Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 121–124,
New York, June 2006. c©2006 Association for Computational Linguistics
Computational Modelling of Structural Priming in Dialogue
David Reitter, Frank Keller, Johanna D. Moore
dreitter | keller | jmoore @ inf.ed.ac.uk
School of Informatics
University of Edinburgh
United Kingdom
Abstract
Syntactic priming effects, modelled as in-
crease in repetition probability shortly af-
ter a use of a syntactic rule, have the
potential to improve language processing
components. We model priming of syn-
tactic rules in annotated corpora of spo-
ken dialogue, extending previous work
that was confined to selected construc-
tions. We find that speakers are more re-
ceptive to priming from their interlocutor
in task-oriented dialogue than in spona-
neous conversation. Low-frequency rules
are more likely to show priming.
1 Introduction
Current dialogue systems overlook an interesting
fact of language-based communication. Speakers
tend to repeat their linguistic decisions rather than
making them from scratch, creating entrainment
over time. Repetition is evident not just on the ob-
vious lexical level: syntactic choices depend on pre-
ceding ones in a way that can be modelled and, ul-
timately, be leveraged in parsing and language gen-
eration. The statistical analysis in this paper aims to
make headway towards such a model.
Recently, priming phenomena1 have been ex-
ploited to aid automated processing, for instance in
automatic speech recognition using cache models,
but only recently have attempts been made at using
1The term priming refers to a process that influences lin-
guistic decision-making. An instance of priming occurs when a
syntactic structure or lexical item giving evidence of a linguistic
choice (prime) influences the recipient to make the same deci-
sion, i.e. re-use the structure, at a later choice-point (target).
them in parsing (Charniak and Johnson, 2005). In
natural language generation, repetition can be used
to increase the alignment of human and computers.
A surface-level approach is possible by biasing the
n-gram language model used to select the output
string from a variety of possible utterances (Brock-
mann et al., 2005).
Priming effects are common and well known. For
instance, speakers access lexical items more quickly
after a semantically or phonologically similar prime.
Recent work demonstrates large effects for partic-
ular synonymous alternations (e.g., active vs. pas-
sive voice) using traditional laboratory experiments
with human subjects (Bock, 1986; Branigan et al.,
2000). In this study, we look at the effect from a
computational perspective, that is, we assume some
form of parsing and syntax-driven generation com-
ponents. While previous studies singled out syntac-
tic phenomena, we assume a phrase-structure gram-
mar where all syntactic rules may receive priming.
We use large-scale corpora, which reflect the reali-
ties of natural interaction, where limited control ex-
ists over syntax and the semantics of the utterances.
Thus, we quantify priming for the general case in
the realistic setting provided by corpus based exper-
iments. As a first hypothesis, we predict that after a a
syntactic rule occurs, it is more likely to be repeated
shortly than a long time afterwards.
From a theoretical perspective, priming opens a
peephole into the architecture of the human lan-
guage faculty. By identifying units in which prim-
ing occurs, we can pinpoint the structures used in
processing. Also, priming may help explain the ease
with which humans engange in conversations.
This study is interested in the differences relevant
to systems implementing language-based human-
121
computer interaction. Often, HCI is a means for
user and system to jointly plan or carry out a task.
Thus, we look at repetition effects in task-oriented
dialogue. A recent psychological perspective mod-
els Interactive Alignment between speakers (Picker-
ing and Garrod, 2004), where mutual understand-
ing about task and situation depends on lower-level
priming effects. Under the model, we expect prim-
ing effects to be stronger when a task requires high-
level alignment of situation models.
2 Method
2.1 Dialogue types
We examined two corpora. Switchboard con-
tains 80,000 utterances of spontaneous spoken con-
versations over the telephone among randomly
paired, North American speakers, syntactically an-
notated with phrase-structure grammar (Marcus
et al., 1994). The HCRC Map Task corpus comprises
more than 110 dialogues with a total of 20,400 ut-
terances (Anderson et al., 1991). Like Switchboard,
HCRC Map Task is a corpus of spoken, two-person
dialogue in English. However, Map Task contains
task-oriented dialogue: interlocutors work together
to achieve a task as quickly and efficiently as pos-
sible. Subjects were asked to give each other direc-
tions with the help of a map. The interlocutors are in
the same room, but have separate, slightly different
maps and are unable to see each other’s maps.
2.2 Syntactic repetitions
Both corpora are annotated with phrase structure
trees. Each tree was converted into the set of phrase
structure productions that license it. This allows us
to identify the repeated use of rules. Structural prim-
ing would predict that a rule (target) occurs more
often shortly after a potential prime of the same rule
than long afterwards – any repetition at great dis-
tance is seen as coincidental. Therefore, we can cor-
relate the probability of repetition with the elapsed
time (DIST) between prime and target.
We considered very pair of two equal syntactic
rules up to a predefined maximal distance to be a
potential case of priming-enhanced production. If
we consider priming at distances 1...n, each rule
instance produces up to n data points. Our binary
response variable indicates whether there is a prime
for the target between n−0.5 and n + 0.5 seconds
before the target. As a prime, we see the invocation
of the same rule. Syntactic repetitions resulting from
lexical repetition and repetitions of unary rules are
excluded. We looked for repetitions within windows
(DIST) of n = 15 seconds (Section 3.1).
Without priming, one would expect that there is a
constant probability of syntactic repetition, no mat-
ter the distance between prime and target. The anal-
ysis tries to reject this null hypothesis and show a
correlation of the effect size with the type of corpus
used. We expect to see the syntactic priming effect
found experimentally should translate to more cases
for shorter repetition distances, since priming effects
usually decay rapidly (Branigan et al., 1999).
The target utterance is included as a random fac-
tor in our model, grouping all 15 measurements of
all rules of an utterance as repeated measurements,
since they depend on the same target rule occurrence
or at least on other other rules in the utterance, and
are, thus, partially inter-dependent.
We distinguish production-production priming
within (PP) and comprehension-production priming
between speakers (CP), encoded in the factor ROLE.
Models were estimated on joint data sets derived
from both corpora, with a factor SOURCE included
to discriminate the two dialogue types.
Additionally, we build a model estimating the ef-
fect of the raw frequency of a particular syntactic
rule on the priming effect (FREQ). This is of par-
ticular interest for priming in applications, where a
statistical model will, all other things equal, prefer
the more frequent linguistic choice; recall for com-
peting low-frequency rules will be low.
2.3 Generalized Linear Mixed Effect
Regression
In this study, we built generalized linear mixed ef-
fects regression models (GLMM). In all cases, a rule
instance target is counted as a repetition at distance
d iff there is an utterance prime which contains the
same rule, and prime and target are d units apart.
GLMMs with a logit-link function are a form of lo-
gistic regression.2
2We trained our models using Penalized Quasi-Likelihood
(Venables and Ripley, 2002). We will not generally give classi-
calR2 figures, as this metric is not appropriate to such GLMMs.
The below experiments were conducted on a sample of 250,000
122
SWBD PP MT PPMT CP
!
0.10
!
0.05
0.00
0.05
0.10
0.15
0.20
Switchboard Map Task
PP PP CPCP
*
*
*
-
-
-
-
0 5 10 15
0.010
0.012
0.014
0.016
0.018
0.020
distance: Temporal Distance between prime and target (seconds)
p(prime=target|target,distance)
Map Task: 
production-production
Switchboard: 
production-production
Map Task: 
comprehension-production
Switchboard: 
comprehension-production
Figure 1: Left: Estimated priming strength (repetition probability decay rate) for Switchboard and Map
Task, for within-speaker (PP) and between-speaker (CP) priming. Right: Fitted model for the development
of repetition probability (y axis) over time (x axis, in seconds). Here, decay (slope) is the relevant factor for
priming strength, as shown on the left. These are derived from models without FREQ.
Regression allows us not only to show that prim-
ing exists, but it allows us to predict the decline of
repetition probability with increasing distance be-
tween prime and target and depending on other vari-
ables. If we see priming as a form of pre-activation
of syntactic nodes, it indicates the decay rate of pre-
activation. Our method quantifies priming and cor-
relates the effect with secondary factors.
3 Results
3.1 Task-oriented and spontaneous dialogue
Structural repetition between speakers occured in
both corpora and its probability decreases logarith-
mically with the distance between prime and target.
Figure 1 provides the model for the influence
of the four factorial combinations of ROLE and
SOURCE on priming (left) and the development of
repetition probability at increasing distance (right).
SOURCE=Map Task has an interaction effect on the
priming decay ln(DIST), both for PP priming (β =
−0.024,t = −2.0, p < 0.05) and for CP priming
(β = −0.059,t = −4.0, p < 0.0005). (Lower coef-
ficients indicate more decay, hence more priming.)
data points per corpus.
In both corpora, we find positive priming effects.
However, PP priming is stronger, and CP priming is
much stronger in Map Task.
The choice of corpus exhibits a marked interac-
tion with priming effect. Spontaneous conversation
shows significantly less priming than task-oriented
dialogue. We believe this is not a side-effect of vary-
ing grammar size or a different syntactic entropy in
the two types of dialogue, since we examine the de-
cay of repetition probability with increasing distance
(interactions with DIST), and not the overall proba-
bility of chance repetition (intercepts / main effects
except DIST).
3.2 Frequency effects
An additional model was built which included
ln(FREQ) as a predictor that may interact with the
effect coefficient for ln(DIST).
ln(FREQ) is inversely correlated with
the priming effect (Paraphrase: βlnDist =
−1.05,βlnDist:lnFreq = 0.54, Map Task:
βlnDist = −2.18,βlnDist:lnFreq = 0.35, all
p < 0.001). Priming weakens with higher
(logarithmic) frequency of a syntactic rule.
123
4 Discussion
Evidence from Wizard-of-Oz experiments (with sys-
tems simulated by human operators) have shown
that users of dialogue systems strongly align their
syntax with that of a (simulated) computer (Brani-
gan et al., 2003). Such an effect can be leveraged
in an application, provided there is a priming model
interfacing syntactic processing.
We found evidence of priming in general, that is,
when we assume priming of each phrase structure
rule. The priming effects decay quickly and non-
linearly, which means that a dialogue system would
best only take a relatively short preceding context
into account, e.g., the previous few utterances.
An important consideration in the context of di-
alogue systems is whether user and system collab-
orate on solving a task, such as booking tickets or
retrieving information. Here, syntactic priming be-
tween human speakers is strong, so a system should
implement it. In other situations, systems do not
have to use a unified syntactic architecture for pars-
ing and generation, but bias their output on previous
system utterances, and possibly improve parsing by
looking at previously recognized inputs.
The fact that priming is more pronounced within
(PP) a speaker suggests that optimizing parsing and
generation separately is the most promising avenue
in either type of dialogue system.
One explanation for this lies in a reduced cog-
nitive load of spontaneous, everyday conversation.
Consequently, the more accessible, highly-frequent
rules prime less.
In task-oriented dialogue, speakers need to pro-
duce a common situation model. Interactive Align-
ment Model argues that this process is aided by syn-
tactic priming. In support of this model, we find
more priming in task-oriented dialogue.3
5 Conclusions
Syntactic priming effects are reliably present in di-
alogue even in computational models where the full
range of syntactic rules is considered instead of se-
lected constructions with known strong priming.
This is good news for dialogue systems, which
tend to be task-oriented. Linguistically motivated
3For a more detailed analysis from the perspective of inter-
active alignment, see Reitter et al. (2006).
systems can possibly exploit the user’s tendency to
repeat syntactic structures by anticipating repetition.
Future systems may also align their output with their
recognition capabilities and actively align with the
user to signal understanding. Parsers and realizers in
natural language generation modules may make the
most of priming if they respect important factors that
influence priming effects, such as task-orientation of
the dialogue and frequency of the syntactic rule.
Acknowledgements
The authors would like to thank Amit Dubey, Roger Levy and
Martin Pickering. The first author’s work is supported by a grant
from the Edinburgh Stanford Link.
References
A. Anderson, M. Bader, E. Bard, E. Boyle, G. M. Doherty,
S. Garrod, S. Isard, J. Kowtko, J. McAllister, J. Miller,
C. Sotillo, H. Thompson, and R. Weinert. 1991. The HCRC
Map Task corpus. Language and Speech, 34(4):351–366.
J. Kathryn Bock. 1986. Syntactic persistence in language pro-
duction. Cognitive Psychology, 18:355–387.
Holly P. Branigan, Martin J. Pickering, and Alexandra A. Cle-
land. 1999. Syntactic priming in language production: Ev-
idence for rapid decay. Psychonomic Bulletin and Review,
6(4):635–640.
Holly P. Branigan, Martin J. Pickering, and Alexandra A. Cle-
land. 2000. Syntactic co-ordination in dialogue. Cognition,
75:B13–25.
Holly P. Branigan, Martin J. Pickering, Jamie Pearson, Janet F.
McLean, and Clifford Nass. 2003. Syntactic alignment be-
tween computers and people: the role of belief about mental
states. In Proceedings of the Twenty-fifth Annual Conference
of the Cognitive Science Society.
Carsten Brockmann, Amy Isard, Jon Oberlander, and Michael
White. 2005. Modelling alignment for affective dialogue. In
Workshop on Adapting the Interaction Style to Affective Fac-
tors at the 10th International Conference on User Modeling
(UM-05). Edinburgh, UK.
Eugene Charniak and Mark Johnson. 2005. Coarse-to-fine n-
best parsing and MaxEnt discriminative reranking. In Proc.
43th ACL.
M. Marcus, G. Kim, M. Marcinkiewicz, R. MacIntyre, A. Bies,
M. Ferguson, K. Katz, and B. Schasberger. 1994. The Penn
treebank: Annotating predicate argument structure. In Proc.
ARPA Human Language Technology Workshop.
Martin J. Pickering and Simon Garrod. 2004. Toward a mech-
anistic psychology of dialogue. Behavioral and Brain Sci-
ences, 27:169–225.
David Reitter, Johanna D. Moore, and Frank Keller. 2006. Prim-
ing of syntactic rules in task-oriented dialogue and sponta-
neous conversation. In Proceedings of the 28th Annual Con-
ference of the Cognitive Science Society.
William N. Venables and Brian D. Ripley. 2002. Modern Ap-
plied Statistics with S. Fourth Edition. Springer.
124
