Syntactic Priming
Syntactic Priming
Syntactic Priming
INTRODUCTION
As a variety of studies has shown, speakers tend to repeat syntactic structures they have just encountered (produced or comprehended) before. This
tendency has been referred to as structural priming, syntactic persistence
or syntactic priming; I will use the latter term throughout the remainder
366
of this paper.2 Levelt and Kelter (1982) and Weiner and Labov (1983)
belong to the earliest studies to which the identication of this phenomenon is commonly attributed. The former study found that merchants in
the Netherlands tended to formulate their answers to questions such that
the syntactic structure of their answers was identical to that of the questions; for example, the Dutch equivalents of the questions in (1a) and (1b)
tended to trigger (2a) and (2b) respectively as answers.
(1) a.
b.
(2) a.
b.
I adopt the denition of syntactic priming proposed by Branigan et al. (1995: 490): We
dene syntactic priming as the proposal that processing a particular syntactic structure
within a sentence affects the processing of the same (or a related) syntactic structure within
a subsequently presented sentence; cf. Szmrecsanyi (2005) for discussion of whether persistence is the more appropriate cover term (specially for corpus-based studies).
367
the dative alternation just referred to, are identical in terms of their metrical structure but differ in terms of their different syntactic structures
(adopted from Bock & Loebell, 1990:24), which is why (3a) results in a
signicant preference of prepositional datives (relative to a baseline/control
condition) whereas (3b) does not.
(3) a. Susan [VP brought [NP a book] [PP to Stella]]
b. Susan [VP brought [NP a book] [S to study]]
Similarly, syntactic priming does not derive from the presence of
closed-class lexical items in particular slots of the sentences or event-structural or thematic utterance characteristics. As to the former, Bock (1989),
for example, provides experimental evidence showing that the priming
effects she obtained cannot be explained by reference to closed-class lexical items involved in the dative alternation, viz. to and for. As to the latter claim, consider the examples in (4) and (5). While (4a) and (4b) are
different in terms of their thematic structure, they are identical in terms of
their syntactic structure and, thus, both (4a) and (4b) prime prepositional
datives (relative to control conditions); the same holds for (5a) and (5b),
both of which prime passives.
(4) a. The wealthy widow [VP gave [NP an old Mercedes] [NPrecipient
to the church]]
b. The wealthy widow [VP drove [NP an old Mercedes] [NPlocative
to the church]]
(5) a. The 747 was alerted [PP agentive by the airports control tower]
b. The 747 was landing [PP locative by the airports control tower]
Subsequent experimental work has mainly focused on spoken English,
but has also been concerned with written English as well as Dutch
(cf. Hartsuiker & Kolk, 1998; Hartsuiker, et al., 1999; Hartsuiker &
Westenberg, 2000) and German (cf. Scheepers & Corley, 2000) in both
speaking and writing. The range of experimental methodologies has
also been broadened considerably and now includes a wide variety of
ofine experimental paradigms such as sentence completion tasks (cf., e.g.,
Pickering & Branigan, 1998; Hartsuiker & Westenberg, 2000 etc.), sentence
recall tasks (Potter & Lombardi, 1998), and picture descriptions in dialogs (cf. Branigan et al., 2000). In addition, Smith & Wheeldon (2001) and
Corley & Scheepers (2002) did online studies where priming effects were
also measured in terms of production latencies. While most studies have
investigated the dative alternation and the activepassive alternation in
English (as in (4) and (5)), more recent work has also looked at the equivalent constructions in Dutch as well as Dutch locative PP alternations, the
368
order of auxiliary verb and past participle in Dutch as well as dative-accusative verb alternation in German (cf. Scheepers & Corley, 2000), and the
order of syntactic functions in Japanese (cf. Yamashita et al. [2002]).
The currently most pressing issues concerning syntactic priming (many
of which will also be addressed in the present approach) are the following:
(i) the duration of syntactic priming: on the one hand, Levelt and
Kelter (1982) and Branigan et al. (1999) report that priming (in
spoken and written production respectively) is fairly short-lived.
On the other hand, other studies report priming effects across
longer time interval or more intervening material (cf. Bock &
Grifn, 2000; Pickering et al., 2000; Chang et al., 2000).
(ii) the directionality of syntactic priming: Branigan et al. (1995)
discuss a variety of different studies which, taken all together,
support the assumption that syntactic priming can operate from
production to production (cf. Bock, 1986; Bock & Loebell, 1990),
from comprehension to comprehension (cf. Branigan et al.,
1995 for an overview) and from comprehension to production
(cf. Branigan et al., 2000; Bock, 2002).
(iii) the grammatical characteristics of the priming verb: Pickering and
Branigan (1998) found that (a) syntactic priming is stronger if
the priming verb lemma and the target verb lemma are identical
(compared to different lemmas in prime and target) and that (b)
morphological differences between the priming verb and the target verb (in terms of tense, aspect and number) do not result in
strongly varying priming strengths.
(iv) the degree to which syntactic priming is asymmetric and construction-(pair)-specic: From a between-alternations perspective,
Bock (1986: Exp. 1) found stronger priming for the two syntactic
frames involved in the dative alternation than for those involved
in the activepassive alternation in English; a similar prominence
of datives over transitives was found for English by Potter and
Lombardi (1998: Exp. 3) and for Dutch by Hartsuiker and Kolk
(1998). In addition, from a within-alternation perspective, further
asymmetries were sometimes obtained: Bock (1986: Exp. 1) found
there was stronger priming for ditransitives than for prepositional
datives while Potter and Lombardi (1998) report the opposite
(and Pickering et al., 2002: 587 mention evidence for symmetric/balanced priming).
