Proceedings of the Fourth International Natural Language Generation Conference, pages 25–32,
Sydney, July 2006. c©2006 Association for Computational Linguistics
Individuality and Alignment in Generated Dialogues
Amy Isard and Carsten Brockmann and Jon Oberlander
School of Informatics, University of Edinburgh
2 Buccleuch Place
Edinburgh EH8 9LW, UK
{Amy.Isard, Carsten.Brockmann, J.Oberlander}@ed.ac.uk
Abstract
It would be useful to enable dialogue
agents to project, through linguistic
means, their individuality or personality.
Equally, each member of a pair of agents
ought to adjust its language (to a greater or
lesser extent) to match that of its interlocu-
tor. We describe CRAG, which generates
dialogues between pairs of agents, who are
linguistically distinguishable, but able to
align. CRAG-2 makes use of OPENCCG
and an over-generation and ranking ap-
proach, guided by a set of language mod-
els covering both personality and align-
ment. We illustrate with examples of out-
put, and briefly note results from user stud-
ies with the earlier CRAG-1, indicating
how CRAG-2 will be further evaluated.
Related work is discussed, along with cur-
rent limitations and future directions.
1 Introduction
A computer agent should be individual. Nass
and collaborators find that users’ responses
to computer-agents are influenced by whether
the agent’s linguistic personality matches—or
mismatches—the personality of the user (Moon
and Nass, 1996; Nass and Lee, 2000). Similarly,
characters in virtual environments should be dis-
tinctive (Ball and Breese, 2000; Rist et al., 2003).
But an aspect of personality is how well you adjust
to other people (and their language use): align-
ment. Pickering and Garrod’s Interactive Align-
ment Model suggests that people tend to automat-
ically converge on lexical and syntactic choices,
via a low-level mechanism of interpersonal prim-
ing (Pickering and Garrod, 2004), and Brennan
has shown that people will align their language to-
wards that of computer agents (Brennan, 1996).
But it is an open issue as to whether some peo-
ple are better ‘aligners’ than others. Conversely,
alignment is only visible and interesting (among
computer agents) if they start out being individual.
We therefore set out to simulate both individ-
uality and alignment. The paper briefly surveys
the evidence for linguistic personality, for inter-
personal alignment, and for interaction between
them. It then sketches the current version of
CRAG. CRAG-2 makes use of OPENCCG and
an over-generation and ranking approach, guided
by a set of language models for personality and
alignment. We illustrate the differing linguis-
tic behaviours that it generates, and briefly note
promising results from user studies with the ear-
lier CRAG-1 system, indicating how CRAG-2 will
be further evaluated. Related work is discussed,
along with possible directions for future work.
2 Background
2.1 Personality and Language
Current work on personality traits is dominated by
Costa and McCrae’s five-factor model (Costa and
McCrae, 1992). The five factors, or dimensions,
are: Extraversion; Neuroticism; Openness; Agree-
ableness; and Conscientiousness (Matthews et al.,
2003). It has been shown that scores on these di-
mensions correlate with some aspects of language
use (Scherer, 1979; Dewaele and Furnham, 1999).
In studies of text, the focus has been on lexical
choice, and Pennebaker and colleagues have anal-
ysed relative frequencies of use of word-stems in
a dictionary structured into semantic and syntac-
tic categories (Pennebaker et al., 2001). Amongst
other results, they have shown that High Extraverts
25
use: more social process talk, positive emotion
words and inclusives; and fewer negations, ten-
tative words, exclusives, causation words, nega-
tive emotion words, and articles (Pennebaker and
King, 1999; Pennebaker et al., 2002).
