Spoken Word Recognition Flashcards
mental lexicon
a dictionary of the lexical units of their language
words play a pivotal role in the acquisition of language
• Functional units for translating between
perception, production and comprehension
– Lexical symbols convey both grammatical and semantic information
– Sublexical structure ‘grounds’ language in the motor processes involved in production
definition of scientific model
-a representation (a diagram, verbal
description, computer program, etc.)
-of an empirical phenomenon
that captures
-the fundamental characteristics of that phenomenon
-simplifies and substitutes for the phenomenon
models of cognitive processes in spoken word recognition
- Logogen model
- Cohort model
- TRACE model
the logogen model (morton, 1970)
- the same lexical (logogen) system accessed by both spoken and written words
- bi-directional links between logogen system and the cognitive system
- separate logogens (evidence collecting devices) for each word that are activated in parallel according to match with input
- evidence can come from visual, auditory or cognitive sources
- when evidence in a logogen exceeds a threshold it fires –> word is identified
implications of logogen model
- Frequency effects: Less evidence required to recognise a common (high frequency) word than a low frequency word (eg bright vs blight)
➔Auditory recognition faster for more frequent words (Soloman & Howes
1951 etc etc)
Threshold for each logogen to fire is permanently reduced each time the logogen is activated ➔ lower for more frequent words
- Context effects: Semantic information from cognitive system contributes evidence for a particular word
• Easier to recognise word at the end of a sentence than in isolation
• Words in noise more easily recognised in sentences (Miller et al., 1951)
limitations: did not specify how evidence accumulates, or how the system ensures that the correct word node reaches threshold first
eg why we do not ALWAYS confuse a low frequency word with similar higher frequency word eg mud with mum; blight with bright
the cohort model
- based on auditory shadowing studies: close shadowers identified words before they finished, particularly in context, restored mispronunciations if not in first syllable (eg asfinaut)
- -> speech signal processed sequentially
- early in the presentation of word, all known words that begin with the presented sounds are activated: word initial cohort
- as more perceptual evidence becomes available, words are eliminated from the cohort until only one left –> successful word identification
empirical evidence for cohort model
- Gating paradigm
- slices of information given to identify a word, more information, less choice
- Words with early uniqueness points recognized faster (ie earlier uniqueness/recognition points) than words with late uniqueness points ➔ earlier elimination of competitors from cohort
- Predictive context: faster recognition point➔ context reduces cohort - Phoneme monitoring task
- Faster identification for phonemes later in word eg alligator
problems with cohort model
Problem 1: the cohort is NOT defined solely by sequence of phonemes
• people successfully recognise words with incorrect initial phoneme eg bleasant for pleasant
• High frequency words faster than low frequency when uniqueness point matched eg street> streak
• Words with more ‘phonological neighbours’ slower to identify
➔Revised model: Cohort membership graded, not all-or-none
• cohort includes words with similar initial phonemes
• strength of activation of cohort members depends on frequency, number of neighbours
Problem 2: the effects of context can override the cohort: contextually predictable words activated even if incompatible with cohort
-cross-modal priming vs visual world paradigm
➔Bottom-up priority: Word initial cohort defined by bottom-up information only, but context influences whether words remain in
cohort, despite contradictory/ambiguous phonetic evidence
Cross-modal priming
Present auditory sentence
• Occasional visual words that participants must classify as word or nonword
eg Zwitserlood (1989):
Participants hear:
With dampened spirits, the sailors stood around the grave to mourn the loss of their captain
Visual probe at cap or after captain
Lexical decision to SHIP or SLAVE
/kap/ primes both ship and slave
BUT after uniqueness point /kaptn/ only ship is primed
➔Semantic associates of competitor words are active before uniqueness point , even if incompatible with context
➔Context only used to select between candidate words late in processing
Visual world paradigm
Eye movements monitored while Ps responded to auditory instructions like:
‘Click on the beaker’
German listeners heard: The woman irons the: Bluse (blouse) Blume (flower Wolke (cloud) ➔ More fixations on Bluse than Blume from very early in processing, even though both compatible with cohort
the TRACE model
connectionist/neuronetwork model
-hierarchical architecture:
phonetic features –> phonemes –> words
-activation spreads to matching nodes at higher and lower levels
- identification occurs when activation in a node exceeds threshold
- threshold depends on frequency, less evidence needed for common words
- competitive selection
- facilitatory activation spreads both bottom-up and top-down through the network
- Lateral inhibition between representations within levels
- Competition between activated nodes ➔ survival of the fittest
- select best matching node
- -> interactive activation
- -> winner takes all network
evidence for TRACE model
- Interactions between top-down and bottom-up activation
eg Ganong effect: Whether /?ask/ is heard as task or dask depends on strength of bottom-up information (how much of d/t present) - Interactions between letters and word levels
- Eye tracking in visual world paradigm, people are sensitive to rhyme as well as initial competitors(eg Allopenna et al., 1998)*
-Eye movements monitored while Ps responded to instructions like: ‘Click on
the beaker’
-Distractors chosen to assess initial (eg beetle) and rhyme (eg speaker) overlap
➔ All information processed in parallel – not just serially defined ‘cohort’
Cohort neighbour active first, then rhyme more active
➔Connectionist model accurately simulates human data
Conclude: Words that are not part of initial cohort can still compete with target word – not eliminated from the cohort
problems for TRACE model
Predicts stronger top down reliance than people show:
• prediction of poor phoneme detection in nonwords closely matched to real words (e.g. vocabutary) not empirically supported
• Poorer at detecting mispronunciations than people are ➔too sensitive to top-down influences
• People only show top-down effects when stimuli are degraded
Model is BOTH too limited and “too flexible”
• Computer simulations based on small vocabulary of 1-syllable words – may not generalise to realistic vocabularies
• Other factors also influence word recognition e.g. orthographic information, syllable duration
• Model has lots of parameters and can produce many outcomes
➔ difficult to falsify
Summary
Cohort & TRACE model remain influential, heuristically useful frameworks and underpin many later models
Agree
• several candidates are activated by perceptual input
• top-down information contributes to selecting between them
Disagree
• which acoustic “chunks” are analysed (phonemes vs phonetic features) and which candidates are activated
• Role of competition in selecting best candidate
• extent and timing of top-down influence of context
Similar issues arise in models of visual word
recognition
