) adj. 1. not able to speak.
  2. temporarily deprived of speech. 3. not
  expressed.......
• speed(spid) n. 1. The act or quality of acting
  or moving fast rapidity. 2. the rate......

15
Forsters (1976) Autonomous Search Model
16
Forsters (1976) Autonomous Search Model

• Library analogy, a word (book) is found in only
  one place in the mental lexicon (library) but can
  be located by using various resources
• Access files Orthographic - words are accessed
  via visual features (for visual word
  identification)
• Phonological - sound (for spoken word
  identification)
• Syntactic/semantic - meaning and grammatical
  class (for production)
• Master lexicon where all linguistic information
  about words are stored pronunciation, spelling,
  grammatical class, meaning, etc.

17
Forsters (1976) autonomous search model

• Frequency Effects can be explained by model as
  the access files are organised into binsin which
  the most frequent words are stored at the top,
  and hence searched first.
• Non-Word Effects in Lexical Decision Tasks -
  strings such as xzpqr are known to be rejected
  very quickly (no search?) but model predicts
  non-words should take longer to reject than it a
  takes to accept a word, since a whole bin (at
  least) must be searched.
• Semantic priming - Not fully explained by
  Forsters model, though he suggests that the
  master lexicon has connections between
  semantically associated words which can be used
  to create a supplementary access file.
• Neighbourhood effects - Serial search might
  appear to predict that words with many neighbours
  will take longer to access than those with few
  neighbours because more entries must be
  examined. However, the exact predictions would
  depend on details of organisation into bins,
  etc..

18
Adapted from Morton, 1979, facilitation on word
recognition experiments causing change in the
logogen model, in P Kolers, et al. (eds)
Processing of visible language vol. 1.

• Original logogen model

Visual analysis
Auditory analysis
Cognitive system
Logogen system
Response buffer
19
Mortons (1969) Logogen model

• Each word has its own Logogen an evidence
  detector or scoreboard for a word
• Heated light bulb analogy once word activated
  the bulb is lit and warm, when word is no longer
  activated the bulb does not cool straight away,
  so if the word is present again, it will take
  less time to be activated.

love
lovely
Individual threshold
xxxxxxxxx
xxxxxxxxx
xxxxxxxxx
xxxx
xxxxxxxxxx
xxxxxxxxxx
xxxxxxxxx
l
o
v
e
l
y
20

• Frequency effects explained as increased
  experience of a word resulting in a higher
  resting activation for high frequency relative to
  low frequency words.

love
loathe
Individual threshold
Semantic priming Because of the bi-directional
flow of information between the cognitive system
and logogen system, activation from one logogen
spreads (indirectly) to those for related words.
Because activated logogens do not return to their
resting level immediately, the primed target will
require less perceptual input to be activated to
its individual threshold, and hence less time
than an unrelated target. Neighbourhood effects
original model proposed before such findings were
discovered, therefore not developed to account
for this finding. Logogens gather evidence
independently of one another
21
Morton, 1979, was not able to find cross modal
priming, hence he changed the model so that
information from across modalities entered
separate logogen systems.
The lexicon
Visual logogens
Auditory logogens
Output logogens
Auditory analysis
Visual analysis
Response buffer
Grapheme Phoneme conversion
However

• Adapted from Morton, 1979, Facilitation in word
  recognition experiments causing change in the
  logogen model, in P Kolers, et al. (eds)
  Processing of visible language vol. 1.

22

• Cross modal priming is now a well-established
  phenomenon.
• E.g. Zwitserlood, 1989.

c a p t i ve
auditory prime
c a p t ai n
or
slave
visual probe
ship
shop

• Priming found to both alternatives in early
  condition only
• More priming found to ship a frequency effect

23
Mixed Models

• Beckers (1976) verification model - a small
  number of candidates, activated in parallel, are
  subject to a (serial) verification process
• Norriss (1986) checking model - partial matches
  with the input as analysed so far are checked to
  see if they fit with context

24
Connectionist models

• McClelland and Rumelhart, 1981, 1982, Interactive
  activation model.

Nodes (visual) feature, (positional) letter, and
word detectors which have Inhibitory and
excitatory connections between them.
25
Interactive Activation Model

• Inhibitory connections within levels
• If the first letter of a word is a, it isnt
  b or c or
• Inhibitory and excitatory connections between
  levels (bottom-up and top-down)
• If the first letter is a the word could be
  apple or ant or ., but not book or
  church or
• If there is growing evidence that the word is
  apple that evidence confirms that the first
  letter is a, and not b..

26
Interactive Activation Model

• The model was originally developed to explain the
  Reicher-Wheeler (word superiority) effect - it is
  easier to recognise a letter in a word than in a
  non word or a string of xs.
• Top-down (word?letter) excitatory connections
  assist letter identification in word, but not in
  nonwords
• It is a localist model - features, letters and
  words are represented by individual nodes - their
  representations are not distributed across the
  network

27

• Frequency effects - not originally accounted for,
  but high frequency words could have stronger
  connections to lower level nodes than low
  frequency words. Therefore L would have a
  stronger connection with the word LOVE compared
  to LOATHE. Or they could have higher resting
  levels of activation (as in the logogen model)
• Priming and context effects - again, not
  originally accounted for
• Neighbourhood effects - as neighbours are words
  represented by similar patterns in the feature
  and letter layers, these will all be activated.
  But one word has to win - the activation of many
  similar words could facilitate recognition of the
  correct word but if a high frequency word is
  competing with a lower frequency word, its
  stronger activation may interfere with the
  process of accessing the lower frequency word.

28
Conclusion

• Both serial and parallel models have pros and
  cons, but there is a general consensus towards
  accepting the parallel processing models of
  lexical access.