Language and Communication 2002

1
Language and Communication - 2002

• Word Identification Spoken Word Recognition and
  Lexical Ambiguity

2
Spoken Word Identification The Segmentation
Problem

• How do we divide the speech stream into a series
  of discrete words
• Each part of the speech stream should be part of
  one and only one word (the possible word
  constraint)
• (In English) the main content-bearing words tend
  to begin with stressed syllables (Cutler, Norris
  and colleagues)

3
Evidence for Segmentation in Progress Activating
spurious words

• Shillcock (1990)
• He picked up the trombone

rib

• Lexical context does not constrain spurious access

4
Some Properties of Spoken Word Identification

• Shadowing and word-monitoring tasks
• latencies of 250-275 msec.
• subtract 50-75 msec for response execution
• 200 msec to identify word
• before acoustic offset
• Context apparently aids recognition

5
Models of Spoken Word Identification

• The TRACE (Interactive Activation) Model
• McClelland Elman, 1986
• The Cohort Model
• Marslen-Wilson Welsh, 1978
• Revised, Marslen-Wilson, 1989

6
TRACE

• Like the interactive-activation model of printed
  word recognition, TRACE has three sets of
  interconnected detectors
• Feature detectors
• Phoneme detectors
• Word detectors
• These detectors span different stretches of the
  input (feature detector span small parts, word
  detectors span larger parts)
• The input is divided into time slices which are
  processed sequentially.

7
TRACE - continued

• Within a set (or level) connections are
  inhibitory
• E.g. evidence that a certain stretch of the input
  is the word tip is evidence that it is NOT any
  other word
• Between a set (or level) connections are
  excitatory
• E.g. evidence that a certain stretch of the input
  is the sound /t/ is evidence that it might be the
  beginning of the word tip
• Also, evidence that the word is tip is evidence
  that its parts are /t/ /i/ /p/, so there are
  top-down (feedback) effects in TRACE as in the
  interactive activation model
• Or inhibitory..
• If its a /t/ it isnt the beginning of cat

8
TRACE - continued

• Accounts for context effects
• Can handle (some) acoustic variability (and
  noise)
• Can account for phoneme restoration (Warren -
  Open the ?oor heard as Open the door)
• Can account for co-articulation effects
• Can find word boundaries (using the possible word
  constraint)

9
Marslen-Wilsons Cohort Model

• The mental representations of words are activated
  (in parallel) on the basis of bottom-up input
  (sounds), and can be de-activated by subsequent
  bottom-up (phonological) and top-down
  (contextual) input.

10
Uniqueness and Recognition

• When we hear the beginning of a word this
  activates ALL words beginning with the same
  sound the word initial cohort. Subsequent
  sounds eliminate candidates from the cohort until
  only one remains (failure to fit with context can
  also eliminate candidates)
• t - tea, tree, trick, tread, tressle,
• trespass, top, tick, etc.
• tr - tree, trick, tread, tressle, trespass,
• etc.
• tre - tread, tressle, trespass, etc.
• tres - tressle, trespass, etc.
• tresp - trespass.

11
Uniqueness and Recognition

• The uniqueness point is the point at which a word
  becomes uniquely identifiable from its initial
  sound sequence
• E.g. dial dayl crocodile krokod ayl
• UP UP
• For non-words there is a deviation point a point
  at which the cohort is reduced to zero
• E.g. zn owble would be rejected with a
  faster RT than thousaj ining
• DP DP

12
Uniqueness and Recognition

• The recognition point is the point at which,
  empirically, a word is actually identified
• Empirical studies show that recognition point
  correlates with (and is closely tied to) the
  uniqueness point.
• phoneme monitoring latencies correlate with a
  priori cohort analysis (and one way to recognise
  word initial phonemes is to recognise the word
  and to know it begins with e.g. /p/)

13
Effects of Material beyond the UP / DP

• Auditory lexical decision task, pairs of
  non-words compared with the same Deviation Point,
  but one resembled a real word beyond (and before)
  the DP.
• e.g. rith l ik rith l an
• UPDP UPDP
• The cohort model predicts same RT for both but
  first word (472ms) was slower than the second
  (372ms), and error rate was 3.5 for the first
  and 0.6 for the second.
• Conclude that the cohort model fails to account
  for this phenomenon.

14
Frequency Effects in Spoken Word Identification

• Marslen-Wilson auditory lexical decision task
  with pairs of words with the same length, UP, and
  different frequencies.
• e.g. DIFFIC ULT high frequency (250ms)
• DIFFID ENT low frequency (379ms)
• Not immediately clear how the original version of
  the Cohort Model accounts for this effect

15
The Zwitserlood experiment

• cross-modal priming

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

16
Zwitserlood - Conclusion

• Zwitserloods experiment showed that frequency
  of a word affects the activation level of
  candidates in the early stages of lexical access,
  hence there are relative frequency effects
  within the initial cohort, so that entry in the
  cohort cannot be all-or-none, but varies along a
  continuumsome candidates are more activated than
  others. pp.60 Harley.

17
Need to Revise the Cohort Model- Further Evidence

• We are capable of identifying a word when
  mispronounced (even at the beginning e.g.
  shigarette, and (sometimes) when we only hear a
  word from the middle on.
• The original cohort model cannot account for
  these effects

18
The Revised Cohort Model

• Initial activation is (still) bottom-up
• Competition between active elements leave one
  element standing out above the rest.
  Incompatible bottom-up evidence does not
  eliminate a candidate (as it does in original),
  but partially deactivates it.
• Thus, revised version of model is much more
  similar to TRACE
• The highest ranking elements are assessed in
  parallel with respect to the interpretation the
  best fit is integrated and (hence) recognized.