(v) the degree to which syntactic priming is language-specic:
Hartsuiker et al., (2002) demonstrate syntactic priming from
comprehending Spanish to producing English, Salamoura (2002:
369
370
V NP PPfor/to
V NP/S NP/S
Row totals
spoken
written
Column totals
926
854
1780
1254
759
2013
2180
1613
3793
371
The variable VFormID is of course nested into VLemmaID since, if the lemmas of CPrime
and CTarget are already different, the forms cannot be identical anymore.
4 Benedikt Szmrecsanyi (p.c.) pointed out to me that collapsing switch rates of different
speakers or of different corpus les this way may be dangerous: It is possible that the conation of, for example, two corpus les in which no priming takes place may result in a
summary table in which priming shows up as a statistical artifact. He therefore recommends
using scatterplots of the kind used by Sankoff and Laberge (1978), in which for each
speaker or le the relative frequency of a construction on the x-axis is plotted against the
ratio of switches to one construction on the y-axis; to my mind, this is comparable to byitem statistics as used in ANOVAs. It follows that only if most dots are located below the
372
Table II. Application of the Coding Scheme to Priming from (6a) to (6b)
Variable
Value/level
Medium:
CPrime:
CTarget:
CID:
Distance:
SpeakerID:
VFormPrime:
VLemmaPrime:
VFormTarget:
VLemmaTarget
VFormID
VLemmaID:
spoken
prepositional dative
ditransitive
no
1
no
gave
give
shows
show
no
no
Table III. CPrime CTarget: Observed vs. Expected Frequencies ( 2 (1) = 202.4, p < .001)
Data (ICE-GB)
CPrime: V NP PPfor/to
CPrime: V NP/S NP/S
Column totals
CTarget:
V NP PPfor/to
CTarget:
V NP/S NP/S
Row
totals
830 (647.1)
556 (762.1)
1386
549 (731.9)
1068 (861.9)
1617
1379
1624
3003
for the most general result, namely the interaction of CPrime and CTarget
across all other variables; the expected frequencies are provided in parentheses and are not computed on the basis of row and column totals but
on the basis of row totals and the overall frequencies of the two constructions as listed in Table I.5
main diagonal, switches (from one construction to the other) are rarer than the null hypothesis of the absence of priming would predict. In order to show that the summary tables
used in the present data set do not suffer from such an articial ination, I also provide
corresponding scatterplots; for that of the datives, cf. Fig. (2)
5 The question may arise why the expected frequencies are not computed the usual way.
The reason for this is the following. If one uses the column totals from Table III for the
computation of the expected frequencies, one treats these as given, as an independent variable so to speak, while in the present design the column totals are of course part of the
dependent variable, namely the frequency of one (target) construction as a response to
some (prime) construction. The more appropriate logic underlying the present way of computation is this: after each prime construction, the speaker has two constructional choices,
and the probabilities of each of the two constructional choices are the frequencies with
373
which these constructions occur in the corpus (rather than .5 vs .5). Thus, what is needed
are the overall frequencies of the two constructions in the corpus, which corresponds to the
column totals of Table I.
6 In terms of interpretation, the results are identical to an analogous analysis with Medium,
CPrime, VFormID, VLemmaID and SpeakerID and Distance as independent variables and
CID as dependent variable.
374
the target, priming is slightly stronger. In other words, production-toproduction priming is stronger than comprehension-to-production priming.
Finally, let us turn to Distance. Obviously, the distance between prime
and target is irrelevant to the constructional choice as operationalized by
CTarget. While this is not surprising, the analogous analysis with CID as
dependent variable also results in no signicant effect (F(1, 2, 984) = .411,
p = .521). Does this mean that Distance does not have any inuence on
the strength of the priming effect (as measured by the percentage of cases
where CPrime equals CTarget)? And if so, would this not invalidate the
corpus-based analysis completely (since no effect of Distance would imply
the relatively unlikely situation that priming is equally likely across all distances)? A second, closer look at the data, however, shows that this is not
so because one has to bear in mind that Distance was entered into a linear model, but that the relation between the distance between prime and
target on the one hand and the strength of the priming effect on the other
hand need not be linear. In fact, there is evidence that this relation is logarithmic (cf. Gries (in press) for empirical evidence on the basis of Pickering
and Branigans [1998] conditional probabilities measure). Thus, while the
linear relation between Distance and the strength of the priming effect is
negligible, the logarithmic one is not (adjusted R 2 = .77; R1,13 = 44.08;
P < .001).7 priming is in fact long-lasting (again in accordance with recent
ndings by Bock and Grifn, 2000) and after a decrease from Distance:0 to
Distance:1, there was no consistent decline in the magnitude of priming,
although there were unstable changes at particular lags (i.e., parse units in
the present study) (Bock & Grifn, 2000: 187).
In sum, not only has the corpus-based analysis of syntactic priming revealed signicant priming effects for ditransitives and prepositional
datives, the results are also strikingly similar to those of previous experimental studies in terms of strength of effects, the inuence of morphological characteristics of the verbs, construction-specicity, directionality and
distance effects (i.e. the time course of priming). The following section will
now provide a more detailed picture of how individual verbs gure in the
priming effects.
Verb-specic Investigation
The above investigation has shown that the corpus-derived results are
quite similar in nature to those obtained experimentally. However, as has
7
For this analysis, the Distance values of 0 and >25 were recoded as 1E-06 and 30 respectively; other values for Distance:0 and Distance:>25 yielded identical results in terms of
explained variance; the resulting equation is (% of CID = 1) = .625 .0175ln(Distance).
375
Data (ICE-GB)
CPrime: V NP PPfor/to
CPrime: V NP/S NP/S
Column totals
CTarget:
V NP PPfor/to
CTarget:
V NP/S NP/S
a
c
a+c
b
d
b+d
Row
totals
a+b
c+d
a+b+c+d
already been argued above, the results discussed in the previous section (and
the results of any other study on syntactic priming I am aware of) did not
take into consideration the degree to which priming effects might be sensitive
to particular verbs individual preferences. This is all the more striking since
(i) at least Potter and Lombardi (1998: 278) mention, but do not investigate,
that individual verbs may affect preferences for particular syntactic patterns
and (ii) probabilistic (i.e. frequency-based) properties of words, word senses
and words in particular constructions have proven to be relevant to a variety
of linguistic and psycholinguistic issues and models. Thus, what is necessary
is a rst exploratory study of this issue. Such an exploratory study of this
issue using experiments would be quite an enormous enterprise: Since it is
unclear which verbs to start with in the rst place, one would have to use
such a large number of different stimuli (and llers and subjects, etc.) that
this seems a daunting task. A more attractive alternative is a corpus-based
approach (cf. Branigan et al.s statement on how of corpus approaches can
be useful for hypothesis generation, which was quoted above in the introduction), where part of the analysis can be (semi-)automated. In order to
look at this in more detail, consider Table IV for an abstraction from the
study of the dative alternation.
The null hypothesis (H0 ) that has apparently been assumed in experimental studies on syntactic priming is that (the strengths of) the priming
effects are independent of the verb(s). More technically, in the experimental
paradigms referred to above it was argued that, on the whole, observed a
and observed d (henceforth aobs and dobs ) should be higher than expected
a and d (henceforth aexp and dexp ) respectively across all verbs (the same
argument was put forward in the preceding section on the corpus-based
approach), and the implicit assumption seems the be that this is also true
for each individual verb. But rather than take H0 for granted, let us look at
whether this hypothesis is actually borne out by the data. To that end, let us
look at one stimulus set of one particularly interesting study, namely Pickering and Branigan (1998). Their experimental items (Pickering and Branigan,
1998: Exp. 1) involve the ten verbs listed in alphabetical order in (7), which
376
Data (ICE-GB)
CPrime: V NP PPfor/to
CPrime: V NP/S NP/S
Column totals
CTarget:
V NP PPfor/to
CTarget:
V NP/S NP/S
Row
totals
85 (126.2)
51 (183.5)
136
184 (142.8)
340 (207.5)
524
269
391
660
also occur differently frequently as part of the priming or the target structure in the stimulus set; for example, offer is used as a priming verb in both
ditransitive and prepositional dative priming contexts, but not as a target
verb.
(7)
give, hand, lend, loan, offer, post, sell, send, show, throw
If we look at the results for the rst of these verbs in the present data
set, we obtain the results in Table V, where the expected values of each
row areas abovenot computed on the basis of the column totals, but
on the basis of the overall frequencies of the two constructions in Table I.
Table V indicates that there are 660 occurrences of prime-target pairs
with the verb lemma give in the target position (where it should be subject to
syntactic priming). In 269 of these 660 cases, the prime had a prepositional
dative structure, in the remaining 391 cases it had a ditransitive structure.
Although we have seen a strong priming effect for the dative alternation
across all verbs (cf. Table III), the results for give do not reect this overall tendency. As is obvious, the results are not exactly as predicted by H0 :
While dobs is larger than dexp (indicating syntactic priming of the ditransitive construction), no such effect is found for the prepositional dativeby
contrast, the ditransitive is preferred even if the prime is a prepositional
dative: bobs is higher than bexp . Interestingly, an analogous examination of
the second verb listed in (7), hand, results in a completely different distribution. Consider Table VI for the results concerning this verb. As is clear, in
this case, aobs is higher than aexp (reecting a priming effect for the prepositional dative), but no such effect is obtained for the ditransitiverather
cobs is higher than cexp .
It is only the third verb listed in (7) which appears to behave in
accordance with H0 , namely lend. Consider Table VII, where, at last, the
priming effect is balanced: aobs and dobs are higher than aexp and dexp
respectively.
Interestingly, these are not isolated patterns. Out of the 10 verbs listed
in (7), seven occur in both constructions in the corpus (loan and throw
377
Table VI. CPrime CTarget: Observed vs. Expected Frequencies for hand
Data (ICE-GB)
CPrime: V NP PPfor/to
CPrime: V NP/S NP/S
Column totals
CTarget:
V NP PPfor/to
CTarget:
V NP/S NP/S
Row
totals
7 (3.8)
9 (5.6)
16
1 (4.2)
3 (6.4)
4
8
12
20
Table VII. CPrime CTarget: Observed vs. Expected Frequencies for lend
Data (ICE-GB)
CPrime: V NP PPfor/to
CPrime: V NP/S NP/S
Column totals
CTarget:
V NP PPfor/to
CTarget:
V NP/S NP/S
Row
totals
10 (5.6)
2 (7.5)
12
2 (6.4)
14 (8.5)
16
12
16
28
only occur in the ditransitive, post occurs only in the prepositional dative).
If we look at these seven verbs, we nd that
show and offer pattern like give, which prefers the ditransitive construction, (cf. Tables (A.2) and (A.3) in Appendix A for details);
sell patterns like hand, a verb preferring the prepositional dative
(cf. Table (A.4) in Appendix A for the exact gures); and
send patterns roughly like lend (and, thus, as predicted by H0 )
(roughly because, for send, dobs is only about the same as dexp ).
While it is important to note that this specic nding does not invalidate the general priming effect, some verbs appear more likely to resist
priming. It seems as if they preferred to occur in one construction and the
question arises as to how to motivate this discrepancy. One possible explanation for these ndings is based on recent general research on subcategorization preferences of verbs (and verb senses).
Most previous approaches to subcategorization preferences just quantify the attraction of some word W to some construction C in terms of the
raw frequency of W in C (examples include Connine et al., 1984, Hunston
& Francis, 2000, Lapata et al., 2001; and Hare et al., 2003, to name
but a few). In a series of publications, Stefanowitsch and Gries developed a family of techniques for quantifying the strengths of association
between words and particular (slots of) constructions, the so-called collostructional analysis (cf. for details Stefanowitsch & Gries, 2003; Gries
& Stefanowitsch, 2004a, b). These techniques make it possible to identify
378
Table VIII. The Distribution of give in the Ditransitive and the Prepositional Dative with
to in the ICE-GB (pFisherexact=1.84E120 )
Data (ICE-GB)
give
other verbs
Row totals
ditransitive
construction
prepositional dative
construction (with to)
Column totals
461
(213)
146
(394)
607
574
(822)
1773
(1525)
2347
1035
1919
2954
what they refer to as signicant collexemes, i.e., the verbs that are most
strongly attracted by the V slot of the, say, ditransitive construction, the
passive construction, the verb-particle construction, etc.; elds of application of collostructional analysis include, but are not limited to, the
identication and (more precise) measurement of subcategorization preferences, the investigation of semantic properties of constructions and their
implications for acquisition, etc. In contrast to the previous raw-frequency
approaches just mentioned, collostructional analysis also takes into consideration the overall frequencies of W and C in the corpus to determine whether the distribution of W in the relevant slot of C deviates from
the one already expected by chance alone, a precaution that many of the
above studies have failed to take.8 The method most relevant to our purposes is an extension of the investigation of distinctive collocates called
distinctive collexeme analysis.9 It requires to rst identify how a verb is
distributed across two alternative constructions, as is represented in Table
VIII; gures in parentheses are again expected frequencies.
For every such table (one for each verb), Gries and Stefanowitsch
compute a Fisher exact test to determine to which construction the verb
is more strongly attracted; in the above example, it is immediately obvious
that give is much more strongly attracted to the ditransitive than to the
prepositional dative with to (cf. the ratio of observed to expected frequencies in the upper left cell of Table VIII), and corresponding ndings can
379
be obtained for all verbs occurring at least once in at least one of the two
constructions under investigation.
For the present purposes, we will restrict our attention to the verbs in
(7), and it turns out that they indeed exhibit similar kinds of preferences:
Of the verbs used by Pickering and Branigan (1998), give, show and offer
are signicant collexemes of the ditransitive (all ps < .1E 09) whereas
sell and hand are signicant collexemes of the prepositional datives (all
ps < .01) lend and send do not exhibit a signicant preference for
either construction (p >.13); the computations are based on all 3973 ditransitives and prepositional datives mentioned in Table I.10
Noting the strong distinctive collostruction strengths of Pickering and
Branigans (1998: Exp. 1) stimulus verbs, we can now explain the ndings
of Table VVII by proposing an alternative hypothesis. This alternative to
the above null hypothesis, the collostruction-based hypothesis H1 , is that
a verb strongly associated with a particular construction resists priming
and rather sticks to its associated construction. Again more technically,
for collexemes of the ditransitive, bobs and dobs should be higher than
bexp and dexp , and for collexemes of the prepositional dative, aobs and cobs
should be higher than aexp and cexp . Finally, the verbs without a strong
association to a construction, i.e. where no verb-construction association
would be expected to block the priming effect, should exhibit the distribution predicted by H0 as observed for all verbs together in Table III. As is
obvious, the results obtained from the corpus data are the ones predicted
by the collostruction-based hypothesis H1 .
More specically, there is a majority of (types of) verbs in the corpus which have no strong association to one of the two constructions and
which are thus fully responsive to priming effects;11 since for many of
these the distribution postulated in H0 is found, the picture that emerges is
the overall priming effect of previous studies and of Table III in the present study. On the other hand, a minority of (types of) verbs is strongly
associated with a particular construction and these are therefore more
resistant to priming, which is why their patterns or priming success rates
differ from those of the others. Put differently, the priming rates of these
10
Differences between the present results and those of Gries (in press) or Gries and Stefanowitsch (2004a) are due to the fact that these earlier studies included only prepositional
datives with to and/or instances with nominal objects.
11 Only 86 out of the 316 dative verb types in the ICE-GB (27.2%) have a signicant association to one of the two constructions. A yet more extreme distribution is observed for
the transitive phrasal verbs to be discussed in the following section: Only 40 out of the 700
transitive phrasal verb types in the ICE-GB (5.7%) have a signicant association to one of
the two constructions.
380
PARTICLE PLACEMENT
General Investigation
If syntactic priming can indeed be attributed to the processing of a
particular structure, then it should be manifested in a variety of differ12
One important point must be claried here. To some readers, this approach may seem
somewhat circular; they might object to my line of reasoning by saying, Wait a minute!
You start out by using the corpus data to compute priming effects. And then you use the
very same corpus data to compute collostruction strengths. No wonder you get such a high
correlation of ndings you measure the same thing under two different labels! In the
general discussion below, I will provide evidence why the issue of circularity is unproblematic here.
381
Data (ICE-GB)
V Prt NP
V NP Prt
Row totals
spoken
written
Column totals
698
553
1251
963
229
1192
1661
782
2443
Data (ICE-GB)
CPrime: V Part NP/S
CPrime: V NP/S Part
Column totals
CTarget:
V Part NP/S
CTarget:
V NP/S Part
Row
totals
548 (444)
300 (476.2)
848
319 (423)
630 (453.8)
949
867
930
1797
382
383
For reasons of space, I will not discuss the results for particle placement at the same level of detail as before. As it turns out, the ndings are
quite similar to those for the dative alternation with respect to the verb
specicity of priming. For example, take up is a verb with a strong collostructional attraction to the construction where the verb and the particle are adjacent, and as Table XI illustrates, priming is correspondingly
restricted to this construction; cf. Table (A.6) for analogous results for nd
out.
A similar point can be made for verbs associated with the particlenal construction. A case in point is put in in Table XII, where priming
is restricted to that construction; cf. also Table (A.7) in Appendix A for
the data on take out. Note in this connection that put in and take out
384
Table XI. CPrime CTarget: Observed vs. Expected Frequencies for take up
Data (ICE-GB)
Prime: V Part NP/S
Prime: V NP/S Part
Column totals
Row totals
7 (5.6)
9 (6.1)
16
4 (5.4)
3 (5.9)
7
11
12
23
Table XII. CPrime CTarget: Observed vs. Expected Frequencies for put in
Data (ICE-GB)
Prime: V Part NP/S
Prime: V NP/S Part
Column totals
Row totals
7 (8.2)
7 (14.3)
14
9 (7.8)
21 (13.7)
30
16
28
44
are particularly relevant in this context becauseunlike some other previously discussed verbsthey are distinctive for one construction without
nearly exclusively occurring in that construction: put in occurs in [VP V
Part NP/S] and [VP V NP/S Part ] 21 and 33 times respectively while take
out occurs in [VP V Part NP/S] and [VP V NP/S Part ] 15 and 26 times
respectively. The fact that these verbs priming effects still exhibit the verb
specicity effect illustrates that their priming behavior and collostruction
strength are not automatic reexes of their raw frequencies.
Finally, consider a verb which has absolutely no preference for one
verb-particle construction, namely pick up. As is evident from Table XIII,
priming occurs for both constructions, and the same holds for another
unbiased verb, namely put down (cf. Table (A.8) in Appendix A).
To summarize, we have again obtained a clear priming effect for
both constructions, but also more detailed evidence on the verb-specicity effect: Some verbs association to a verb-particle construction appear
to allow for, or resist, the priming effect much more strongly than others.
Given that the two verb-particle constructions are associated with semantically different groups of verbs (cf. Gries 2003a, Gries & Stefanowitsch
Table XIII. CPrime CTarget: Observed vs. Expected Frequencies for pick up
Data (ICE-GB)
Prime: V Part NP/S
Prime: V NP/S Part
Column totals
Row totals
15 (13.8)
16 (21.5)
31
12 (13.2)
26 (20.5)
38
27
42
69
385
2004a for details), this even invites inferences as to how semantic properties of verbs correlate with the strength of priming effects.
386
Let me rst state something quite explicitly: It is true that, in general, experimental studies are in a better position to single out particular
aspects of priming more easily than corpus-based studies, and the possibility to hold experimental conditions constant across a variety of trials and
(combinations of) conditions should not be underestimated. However, the
exploratory benets of corpus data have been mentioned by Pickering and
Branigan themselves (cf. above), and from a different perspective, the controlled nature of experimental conditions also has some drawbacks.
First, the priming data are usually collected in a very narrowly
dened and articial setting. While this is desirable from the point of view
of delimiting error variance, it does not allow generalizations of the role of
register effects on syntactic primingthe corpus data, by contrast, allow
for a multifactorial analysis of syntactic repetition in natural settings. In
addition, in their discussion of previous experimental approaches to priming, Hartsuiker and Kolk (1998: 148) criticize much previous work for
not taking into consideration the overall frequencies of syntactic constructions, whichif not considered properlymay introduce frequency effects
into the priming results. In the present approach, the corpus data allow
for a natural computation of construction baseline frequencies. Second,
not all experimental studies managed to account for all potential explanatory factors. For example, Bock and Loebells (1990) ndings were interpreted as evidence for the irrelevance of thematic utterance characteristics
and that function words were irrelevant to priming until Hare and Goldbergs (1999) and Bencini, et al.s, (2002) replications showed that this was
not necessarily the case. Of course, this does not invalidate the experimental approach as such, but it points out that the number of factors to be
taken into consideration is so high that it is not always possible to hold
them all constant. Thus, including such confounding factors into a corpus-based evaluation may sometimes be a useful alternative. Finally, by
investigating syntactic priming from a corpus-based perspective, one can
determine to what degree it plays a role for grammatical variation, i.e.
the phenomenon that in a given discourse situation the speaker may have
the choice between two truth-conditionally equivalent, nearly synonymous
constructions (e.g. between the two dative constructions, active vs. passive, or the of-genitive vs. the s-genitive, etc.). Including the priming effects
into the research design may make it possible to increase the accuracy of
predicting the construction the speaker will choose (subconsciously); cf.
Gries (2003a, in press) on verb-particle constructions and Rohdenburg and
Mondorf (2003) for a more general perspective.
Apart from these general methodological arguments, some other more
specic comments on Pickering and Branigans nonsyntactic factors are due
because not all of these lend themselves to an explanation of the present
387
results (and Pickering and Branigan have not provided empirical evidence
for their claims). For example, the fact that one of the two constructions
may be predominant in a particular register is taken into account here since
(a) the corpus data cover a wide variety of registers and (b) the medium
(speaking vs. writing) was included into the analysis. Note also that neither
alternation investigated is inherently related to a particular level of formality, and explaining the frequent cases of syntactic priming by hundreds of
sudden register/formality changes does not seem very plausible.
Similarly, the effects cannot be straightforwardly reduced to, say, the
givenness or semantic characteristics of the direct objects referent: First,
both datives have information structure properties (cf. Thompson, 1990),
so why should only one result in priming in corpora? Second, one might
suggest that the slight priming prominence of the verb-particle construction with a VP-nal particle is due to the fact that this construction is
associated with a given referent of the direct object (cf. Gries, 2003a, Sect.
6.1.4: Once the referent of the direct object has been introduced, the verbparticle constructions in the subsequent discourse will place it before the
particle. However, Gries (2003a, 120121, 131) found priming effects for
VPCs regardless of whether the referent of the direct object NP in the second construction is coreferential with that of the rst. Third, the kind of
animacy/argument effects that might in principle affect datives (such that
animacy affects constructional choices) cannot explain the results on VPCs
where animacy plays no role (cf. Gries, 2003a, 8889]) and the particle
is often aspectual or idiomatically used and can, thus, not be attributed
argument status.
An additional important point is that other non-syntactic factors can
also not be held responsible for the present ndings. For example, those
who would like to attribute the present results to lexical repetition effects
would have to explain why, in the case of dative alternation, it is the
ditransitive construction that primes more strongly although (i) it is the
prepositional dative which allows for the priming of the function words
to and for and (ii) the fact that lexical activation decays too fast makes
it unlikely that the long duration of priming effects observed here and in
other (experimental studies) is just a lexical memory effect.
In sum, much of the present ndings resembles those obtained experimentally so strongly that they cannot be explained away as easily as suggested. While I do not rule out discourse-motivated factors of priming at
all, it is hard to explain all the similarities between the different kinds of
results and still simply uphold the claim that all this is epiphenomenal.
Without doubt, further experimental evidence is necessary, but it seems as
if the utility of corpus-based, explorative results should not be underestimated prematurely.
388
Data (ICE-GB)
CTarget: C1
CTarget: C2
Row totals
CPrime: C1
CPrime: C2
Column totals
12 (8)
0 (2)
12
4 (8)
4 (2)
8
16
4
20
389
Gries, et al. (in press a, b), for example, computed the collostruction
strengths of verbs to the as-predicative (e.g. I regard her as clever) in the
ICE-GB and then, for a later study, of the BNC sampler, a subpart of the
British National Corpus consisting of 2m words of (spoken and written)
English. As it turns out, when the signicant collexemes of the as-predicative are sorted in each corpus, then the signicant collexemes of the as-predicative in one corpus are not only also a signicant collexeme in the other
corpus, but they also tend to occupy similar ranks in the other corpus. For
example, 60% of the 30 most signicant collexemes in the ICE-GB are also
among the 30 most signicant collexemes in the BNC sampler. What is
more, the correlation between the ranks of all verbs occurring in both corpora is highly signicant ( = .535; z = 6.97; p < .001) as is that between
the ranks of all signicant collexemes ( = .591; z = 5.29; p < .001). In
other words, given this high correspondence between different corpora, it is
very likely that computing verb-specicity from another corpus would not
have changed the picture markedly. Once other manually parsed corpora are
available, checking the present results will be easy.
Finally, there is also experimental evidence supporting the verb-specicity effect argued for here. Gries and Wulff (in press) replicated Pickering and
Branigans (1998) experiments on syntactic priming in English with native
speakers of German to determine whether syntactic priming is also obtained
with advanced learners of a foreign language (recall point (v) in the Introduction). In addition to a general priming effect, they found that, just like
in the present study, the strength of the priming effect of the seven dativealternation verbs discussed above is strongly correlated with a general bias
of the subjects to use the experimental verbs in particular constructions.
In Figure 1, their results are summarized: the y-axis portrays the bias of
individual verbs to either the ditransitive or the prepositional dative in the
corpus data, basically as measured by collostruction strength. The x-axis
portrays the preference of individual verbs to be completed using either
the ditransitive or the prepositional dative in the sentence-completion task.
Finally, the strong correlation (r 2 = .8; t (5) = 4.47; p = .007) is indicated
by the slope of the regression line, which shows that one can predict the
outcome in the priming experiment on the basis of the verbs preferences as
measured on the basis of the corpus data.
Even though these results were not obtained with native speakers,
they do point to the fact that experimentally primed sentence completion
is strongly sensitive to verb bias. Thus, I submit, this issue is clearly in
need of further research of which corpus linguistic methods may play an
essential role in determining collostruction strengths for verbs to be tested
(cf., e.g., Gries and Stefanowitsch, 2004a on verbs distinctive for actives
and passives).
390
Fig. 1. The verbs constructional preferences in the corpus data and in thepriming-experiment responses.
391
well integrated into Pickering and Branigans model. Remember, for example, that not only priming was found, the priming effect was also stronger when the verb lemmas were identical. In addition, there was also
a tendency for identical verb forms to result in a stronger priming
effect, and while that was not hypothesized in the above model, similar
tendencies were obtained in Pickering and Branigans (1998: Exp. 5) on
singular-vs.-plural form differences. Also, since the model has been argued
to involve a shared representation in comprehension and production, the
fact that SpeakerID had no strong effect in this study can be explained
naturally. Finally, the fact that the verb-particle constructions exhibit priming supports the idea that order information is encoded within combinatorial nodes.
Given this kind of psycholinguistic model, the second kind of nding of this study, the verb-specicity of priming, can be integrated straightforwardly. Recall that each verb lemma is connected to the combinatorial
nodes of the construction in which the verb can be used. Since syntactic
priming of a construction C involves the repeated activation of Cs combinatorial node (so that its resting level is exceeded), it follows naturally
that when the link between a verb and C is stronger, priming of (only) that
construction should be stronger. This is exactly what we nd: the verbs
which are strongly associated with one construction exhibit priming with
this construction much more strongly than with the other construction.
Thus, we only need to supplement Pickering and Branigans model with
the notion that the links between verb lemmas and combinatorial nodes
they postulated anyway can also be differentially strong to reect their
degree of attraction/repulsion to a construction as measured by collostruction strength. This would allow for the model to accommodate the present
ndings on verb-specicity, but also allows for an economical representation of many of the ndings concerning verb subcategorization preferences, verb bias etc. mentioned above.13 Given the current interest in
the issue of whether syntactic priming is best explained as activation patterns or implicit learning (cf. Chang et al., 2000, 2003), it is even conceivable that the network architectures used to test these different conceptions
could be somehow enriched with the collostructional information.
All in all, the present ndings demonstrate how usefulin spite of
some limitationscorpus-based approaches to priming phenomena can be
to support and extend ndings obtained with other methodologies, promoting once more the ideal of converging evidence.
13
392
APPENDIX A
Table A.1. GLM Results for the Priming of Datives
Effect source
CPrime
VLemmaID CPrime
VFormID CPrime
SpeakerID CPrime
VLemmaID
SpeakerID
Medium
Distance
Medium VFormID CPrime
Medium VLemmaID CPrime
Medium VLemmaIDa
Medium CPrime
VFormID SpeakerID
VLemmaIDa SpeakerID
VFormID
Medium VFormID
VFormID SpeakerID CPrime
VLemmaID SpeakerID CPrime
161.671
31.946
4.429
3.645
2.509
2.346
1.935
1.867
.614
.267
.198
.101
.099
.084
.043
.039
.015
0
<.0001
<.0001
.0354
.0563
.1133
.1257
.1643
.1719
.4333
.6055
.6568
.7511
.7533
.7724
.8354
.8441
.9011
.9983
Partial 2
.0514
.0106
.0015
.0012
.0008
.0008
.0006
.0006
.0002
.0001
.0001
0
0
0
0
0
0
0
Table A.2. CPrime CTarget: Observed vs. Expected Frequencies for show
Data (ICE-GB)
CPrime: V NP PPf or/to
CPrime: V NP/S NP/S
Column totals
CTarget:
V NP PPfor/to
CTarget:
V NP/S NP/S
Row
totals
5 (12.2)
7 (28.6)
12
21 (13.8)
54 (32.4)
75
26
61
87
Table A.3. CPrime CTarget: Observed vs. Expected Frequencies for offer
Data (ICE-GB)
CPrime: V NP PPfor/to
CPrime: V NP/S NP/S
Column totals
CTarget:
V NP PPfor/to
CTarget:
V NP/S NP/S
Row
totals
6 (9.4)
4 (18.3)
10
14 (10.6)
35 (20.7)
49
20
39
59
393
Table A.4. CPrime CTarget: Observed vs. Expected Frequencies for sell
Data (ICE-GB)
CTarget:
V NP PPfor/to
CTarget:
V NP/S NP/S
Row
totals
4 (1.9)
5 (2.8)
9
0 (2.1)
1 (3.2)
1
4
6
10
Effect source
CPrime
Medium
Medium VLemmaID CPrime
VFormID SpeakerID CPrime
PartID
VLemmaID CPrime
VFormID CPrime
PartID SpeakerID
VFormID SpeakerID
VFormID
Medium VLemmaID
PartID CPrime
Medium VFormID CPrime
Medium VFormID
Medium CPrime
SpeakerID
PartID SpeakerID CPrime
Distance
Medium PartID
Medium PartID CPrime
SpeakerID CPrime
VLemmaID
25.451
5.586
5.094
4.192
2.757
2.516
2.083
1.234
1.151
1.008
.983
.722
.506
.449
.436
.435
.203
.189
.052
.045
.004
.003
<.0001
.0182
.0241
.0408
.097
.1129
.1491
.2668
.2836
.3155
.3216
.3955
.4769
.5031
.509
.5096
.6523
.6639
.8193
.8314
.9473
.9541
Partial 2
.0142
.0032
.0029
.0024
.0016
.0014
.0012
.0007
.0007
.0006
.0006
.0004
.0003
.0003
.0002
.0002
.0001
.0001
0
0
0
0
Table A.6. CPrime CTarget: Observed vs. Expected Frequencies for nd out
Data (ICE-GB)
CPrime: V Part NP/S
CPrime: V NP/S Part
Column totals
CTarget:
V Part NP/S
CTarget:
V NP/S Part
Row
totals
22 (12.8)
13 (7.7)
35
3 (12.2)
2 (7.3)
5
25
15
40
394
Table A.7. CPrime CTarget: Observed vs. Expected Frequencies for take out
Data (ICE-GB)
CPrime: V Part NP/S
CPrime: V NP/S Part
Column totals
CTarget:
V Part NP/S
CTarget:
V NP/S Part
Row
totals
7 (7.7)
5 (9.7)
12
8 (7.3)
14 (9.3)
22
15
19
34
Table A.8. CPrime CTarget: Observed vs. Expected Frequencies for put down
Data (ICE-GB)
CPrime: V Part NP/S
CPrime: V NP/S Part
Column totals
CTarget:
V Part NP/S
CTarget:
V NP/S Part
Row
totals
8 (5.1)
5 (8.2)
13
2 (4.9)
11 (7.8)
13
10
16
26
395
Fig. 2. Relative frequencies of ditransitives and prepositional datives plotted against their
switch rates per corpus le.
396
Fig. 3. Relative frequencies of V Part DO and V DO Part plotted against their switch rates
per corpus le.
397
REFERENCES
Bencini, G., Bock, J. K., & Goldberg, A. E. (2002). How abstract is grammar? Evidence from
structural priming in language production. Paper presented at the 15th CUNY Conference on Human Sentence Processing.
Bock, J. K. (1986). Syntactic persistence in language production. Cognitive Psychology 18,
355387.
Bock, J. K. (1989). Closed-class immanence in sentence production. Cognition 31(2), 163186.
Bock, J. K. (1990). Structure in language: Creating form in talk. American Psychologist
45(11), 12211236.
Bock, J. K. (2002). Persistent structural priming from comprehension to production. Paper
presented at the 15th CUNY Conference on Human Sentence Processing.
Bock, J. K., & Loebell, H. (1990). Framing sentences. Cognition 35(1), 139.
Bock, J. K., Loebell, H., & Morey, R. (1992). From conceptual roles to structural relations:
Bridging the syntactic cleft. Psychological Review 99(1), 150171.
Bock, J. K., & Grifn, Z. M. (2000). The persistence of structural priming: Transient activation or implicit learning. Journal of Experimental Psychology: General 129(2), 177192.
Branigan, H. P., Pickering, M. J., & Cleland, A. A. (1999). Syntactic priming in written production: Evidence for rapid decay. Psychonomic Bulletin and Review 6(4), 635640.
Branigan, H. P., Pickering, M. J., & Cleland, A. A. (2000). Syntactic co-ordination in dialogue. Cognition 75(2), B13B25.
Branigan, H. P., Pickering, M. J., Liversedge, S. P., Stewart, A. J., & Urbach, T. P. (1995).
Syntactic priming: Investigating the mental representation of language. Journal of Psycholinguistic Research 24(6), 489506.
Branigan, H. P., Pickering, M. J., Stewart, A. J., & McLean, J. F. (2000). Syntactic priming in
spoken production: Linguistic and temporal interference. Memory and Cognition 28(8),
12971302.
Chang, F., Dell, G. S., Bock, J. K., & Grifn, Z. M. (2000). Structural priming as implicit
learning: A comparison of models of sentence production. Journal of Psycholinguistic
Research 29(2), 217229.
Chang, F., Dell, G. S. & Bock, J. K. (2003). Comprehending structural priming: A connectionist learning account. Paper presented at the 16th CUNY Conference on Human Sentence Processing.
Church, K. W., Gale, W., Hanks, P., & Hindle, D. (1991). Using statistics in lexical analysis.
In Zernik Uri, (ed.). Lexical acquisition: Exploiting On-line Resources to Build up a Lexicon. Hillsdale, NJ: Lawrence Erlbaum, pp. 115164.
Connine, C., Ferreira, F., Jones, C., Clifton, C., & Frazier, L. (1984). Verb frame preferences:
Descriptive norms. Journal of Psycholinguistic Research 13(4), 307319.
Corley, M. & Scheepers, C. (2002). Syntactic priming in English sentence production: Categorical and latency evidence from an internet-based study. Psychonomic Bulletin and
Review 9(1), 126131.
Estival, D. (1985). Syntactic priming of the passive in English. Text 5(1), 722.
Garnsey, S. M., Pearlmutter, N. J., Myers, E., & Lotocky, M. A. (1997). The contributions
of verb bias and plausibility to the comprehension of temporarily ambiguous sentences.
Journal of Memory and Language 37(1), 5893.
Gries, St. Th. (2003a). Multifactorial Analysis in Corpus Linguistics: The Case of Particle
Placement. London/New York: Continuum Press.
Gries, St. Th. (2003b). Testing the sub-test: A collocational-overlap analysis of English -ic
and -ical adjectives. International Journal of Corpus Linguistics 8(1), 3161.
398
Gries, St. Th., (in press). New perspectives on old alternations. In Cihlar, J. E., L. Franklin,
A., Kaiser, D. W., & Kimbara, I. (eds.). Papers from the 39th Regional Meeting of the
Chicago Linguistics Society: Vol. II. The Panels. Chicago, IL: Chicago Linguistics Society.
Gries, St. Th., & Stefanowitsch, A. (2004a). Extending collostructional analysis: A corpusbased perspective on alternations. International Journal of Corpus Linguistics 9(1), 97
129.
Gries, St. Th., & Stefanowitsch, A. (2004b). Co-varying collexemes in the into-causative. In
Achard, M. & Kemmer, S., (eds.). Language, Culture, and Mind (pp. 225236) Stanford,
CA: CSLI.
399