1. Phonological Priming and Lexical Access in Spoken Word Recognition
Christine P. Malone
Minnesota State University Moorhead

2. Problems
How do the processes occurring during early stages of spoken word recognition affect single-word shadowing (naming) performance?
How is a string of incoming phonetic features mapped onto a remembered lexical item?
How does phonological information influence the organization of lexical items in memory?
It is generally agreed that during spoken word recognition, listeners automatically evaluate the unfolding input by activating a set of potential lexical candidates, which then compete for recognition. However, numerous questions remain about how the set of possible candidates is established, what characteristics these candidates share with the incoming stimulus, and how the activated set is evaluated during real-time processing. For example, little is known about how potential candidates are related in terms of phonological information (i.e., shared sounds), especially when the incoming stimulus consists of more than one syllable. Thus, it is not clear whether as the word motivate is heard, the word recognition system considers all words containing matching sounds and matching relative location (e.g., motorist, innovate) or whether all words with the same sounds, regardless of relative location are considered (e.g., demote, innovate). Further, does candidate activation occur regardless of the lexicality of the unfolding stimulus? Questions like these have important implications for how we understand and model the word recognition system.

3. Background on Activation
During spoken word recognition, listeners automatically evaluate the unfolding input by activating a set of potential lexical candidates, which then compete for recognition.
The incoming sound pattern determines the potential candidates.
Degree of activation is determined by match between the potential candidates and the unfolding sensory input.
Explain deviation point.
Explain set of candidates as proposed by cohort theory.

4. Connectionist Models
Difficulty obtaining facilitation following beginning phonological overlap across different tasks (Connine, Blasko, & Titone, 1993).
Theoretical interest turned from cohort to connectionist theory.
Multi-level architecture composed of simple processing units, called nodes. Adapted from visual word recognition.
Connine, Blasko, & Titone (1993)—obtained facilitation in cross-modal priming paradigm when minimum change nonwords were primes, whether initial or medial phonemes were altered. Overall goodness-of-fit explanation. This can be explained with the stimuli
For experiment 1 or at the predictions part.

5. Levels of Speech Processing
Feature Level--Break the speech stream into phonetic features, such as voiced/voiceless.
Phoneme Level--Interpret stream of features and produce a prelexical representation.
Lexical Level--Identify the word.

6. A Simplified Connectionist Model (adapted from Colombo, 1986)
WORD
LETTER
A connectionist model constructed for visual word recognition and eventually applied to spoken word recognition was put forward by Colombo (1986). The model assumes processing takes place at 3 levels—features, letters, and words. Each processing unit at each level is called a node. Activation and inhibition are assumed to operate at each level. However, activation (an excitatory message) can only travel between levels (I.e., feature nodes send activation to letter nodes, or letter nodes to word nodes). Activation cannot spread to nodes at the same level. Inhibition can occur between levels (e.g., from letters ro words) or within the same level (from a word to other words). Inhibition has the effect of decreasing the activation level of the recipient node.
Intra-level inhibition presumably occurs because nodes at the same level are competing with each other for access to output. Upon reaching its threshold level of activation, each unit begins inhibiting all other connected units.
FEATURE
INPUT

7. Levels of Speech Processing
Feature Level—Break speech into phonetic features (e.g., voiced/voiceless)
Phoneme Level--Interpret stream of features and produce a pre-lexical representation.
Lexical Level--Identify the word.
WORD
LETTER
FEATURE
INPUT

8. Priming and the Naming Task
Shadowing task (naming) involves lexical processing, but is relatively unaffected by postlexical processing.
Effect of having recognized the prime on recognizing the target?
Phonological priming--assess differential levels of residual activation when manipulating phonological overlap and lexicality of prime.
Measurement of the activation is taking place immediately after its presumed to occur

9. Experiment 1 Stimuli Match Mismatch Target: motivate
Match Mismatch
Early Overlap motorist demote
Late Overlap innovate atrium
Unrelated vocalist vocalist

10. Hypotheses
If inhibition takes place among beginning phonemes, then Early Overlap/Match targets will have longer latencies compared to Late Overlap/Match targets.
If overall match (and not location) is important in activation, then Mismatch pairs should show same patterns as Match pairs.
In monosyllabic word shadowing studies, get inhibition with 2 or 3 overlapping initial phonemes; get facilitation with 2 or 3 ending phonemes.
Overall goodness of fit explanation—then relative location doesn’t matter? Never been tested.

11. Method Auditory priming paradigm, 100 ms ISI
Single word shadowing (or naming) task, each list contained 8 EM, 8 EMM, 8 LM, 8 LMM, and 8 Unrelated pairs.
Digitally recorded stimuli (22kHz, 16-bit) using SoundEdit and presented via PsyScope.
Voice-activated reaction times recorded from target onset until beginning of vocal response.

12. Graph
EM vs. EMM sig. (Early MISMATCH is 31 ms faster than early match).
LM vs. LMM sig (Late MATCH is 39 ms faster than late mismatch).
EMM vs. UN sig. (facilitation)
EM vs. UN not sig.
LMM vs. UN not sig.
LM vs. UN sig. (facilitation)
These primed naming studies provide evidence for facilitation at the phoneme level
Have yet to obtain evidence for competition (inhibition) at the word level for long words…routinely obtained for single syllable words. Maybe ignoring N is obscuring effects….

13. Shadowing Latencies
Reaction times to LM pairs were consistently faster than those to LMM pairs
Reaction times were significantly faster following LM pairs as compared to Unrelated pairs (i.e., facilitation).
On average, response latencies following Early pairs were equal to those following Unrelated pairs.

14. Conclusions
Shared beginnings slowed naming of target (inhibition) for word and nonword targets.
Potential candidates are inhibited based on matching beginning information, supporting connectionist architecture.

15. Applications
Questions regarding the lexicon architecture have important implications for how we understand and model the word recognition system.
Empirical data is useful for scientists studying language processing, as well as for scientists developing speech recognition systems.

16. Questions?