- 1:00 Welcome by Mike Hammond
- Introductory remarks by Andy Wedel
Friday, April 25, 2008: Harvill rm. 115
- 1:15 - 2:15 Harry Tily
- Diachronic processing preferences and their implications for models of syntactic change
- 2:15 – 3:15 Neal Snider
- An exemplar model of syntactic production
Tea Break
- 3:30 – 4:30 Joan Bresnan
- Predicting Syntax: Processing Dative Constructions in Two Varieties of English
- 4:30 – 5:30 Discussion
Dinner at Poca Cosa
- 8:30 Coffee
Saturday, April 26, 2008: Harvill rm. 102
- 8:45 – 9:45 Rob Malouf, Farrell Ackerman and Jim Blevins
- Inflectional morphology as a complex adaptive system
- 9:45 – 10:45 Melissa Redford
- Meaning and Mechanics in Speech and Language Acquisition
Tea Break
- 11:00 – 12:00 Eduardo Altmann
- Recurrences in processes with long-term memory
Lunch
- 1:30—2:30 Robert Daland
- Language variation: convergence, divergence and death
- 2:30 – 3:30 Colin Dawson
- 'Second-Order Learning' as a Source of Structure Stabilization in Both Individual Learning and Cultural Evolution
Tea Break
- 3:45 – 4:45 Andy Wedel
- Modeling sublexical contrast maintenance as an emergent effect of lexical category competition
- 4:45 – 5:45 Clay Beckner and Andy Wedel
- Modeling contributions of usage versus acquisition to language change
Department Potluck party at 7 at Adam Ussishkin and Andy Wedel’s house
An exemplar model of syntactic production
Neal Snider
Exemplar models of syntactic storage predict that there should be little or no difference between lexical and syntactic production, because the same representations are being accessed. I present an exemplar production model and supporting evidence that structural priming and lexical priming are influenced by some of the same factors. Many studies have shown that lexical priming exhibits sensitivity to three key effects: the frequency of the prime, the neighborhood density of the prime, and the similarity between the prime and the target. Words are primed less by their high frequency orthographic and semantic neighbors than by low frequency neighbors (Scarborough etal. 1977; Thomsen etal. 1996). Also, words with fewer orthographic and semantic neighbors prime more (Perea and Rosa 2000; Anaki and Henik 2003). Finally, the more similar the prime and target words, the greater the magnitude of the priming effect (Ratcliff and McKoon 1988). All of these effects have been argued to follow from an activation-based model (Anderson 1983; Kapatsinski 2006). Inverse frequency effects arise because less activation is left in the prime type when the prime is more token frequent because activation also spreads to the stored tokens. Similarly, primes in more dense neighborhoods have less activation left in the prime type because more activation spreads out to other similar types. Finally, more activation spreads between more similar types because of the greater resting activation between them due to their similarity. If the increasing evidence for storage of linguistic representations beyond the lexical level is right (Bod 2001; Schmitt and Galpin 2004; Schmitt, Grandage, and Adolphs 2004; Tremblay et al 2007), then one would expect lexical and syntactic priming to behave similarly, because the same representations are being accessed. I have new evidence from corpus-based studies on spontaneous speech that syntactic priming exhibits the same three effects.
In two studies, one using the dative alternation:
- "Šwe give a country moneyŠ" (Switchboard corpus)
- we give money to a country
and another using the passive alternation:
- "Šshe was charged with murderŠ" (Switchboard corpus)
- they charged her with murder
I use statistical modeling of naturalistic corpus data to present evidence for an inverse frequency effect in structural priming: prime structures that contain verbs that occur very frequently in that structure are less likely to prime it. In a corpus, the factors that lead a speaker to produce one structure over another are necessarily complex, so for each study, I used a database that contained many factors that had been reported in the literature to affect construction choice, analyzed with advanced statistical models (Logit Generalized Linear Mixed Models) to show that the effects of priming were not merely due to the prime and target structures being in similar syntactic and pragmatic environments. For the ditransitive study, I used the Spoken Dative Database (Bresnan et al 2007). For the passive study, I automatically extracted a new database from the Switchboard corpus consisting of 1757 actives and 557 passives. To ensure that I was modeling a true syntactic variable, I only included actives that had generic subjects (e.g. "you get to meet different people", "where they have the river patrol cops", etc.), so both alternants in the database had the same number of expressed arguments (namely, one). I automatically annotated these sentences for various information structural properties of the argument (givenness, animacy, length, definiteness, pronominality), as well as social factors (age and gender). These factors were controlled in the analysis by the mixed model, as well as speaker effects, to allow for each speaker to possibly have a different rate of passivization. The tendency for speakers to be more likely to repeat infrequent passive structural primes persists, just as in lexical priming, when all these factors are controlled (p<.02). I also find this effect in priming in the ditransitive alternation (Jaeger and Snider 2007). The inverse frequency effect in ditransitive structural priming has been experimentally reproduced by Bernolet and Hartsuiker (2007).
In another study, I find an effect of similarity on structural priming. Structural and semantic similarity of the prime and target structures are modeled using a nearest-neighbor (NN) similarity metric and the database of passives and ditransitives extracted from the Switchboard corpus. More similar prime and target exemplars are more likely to occur in the same construction (p<.03). This effect is in addition to the known similarity effect of verb identity (Pickering and Branigan 1998), which is controlled through simultaneous multiple regression and model comparison. This effect is also found in priming the ditransitive (Snider 2007). These results mirror the similarity effects found in lexical priming.
Finally, I find that neighborhood density, defined as the number of constructions in which the verb appears in a large corpus, has an effect on the likelihood of priming. Using the same dataset, passive prime verbs that occur in fewer other constuctions are more likely to cause a passive to be produced in the target (p < 0.03). Structural priming seems to be affected by the same kinds of neighborhood density effects as lexical priming.
In summary, I have found evidence that both lexical and structural priming are affected by prime frequency, prime-target similarity, and prime neighborhood density. This suggests that lexical and structural priming could be the same process due to the similar representations being accessed in their production.