19
Activation in the Revised Cohort Model
elephant
energise
dog
activation
wombat
elegant
captain
time
captive
c a p t i n
20
Neighbourhood effects

• People are faster to recognise a high frequency
  word which only has low frequency neighbours than
  vice versa. This effect is compatible with the
  revised cohort model.
• However, the model predicts that the size of the
  cohort at any point (number of competitors) does
  not affect the speed at which a target is
  recognised, only the time to reach uniqueness.
  However, cohort size does affect the time course
  of word recognition (Luce et al.).

21
Spoken Word Recognition Conclusions

• The two leading models, TRACE and the Revised
  Cohort Model, have much in common
• Both depend on competition between partially
  activated candidates for the words identity

22
Language and Communication

• Lexical Ambiguity

23
Lexical Ambiguity

• What happens when a word form (visual or
  auditory) is associated with two (or more)
  meanings, rather than one (e.g. bank, straw)?
• The appropriate meaning is usually determined by
  context
• As readers or listeners we dont usually notice
  the ambiguity, but what effect does it have?

24
Lexical ambiguity

• Where is the ball?
• Look at that chip.

25
Lexical Ambiguity - MacKay 1966

• Sentence completion task
• After talking the right/left turn at the
  intersection, I.
• Harder to complete after right (ambiguous) than
  left (unambiguous)

26
Lexical Ambiguity - Models

• Context-guided (selective) access (Schvaneveldt)
• Appropriate meaning is chosen by context, others
  are not considered
• But how could it work?
• Ordered Access (Hogaboam Perfetti, 1975)
• Most common meaning checked first
• Accepted if it fits context
• Otherwise other meanings are checked
• Multiple Access (Swinney, 1979)
• All meanings accessed, context selects among them
• Reordered Access Model (Duffy, Morris, Rayner,
  1988)

27
Contradictory evidence from the 1970s

• For selective access e.g. Hogaboam Perfetti,
  1975, ambiguity detection task Found longer RT
  when word used with it more common meaning (e.g.
  ink pen, rather than sheep pen)
• The accountant filled his pen with ink.
• The farmer put the sheep in the pen.
• For multiple access e.g. Foss, 1970, phoneme
  monitoring People are slower to detect /b/ in A)
  than in B) because straw is ambiguous, even
  though the context (farmer) strongly favours
  one meaning. Suggests both meanings accessed.
  (compare the old MacKay finding)
• A) The farmer put his straw beside the machine.
• B) The farmer put his hay beside the machine.
• However, this task is sensitive to the length of
  preceding words. A short word may not be fully
  processed before the next word begins. A longer
  word can be identified before their end. Problem
  majority of polysemous English words are short.
  If this is controlled for, the effect disappears
  (Mehler et al., 1978).
• So which view is correct?

28
Swinney, 1979
Context none (bugs or insects) or biasing
(spiders, roaches, and other bugs/insects)
Words Ambiguous (bug) or unambiguous (insect)
Rumour had it that for many years, the government
building had been plagued with problems. The man
was not surprised when he found several (spiders,
roaches, and other) bugs (insects) in the corner
of his room.
In both contextsambiguous 1. Ant spy, gt sew
ant
sew
spy
2. Ant gt spy and sew
29
Replicated by..

• Onifer Swinney, 1981 with biased ambiguous
  words
• Tanenhaus, et al., 1979, naming task -
  context-independent meaning fades after ? 200
  msec
• Seidenberg, et al., 1982, for a few 100 msec all
  meanings of an ambiguous word are activated
  regardless of semantic and syntactic constraints.
• Results support a modular view of sentence
  processing. Challenges by some, but generally
  accepted view that lexical ambiguity is resolved
  by an interaction b/w frequency and context

30
Syntactic context (I)

• Tanenhaus et al. (1979)
• John began to watch ...

look / time

• Syntactic context does not constrain multiple
  access

31
Syntactic context (II)

• Shillcock Bard (1993)
• John decided that he would ...
• John decided that wood ...

plank / blank

• ??????????????

32
Lexical Ambiguity - Balanced and Biased
Ambiguities

• The meanings of some ambiguous words (balanced)
  are roughly equally common
• For others (biased) one meaning is much more
  common than the other(s)
• Onifer Swinney (1981) replicated Swinneys
  (1979) results for biased ambiguities
• However, others have claimed that only balanced
  ambiguities show multiple access

33
Lexical Ambiguity - Types of Context

• Contexts may be more or less strongly biased
  towards one or other meaning of an ambiguous
  word.
• Maybe selective access occurs only with strongly
  biasing contexts
• Contexts may be consistent with specific
  properties of one or other meaning of an
  ambiguous word (Tabossi)

34
Lexical Ambiguity - Types of Context

• Relevant context may come either before or after
  ambiguous word
• The footballer asked where is the ball?
• Where is the ball? asked the footballer.
• A man in a tuxedo asked where is the ball?
• Where is the ball? asked the man in a tuxedo.
• Context that follows an ambiguous word is
  unlikely to affect the process of word
  identification

35
Lexical Ambiguity Conclusions

• The Swinney 1979 study provided striking evidence
  for multiple access
• Later studies have found that evidence for
  multiple access is clearest with balanced
  ambiguities and contexts that are not too
  constraining