Computational linguistic exploitation of such
empirically-derived features has been limited. On
the one hand, in generation, there has been work
on personality-based generation. For instance, in
developing embodied conversational agents, re-
searchers have designed agents or teams of agents
with distinguishable linguistic personalities (Ball
and Breese, 2000; Rist et al., 2003; Piwek and
van Deemter, 2003; Gebhard, 2005). However,
the linguistic behaviour is usually informed by
rules based on personality stereotypes, rather than
on language statistics themselves. On the other
hand, in interpretation, more empirical work has
recently been carried out, to enable text classifi-
cation. Argamon et al. (2005) attempted to clas-
sify authors as High or Low Extravert and High
or Low Neurotic, using Pennebaker and King’s
(1999) data. They report classification accuracies
of around 58% (with a 50% baseline). Oberlander
and Nowson (2006) undertake a comparable task,
using weblog data. They report classification ac-
curacies of roughly 85% (Neuroticism) and 94%
(Extraversion), and comparable figures for Agree-
ableness and Conscientiousness. Such studies can
provide ordered lists of linguistic features which
are useful for distinguishing language producers,
and we will return to this, below.
2.2 Alignment and Language
People converge with their interlocutors in linguis-
tic choices at a number of levels (Pickering and
Garrod, 2004). The phenomena can be seen in
both social and cognitive terms. On the social side,
co-operative processes such as audience design
are usually considered to be conscious, at least in
part (Bell, 1984). But on the cognitive side, co-
ordinative processes such as alignment are usu-
ally considered to be largely automatic (Garrod
and Doherty, 1994). Alignment can be probed
by psycholinguistic tests for interpersonal prim-
ing, establishing the extent to which participants
are more likely to use a lexical item or syntac-
tic construction after hearing their conversational
partner use it. Syntactic priming experiments in-
volve constructions such as passives, and ditransi-
tives (Pickering and Branigan, 1998).
It is possible that some people are stronger
aligners than others. Gill et al. (2004) probed
syntactic priming for passives, and investigated
whether levels of Extraversion or Neuroticism
would affect the strength of priming effects. It
was found that Extraversion has no effect, but that
Neuroticism has a non-linear effect: both High and
Low levels of Neuroticism led to weaker priming;
Mid levels led to significantly stronger priming.
Given this, if a generation system is going to simu-
late alignment, it is probably worth designing it so
that it can simulate agents with differing propensi-
ties to align.
3 The CRAG System Overview
The system described in the following sections
(CRAG-2) is the successor to CRAG-1 which is
detailed in Isard et al. (2005). The system gener-
ates a dialogue between two computer agents on
the subject of opinions about a film. CRAG-2 uses
the OPENCCG parsing and generation framework
(White, 2004; White, 2006). The realiser com-
ponent takes a logical form as input and outputs
a list of candidate sentences ranked using one or
more language models. In CRAG-2, we use the
OPENCCG generator to massively over-generate
paraphrases, and the combination of n-gram mod-
els described in Section 4 to choose the best ut-
terance according to a character’s personality and
agenda, and the dialogue history.
4 N-Grams: Personality and Alignment
Modelling
4.1 N-Gram Language Models
The basic assumption underlying CRAG-2 is that
personality, as well as alignment behaviour, can
be modelled by the combination of a variety of n-
gram language models.
Language models are trained on a corpus and
subsequently used to compute probability scores
of word sequences. An n-gram language model
approximates the probability of a word given its
history of the preceding n−1 words. According
to the chain rule, probabilities are then combined
by multiplication. Equation (1) shows a trigram
model that takes into account two words of context
to predict the probability of a word sequence wn1:
(1) P(wn1)≈
n∏
i=1
P(wi|wi−1i−2)
26
4.2 Avoiding the Length Effect
Because word probabilities are always less than 1
and therefore each multiplication decreases the to-
tal, if we use this standard model, longer sentences
will always receive lower scores (this is known as
the length effect). We therefore calculate the prob-
ability of a sentence as the geometric mean of the
probability of each word in the sentence as shown
in (2):
(2) P(wn1)≈
n∏
i=1
P(wi|wi−1i−2)1/n
4.3 Linear Combination of Language Models
OPENCCG supports the linear combination of
language models, where each model is assigned a
weight. For uniform interpolation of two language
models Pa and Pb, each receives equal weight:
(3) P(wi|wi−1i−2)= Pa(wi|w
i−1
i−2)+Pb(wi|w
i−1
i−2)
2
In the more general case, the language models
are assigned weights λi, the sum of which has to
be 1:
(4) P(wi|wi−1i−2)=λ1Pa(wi|wi−1i−2)+λ2Pb(wi|wi−1i−2)
For example, setting λ1 = 0.9 and λ2 = 0.1 assigns
a high weight to the first language model.
4.4 OPENCCG N-Gram Ranking
In the OPENCCG framework, language models
can be used to influence the chart-based realisation
process. The agenda of edges is re-sorted accord-
ing to the score an edge receives with respect to a
language model. For CRAG-2, many paraphrases
are generated from a given logical form, and they
are then ranked in order of probability according
to the combination of n-gram models appropriate
for the character and stage of the dialogue.
5 CRAG-2 Personality and Alignment
Models
We use the SRILM toolkit (Stolcke, 2002) to com-
pute our language models. All models (except
for the cache language model described in Sec-
tion 5.4) are trigram models with backoff to bi-
grams and unigrams.
We have experimented with two strategies for
creating personality models. Since we want to
study the effects of alignment as well as person-
ality, it is essential that the two characters in a di-
alogue be distinct from one another, so that the ef-
fects of alignment can be seen. The first strategy
involves using typical language for each personal-
ity trait, and the second uses the language of one
individual. In both cases, the language models de-
scribed in the following sections are combined as
described in Section 5.5.
5.1 Building a Personality
Nowson (2006) performed a study on language
use in weblogs. The weblog authors were asked to
complete personality questionnaires based on the
five-factor model (see Section 2.1). All weblog au-
thors scored High or Medium on the Openness di-
mension, so we have no data for typical Low Open
language.
We divided the data into High, Medium and
Low for each personality dimension, and trained
language models so that we would be able to as-
sess the probability of a word sequence given a
personality type. This means that each individual
weblog is used 5 times, once for each dimension.
For each personality dimension, the system sim-
plifies a character’s personality setting x by assign-
ing a value of High (x > 70), Medium (30 < x ≤
70) or Low (x ≤ 30). The five models correspond-
ing to the character’s assigned personality are uni-
formly interpolated to give the final personality
model. If the character has been given a low Open-
ness score, since we do not have a model for this
personality type, we simply interpolate the other
four models.
5.2 Borrowing a Personality
Our second strategy was to train n-gram models
on language of the individuals from the CRAG-1
corpus (Isard et al., 2005) and to use one of these
models for each character in the dialogue.
5.3 Base Language Model
In the case of building a personality, a base lan-
guage model is obtained by combining a language
model computed from the corpus collected for the
CRAG-1 system and a general language model
based on data from the Switchboard corpus (Stol-
cke et al., 2000). The combined base model alone
would rank the utterances without any bias for per-
sonality or alignment. When we are borrowing a
personality, the base model is calculated from the
Switchboard corpus alone.
27
5.4 Cache Language Model
We simulate alignment by computing a cache lan-
guage model based on the utterance that was gen-
erated immediately before. This dialogue history
cache model is the uniform interpolation of word-
and class-based n-gram models, where classes act
as a backoff mechanism when there is no exact
word match. Classes group together lexical items
with similar semantic properties, e.g.:
• good, bad: quality-adjective
• loved, hated: opinion-verb
Details of this approach can be found in Brock-
mann et al. (2005).
5.5 Combining the Language Models
The system uses weights to combine all the mod-
els described above. First the base and person-
ality models are interpolated to produce a base-
personality model, and finally the cache model is
introduced to add alignment effects.
6 Dialogue and Utterance Specifications
6.1 Character Specification
Two computer characters are parameterised for
their personality by specifying values (on a scale
from 0 to 100) for the five dimensions: Extraver-
sion (E), Neuroticism (N), Openness (O), Agree-
ableness (A), and Conscientiousness (C). Their
alignment behaviour is set to a value between 0
(low propensity to align) and 1 (high propensity
to align). Also, each character receives an agenda
of topics they wish to discuss, along with polari-
ties (positive/negative) that indicate their opinion
on the respective topic.
6.2 Utterance Design
The character with the higher Extraversion score
begins the dialogue, and their first topic is se-
lected. Once an utterance has been generated, the
other character is selected, and the system applies
the algorithm shown in (5) to decide which topic
should come next. This process continues until
there are no topics left on the agenda of the cur-
rent speaker.
(5) if (A < 46) or (C < 46) or
(no. of utts about this topic = 2)
then take next topic from own agenda
else continue on same topic
The system creates a simple XML representa-
tion of the character’s utterance, using the speci-
fied topic and polarity. An example using the topic
music and polarity negative is shown in Figure 1.
At this point the system also decides which dis-
course connectives may be appropriate, based on
the previous topic and polarity.
<utterance>
<utt topic="music" polarity="dislike"
opp-polarity="like" so="no" right="no"
also="no" well="yes" and="no" but="no">
<pred adj="bad"/>
<opp-pred adj="good"/>
</utt>
</utterance>
Figure 1: Simple Utterance Specification
6.3 OPENCCG Logical Forms
Following the method described in Foster and
White (2004), the basic utterance specification is
transformed, using stylesheets written in the XSL
transformation language, into an OPENCCG log-
ical form. We make use of the facility for defin-
ing optional and alternative inputs and underspec-
ified semantics to massively over-generate candi-
date utterances. A fragment of the logical form
which results from the transformation of Figure 1
is shown in Figure 2. We also include some frag-
ments of canned text from the CRAG corpus in our
OPENCCG lexicon.
We also add optional interjections (i mean, you
know, sort of ) and conversational markers (right,
but, and, well) where appropriate given the dis-
course history.
When the full logical form is processed by the
OPENCCG system, the output consists of sen-
tences of the types shown below:
(I think) the music was bad.
(I think) the music was not (wasn’t)
good.
I did not (didn’t) like the music.
I hated the music.
One thing I did not (didn’t) like was the
music.
One thing I hated was the music.
The fragmentary logical form in Figure 2 would
create all possible paraphrases from:
(well) (you know) I (kind of) [liked/loved] the
[music/score]
By using synonyms (e.g., plot=story, com-
edy=humour) and combining the sentence types
28
<node id="l1:opinion" pred="like" tense="past">
<rel name="Speaker">
<node id="p1:person" pred="pro1" num="sg"/>
</rel>
<rel name="Content">
<node id="f1:cragtopic" pred="music"
det="the" num="sg"/>
</rel>
<opt>
<rel name="Modifier">
<node id="w1:adv" pred="well"/>
</rel>
<opt>
<opt>
<rel name="HasProp">
<node id="a2:proposition" pred="kind-of"/>
</rel>
</opt>
<opt>
<rel name="Modifier">
<node id="a1:adv" pred="you-know"/>
</rel>
</opt>
</node>
Figure 2: Fragment of Logical Form
Stan: E:53 N:48 A:57 C:46 O:65
agenda: film(neg), dialogue(neg),
music(pos)
other opinions: plot(neg), comedy(neg)
Eddie: E:51 N:43 A:57 C:41 O:65
agenda: plot(neg), comedy(neg),
dialogue(neg)
other opinions: music(pos), film(neg)
Figure 3: Stan and Eddie
and optional expressions, we create up to 3000
possibilities per utterance, and the best candidate
is chosen by the specific combination of n-gram
models appropriate for the given personality and
dialogue history, as described in Section 4.
Our OPENCCG lexicon is based on the core
English lexicon included with the system and we
have added vocabulary appropriate to the movie
domain, and extended the range of grammatical
constructions where necessary.
7 Output and Evaluation
7.1 Output
In this section, we provide some example out-
puts from the CRAG-2 system, using characters
based on participants from our corpus (see Sec-
tion 5.2). Stan is higher on the Extraversion, Neu-
roticism, and Conscientiousness scales than Ed-
die. The characters’ personalities and agendas are
summarised in Figure 3.
We show three example dialogues between Stan
and Eddie. In the first (Figure 4) neither charac-
ter aligns with the other at all, while in the sec-
ond (Figure 5) Stan has a slight tendency towards
alignment and in the third (Figure 6) a more pro-
nounced tendency. In system terms, this means
that in the first dialogue the cache model was given
weight 0, while in the second and third the cache
model was given weights 0.05 and 0.1 respectively
for Stan’s utterances. It can be seen that in the
first dialogue, Eddie’s use of humour in utterance
4 is followed by Stan’s comedy, and dialogue in
utterance 6 is followed by script. In the second
dialogue, Stan aligns his first lexical choice with
Eddie’s but not his second, while in the third di-
alogue he aligns both. The syntactic structures of
Stan’s utterances in the second and third dialogue
also mirror Eddie’s progressively more closely.
1. Stan: I didn’t think much of the film you know.
2. Eddie: I mean I didn’t like kind of like the story.
3. Stan: Yeah there were so many little pathways that they
didn’t take you know.
4. Eddie: I mean I didn’t think the humour was really funny.
5. Stan: Yeah you know the comedy was a bit dull.
6. Eddie: I mean I didn’t like kind of like the dialogue.
7. Stan: I mean the script was a bit dull.
Figure 4: Zero Alignment
1. Stan: I didn’t think much of the film you know.
2. Eddie: I mean I didn’t like kind of like the story.
3. Stan: Yeah there were so many little pathways that they
didn’t take you know.
4. Eddie: I mean I didn’t think the humour was really funny.
5. Stan: I mean the humour was a bit dull.
6. Eddie: I mean I didn’t like kind of like the dialogue.
7. Stan: I mean the script was a bit dull.
Figure 5: Little Alignment from Stan
1. Stan: I didn’t think much of the film you know.
2. Eddie: I mean I didn’t like kind of like the story.
3. Stan: I mean the story was a bit dull.
4. Eddie: I mean I didn’t think the humour was really funny.
5. Stan: I mean the humour was a bit dull.
6. Eddie: I mean I didn’t like kind of like the dialogue.
7. Stan: I mean the dialogue was a bit dull.
Figure 6: More Alignment from Stan
To further illustrate the differences between the
dialogues with and without alignment, we provide
some utterance rankings. We show candidates
for the fifth utterance in each dialogue. Table 1
shows sentences from the example generated with-
out alignment, corresponding to utterance 5 (Stan)
29
1 .03317 Yeah you know the comedy was a
bit dull.
3 .03210 Yeah you know the humour was a bit
dull.
6 .03083 Yeah to be honest I didn’t think that
the comedy was very good either.
15 .02938 I didn’t think much of the comedy
either.
24 .02861 I thought that the comedy was a bit
dull too you know.
Table 1: Ranked Sentences with Zero Alignment
1 .05384 I mean the humour was a bit dull.
8 .05239 The humour wasn’t really funny you
know.
15 .04748 I mean I didn’t think that the humour
was very good either.
19 .04518 I didn’t think much of the humour
either you know.
21 .04478 I thought the humour was a bit dull
too you know.
Table 2: Ranked Sentences with Little Alignment
from Stan
from Figure 4. We show the first five occurrences
of different sentence structures (see Section 6.3),
with their rank and their geometric mean adjusted
scores.
Table 2 shows the the top five sentences from
the fifth utterance from Figure 5 (little alignment),
and Table 3 those from Figure 6 (more align-
ment). It can be seen that when more alignment
is present, the syntactic structure used by the pre-
vious speaker rises higher in the rankings.
7.2 Evaluation
We have not evaluated CRAG-2. However, we
have evaluated CRAG-1. The method was to gen-
erate a set of dialogues, systematically contrasting
characters with extreme settings for the personal-
ity dimensions (High/Low Extraversion, Neuroti-
cism, and Psychoticism1).
1CRAG-1 used the simpler PEN three factor personality
model.
1 .07081 I mean the humour was a bit dull.
2 .06432 The humour wasn’t really funny you
know.
15 .05516 I mean I didn’t think that the humour
was really funny either.
27 .05000 I thought the humour was a bit dull
too you know.
36 .04884 I mean I didn’t think much of the hu-
mour either.
Table 3: Ranked Sentences with More Alignment
from Stan
Human subjects were asked to fill in a question-
naire to determine their personality. They were
then given a selection of dialogues to read. After
each dialogue, they were asked to rate their per-
ception of the interaction and of the characters in-
volved by assigning scores to a number of adjec-
tives related to the personality dimensions.
It was found that subjects could recognise dif-
ferences in the Extraversion level of the language.
Also, the personality setting of a character influ-
enced the perception of its and its dialogue part-
ner’s personality (Kahn, 2006).
We plan a similar evaluation for CRAG-2 to be
able to compare human raters’ impressions of di-
alogues generated by the two systems. We also
plan to evaluate CRAG-2 internally by varying the
weight given to the underlying language models,
and observing the effects this has on the resulting
ranking of the generated utterances.
8 Related Work
Related work in NLG involves either personality
or alignment. So far as we can tell, there is little
work on the latter. Varges (2005) suggests that “a
word similarity-based ranker could align the gen-
eration output (i.e. the highest-ranked candidate)
with previous utterances in the discourse context”,
but there is no report yet on an implementation of
this proposal. A rather different approach is sug-
gested by Bateman and Paris (2005), who discuss
initial work on alignment, mediated by a process
of register-recognition. Regarding generation with
personality, the most influential work is probably
Hovy’s PAULINE system, which varies both con-
tent selection and realisation according to an indi-
vidual speaker’s goals and attitudes (Hovy, 1990).
In her extremely useful survey of work on affective
(particularly, emotional) natural language gener-
ation, Belz (2003) notes that the complexity of
PAULINE’s rule system means that numerous rule
interactions can lead to unpredictable side effects.
In response, Paiva and Evans (2004) take a more
empirical line on style generation, which is closer
to that pursued here. Other relevant work includes
Loyall and Bates (1997), who explicitly propose
that personality and emotion could be used in
generation, but Belz observes that technical de-
scriptions of Hap and the Oz project suggest that
the proposals were not implemented. Walker et
al.’s (1997) system produces linguistic behaviour
which is much more varied than our current sys-
30
tem is capable of; but there, variation is driven by
a model of social relations (based on Brown and
Levinson), rather than on personality. The NECA
project subsequently developed methods for gen-
erating scripts for pairs of dialogue agents (Piwek
and van Deemter, 2003), supported by the MIAU
platform (Rist et al., 2003). The VIRTUALHU-
MAN project is a logical successor to this work,
and its ALMA platform provides an integrated ap-
proach to affective generation, covering emotion,
mood and personality (Gebhard, 2005).
9 Conclusion and Next Steps
Our current system takes a much coarser-grained
approach to semantics and discourse goals than
the recent projects described above, in order to
take advantage of empirically-derived relations
between language and personality. It should be
feasible in principle to move to a more sophisti-
cated semantics, but still retain the massive over-
generation and ranking method. However, to
support more perceptible variation, we need to
exploit much larger personality-corpus resources
than have been available up to now, and our cur-
rent priority is to obtain a corpus at least an order
of magnitude larger than what is currently avail-
able. This interest in individual differences and
what corpora can (and cannot) tell us about them
is one we share with Reiter and colleagues (Reiter
and Sripada, 2004).
We also plan to integrate techniques from
CRAG-1 and CRAG-2, by passing the ranked out-
put of CRAG-2 through further processing and
ranking stages. Furthermore, we intend to inves-
tigate longer-ranging alignment processes, taking
into account more than one previous utterance,
with reduced weight by distance, to emulate mem-
ory effects.
With these enhancements, we will take further
steps towards our goal of simulating both individu-
ality and alignment in believable computer agents.
10 Acknowledgements
This research has been funded by Scottish Enter-
prise through the Edinburgh-Stanford Link project
“Critical Agent Dialogue” (CRAG). We would
like to thank Michael White and Scott Nowson for
their assistance and our anonymous reviewers for
their helpful comments.

References
Shlomo Argamon, Sushant Dhawle, Moshe Koppel,
and James W. Pennebaker. 2005. Lexical predic-
tors of personality type. In Proceedings of the 2005
Joint Annual Meeting of the Interface and the Clas-
sification Society of North America.
Gene Ball and Jack Breese. 2000. Emotion and per-
sonality in a conversational agent. In J. Cassell,
J. Sullivan, S. Prevost, and E. Churchill, editors, Em-
bodied Conversational Agents, pages 189–219. MIT
Press, Cambridge, MA, USA.
John A. Bateman and C´ecile L. Paris. 2005. Adap-
tation to affective factors: architectural impacts for
natural language generation and dialogue. In Pro-
ceedings of the Workshop on Adapting the Interac-
tion Style to Affective Factors at the 10th Interna-
tional Conference on User Modeling (UM-05), Ed-
inburgh, UK.
Allan Bell. 1984. Language style as audience design.
Language in Society, 13(2):145–204.
Anja Belz. 2003. And now with feeling: Develop-
ments in emotional language generation. Techni-
cal Report ITRI-03-21, Information Technology Re-
search Institute, University of Brighton, Brighton.
Susan E. Brennan. 1996. Lexical entrainment in spon-
taneous dialog. In Proceedings of the 1996 Inter-
national Symposium on Spoken Dialogue (ISSD-96),
pages 41–44, Philadelphia, PA.
Carsten Brockmann, Amy Isard, Jon Oberlander, and
Michael White. 2005. Modelling alignment for af-
fective dialogue. In Proceedings of the Workshop on
Adapting the Interaction Style to Affective Factors at
the 10th International Conference on User Modeling
(UM-05), Edinburgh, UK.
Paul T. Costa and Robert R. McCrae, 1992. Re-
vised NEO Personality Inventory (NEO-PI-R) and
NEO Five-Factor Inventory (NEO-FFI): Profes-
sional Manual. Odessa, FL: Psychological Assess-
ment Resources.
Jean-Marc Dewaele and Adrian Furnham. 1999. Ex-
traversion: The unloved variable in applied linguis-
tic research. Language Learning, 49:509–544.
Mary Ellen Foster and Michael White. 2004. Tech-
niques for Text Planning with XSLT. In Proc. of the
4th NLPXML Workshop.
Simon Garrod and Gwyneth Doherty. 1994. Conver-
sation, co-ordination and convention: an empirical
investigation of how groups establish linguistic con-
ventions. Cognition, 53(3):181–215.
Patrick Gebhard. 2005. Alma: a layered model of af-
fect. In AAMAS ’05: Proceedings of the Fourth In-
ternational Joint Conference on Autonomous Agents
and Multiagent Systems, pages 29–36, New York,
NY, USA. ACM Press.
Alastair J. Gill, Annabel J. Harrison, and Jon Ober-
lander. 2004. Interpersonality: Individual differ-
ences and interpersonal priming. In Proceedings of
the 26th Annual Conference of the Cognitive Science
Society, pages 464–469.
Eduard Hovy. 1990. Pragmatics and natural language
generation. Artificial Intelligence, 43.
Amy Isard, Carsten Brockmann, and Jon Oberlander.
2005. Re-creating dialogues from a corpus. In
Proceedings of the Workshop on Using Corpora for
Natural Language Generation at Corpus Linguistics
2005 (CL-05), pages 7–12, Birmingham, UK.
Adam S. Kahn. 2006. Master’s thesis, Stanford Uni-
versity.
A. Bryan Loyall and Joseph Bates. 1997. Personality-
rich believable agents that use language. In J. Lewis
and B. Hayes-Roth, editors, Proceedings of the
1st International Conference on Autonomous Agents
(Agents’97). ACM Press.
Gerald Matthews, Ian J. Deary, and Martha C. White-
man. 2003. Personality Traits. Cambridge Univer-
sity Press, Cambridge, 2nd edition.
Youngme Moon and Clifford Nass. 1996. How “real”
are computer personalities? Communication Re-
search, 23:651–674.
Clifford Nass and Kwan Min Lee. 2000. Does
computer-generated speech manifest personality?
an experimental test of similarity-attraction. In
Proceedings of CHI 2000, The Hague, Amsterdam,
2000, pages 329–336.
Scott Nowson. 2006. The Language of Weblogs: A
study of genre and individual differences. Ph.D. the-
sis, University of Edinburgh.
Jon Oberlander and Scott Nowson. 2006. Whose
thumb is it anyway? Classifying author personality
from weblog text. In Proceedings of COLING/ACL-
06: 44th Annual Meeting of the Association for
Computational Linguistics and 21st International
Conference on Computational Linguistics, Sydney.
Daniel S. Paiva and Roger Evans. 2004. A framework
for stylistically controlled generation. In Proceed-
ings of the 3rd International Conference on Natural
Language Generation, pages 120–129.
James W. Pennebaker and Laura King. 1999. Lin-
guistic styles: Language use as an individual differ-
ence. Journal of Personality and Social Psychology,
77:1296–1312.
James W. Pennebaker, Martha E. Francis, and Roger J.
Booth. 2001. Linguistic Inquiry and Word Count
2001. Lawrence Erlbaum Associates, Mahwah, NJ.
James W. Pennebaker, Matthias R. Mehl, and Kate G.
Neiderhoffer. 2002. Psychological aspects of nat-
ural language use: Our words, our selves. Annual
Review of Psychology, 54:547–577.
Martin J. Pickering and Holly P. Branigan. 1998. The
representation of verbs: Evidence from syntactic
priming in language production. Journal of Mem-
ory and Language, 39(4):633–651.
Martin J. Pickering and Simon Garrod. 2004. Towards
a mechanistic psychology of dialogue. Behavioral
and Brain Sciences, 27:169–225.
Paul Piwek and Kees van Deemter. 2003. Dialogue as
discourse: Controlling global properties of scripted
dialogue. In Proceedings of the AAAI Spring Sym-
posium on Natural Language Generation in Spoken
and Written Dialogue.
Ehud Reiter and Somayajulu Sripada. 2004. Contex-
tual influences on near-synonym choice. In Pro-
ceedings of the Third International Conference on
Natural Language Generation, pages 161–170.
Thomas Rist, Elisabeth Andr´e, and Stephan Baldes.
2003. A flexible platform for building applications
with life-like characters. In IUI ’03: Proceedings of
the 8th International Conference on Intelligent User
Interfaces, pages 158–165, New York, NY, USA.
ACM Press.
Klaus Scherer. 1979. Personality markers in speech.
In K. R. Scherer and H. Giles, editors, Social Mark-
ers in Speech, pages 147–209. Cambridge Univer-
sity Press, Cambridge.
Andreas Stolcke, Harry Bratt, John Butzberger, Hora-
cio Franco, Venkata Ramana Rao Gadde, Madelaine
Plauch´e, