Laura M. Gonnerman and David C. Plaut
Department of Psychology, Carnegie Mellon University
Center for the Neural Basis of Cognition
Poster presented at the Cognitive Neuroscience Society Annual Meeting
April 2000
Morphological structure is widely assumed to govern the organization of the mental lexicon. Thus, words such as teacher and baker are thought to be represented and/or processed in terms of their component morphemes. Evidence for this position has been obtained using several experimental paradigms across a wide number of languages (see papers in Feldman, 1995).
A number of researchers have begun to question assumptions underlying traditional accounts of morphological representations and processing (e.g., Gonnerman et al., 1999; Plaut & Gonnerman, 1999; Rueckl et al., 1997). On the alternative view, lexical knowledge is distributed across simple, neuron-like processing units that encode information about the sound and meanings of words and morphology arises from the correlations between these codes. In earlier work, Gonnerman, Andersen, and Seidenberg (1999) found that the degree of semantic similarity between primes and targets predicted the magnitude of priming in a series of cross-modal lexical decision studies.
Data from masked priming experiments have been used to support the traditional view of the mental lexicon where morphemes are stored discretely and complex words are decomposed in processing. In masked priming, visual primes are presented briefly and masked so to that subjects cannot consciously process them. This paradigm, sometimes called ‘form priming’, is generally considered insensitive to pure semantic effects (e.g., Frost, 1999). Thus, results showing greater facilitation for morphologically complex (cars-car) compared to orthographically overlapping prime-target pairs (card-car) are argued to arise from morphological rather than semantic factors.
Such morphological effects have been found using the masked priming paradigm for several languages, including English (Forster, Davis, Schoknecht, & Carter, 1987), French (Grainger, Colé, & Segui, 1991), Dutch (Drews & Zwitserlood, 1995), and Hebrew (Frost, Forster, & Deutsch, 1997).
However, there have been recent reports of reliable masked priming effects for semantically related items in both English (Williams, 1996) and Spanish (Perea and Gotor, 1997). In the two experiments reported here, we investigate the possibility that morphological effects in masked priming arise from the conjunction of orthographic and semantic information. If this is the case, then semantic relatedness should modulate the magnitude of priming effects among orthographically related words, irrespective of the presence of pure semantic priming.
133 USC undergraduates rated word-pairs for similarity in meaning on a 7-point scale. Results show that subjects are sensitive to degrees of semantic similarity between related words. Sample ratings are presented below.
|
_____________________ |
||
|
message-mess 1.1 |
Less related |
|
|
pillage-pill 1.5 |
• |
|
|
• |
• |
|
|
• |
• |
|
|
lately-late 3.4 |
• |
|
|
shortage-short 4.1 |
• |
|
|
• |
• |
|
|
• |
• |
|
|
darken-dark 6.2 |
• |
|
|
hunter-hunt 6.5 |
More related |
|
Materials: We used semantic relatedness ratings to select prime-target pairs (e.g., likely-like). For each prime word (e.g., likely), a control word (e.g., senior) was selected that matched in frequency, number of syllables, number of letters, and part of speech. An equal number of non-word fillers were included, some phonologically related (e.g., slither-slith), others not (e.g., basil-grook).
Targets were matched across the five conditions for:
Sample stimuli and mean relatedness ratings for each condition. A score of 1 is unrelated, 7 is highly related.
|
Condition |
Prime-Target Example |
N |
Mean semantic relatedness |
|
|
1. hi orth, lo sem |
yellow-yell |
28 |
1.2 |
|
|
2. hi orth, lo sem |
hardly-hard |
28 |
1.9 |
|
|
3. hi orth, mid sem |
lately-late |
28 |
3.9 |
|
|
4. hi orth, hi sem |
boldly-bold |
28 |
6.1 |
|
|
5. lo orth, hi sem |
profit-gain |
28 |
6.0 |
|
Participants: 46 CMU undergraduate students.
Procedure: There were 20 practice trials, followed by 4 warm-up trials, and the 280 test trials. Participants were run in one block that lasted about 12 minutes. A trial proceeded as follows:
|
Fixation 1000 ms |
* |
|
Mask 500 ms |
$\%\#\#\$\%\$\#\&@\#@ |
|
Prime 39 ms |
boldly |
|
Target 200 ms |
bold |
|
1000 ms delay after response |
lexical decision |
Following the procedure used in Frost et al. (1997), the primes were masked by targets presented in lower case in a 20% larger font than the primes.
 |
Experiment 1. Priming effects for orthographically transparent suffixed-stem pairs with different degrees of semantic relatedness: Low (Conditions 1 and 2), Mid (Condition 3), High (Condition 4: orthographically related; & 5: orthographically unrelated). |
Semantically unrelated words did not prime, whether there was an orthographic segment resembling a suffix (e.g., yellow-yell) or not (e.g., hardly-hard).
Semantically related words did prime, and the degree of relatedness affected the magnitude of the priming, with moderately related words (e.g., lately-late) priming less than highly related words (e.g., boldly-bold), that is, 20 and 30 msec, respectively.
Priming was also found for semantically related pairs with no orthographic overlap (i.e., 21 msec for profit-gain pairs).
Hypothesis: the semantic priming found in Experiment 1 may have arisen due to explicit recognition of the prime. Therefore, in Experiment 2 we change the masking to ensure that participants cannot consciously process the primes; targets were presented in uppercase letters rather than in a larger font, as in Experiment 1. The stronger masking should eliminate semantic priming, if the effect is due to conscious priming.
Materials: The same materials were used as in Experiment 1.
Participants: 48 CMU undergraduate students.
Procedure: The procedure was the same as in Experiment 1, except that the targets were presented in uppercase letters in the same size font as the primes. Also, the delay between trials was reduced to 500 ms.
 |
Experiment 2: Priming effects for orthographically transparent suffixed-stem pairs with different degrees of semantic relatedness: Low (Conditions 1 and 2), Mid (Condition 3), High (Condition 4: orthographically related; & 5: orthographically unrelated). Primes are more effectively masked by targets compared to Experiment 1. |
After completing the experimental task, participants were asked to describe a trial as if they were explaining what they just did to someone who was not present. None of the 48 participants reported seeing the primes. When then explicitly queried about what appeared just prior to the target, some participants responded that they saw something flash, and a few mentioned letters. Therefore, we are reasonably confident that the stronger masking prevented conscious processing of the primes.
The pattern of results parallels that found in Experiment 1, although the magnitude of the priming decreased overall. Thus, when participants are unaware of the presence of the prime, the degree of semantic similarity between the primes and targets influences the magnitude of the priming effect.
In addition, the finding of pure semantic priming was replicated. Highly semantically related items with little form overlap (e.g., profit-gain) produced significant facilitation.
In two experiments we found that the degree of semantic similarity between orthographically transparent primes and targets effects the magnitude of facilitation (i.e., boldly-bold primes more than lately-late or hardly-hard). These graded priming effects support a distributed connectionist approach to morphological processing.
The graded priming effects reflecting degree of both semantic and orthographic similarity are awkward for accounts in which morphological decomposition is an all-or-none phenomenon. Given the continuous nature of relatedness and priming, it is unclear how proponents of hybrid models (e.g., Marslen-Wilson, Tyler, Waksler, & Older, 1994) might delimit the set of complex words to be decomposed from those that are stored or accessed as wholes.
In addition to the finding that degrees of semantic relatedness affect the magnitude of priming, we also found priming for semantically related items that shared no morphology (see also Williams, 1996). The word pairs that produced this facilitation were all synonyms, and as such were very highly related.
This surprising pattern of results suggests that 1) semantic information is available very early on, even in the absence of conscious processing, and 2) the failure to find pure semantic priming in previous studies may result from using less strongly related word pairs.
Moreover, the graded priming effects support a view of morphology as an interlevel representation that mediates mappings between semantics and phonology and that emerges in the service of language acquisition and processing.
On this view, morphology reflects structure present in the world: language input contains patterns that are picked up on by language learners to the extent that they are useful in solving the primary tasks of competent speakers, that is comprehending and producing speech.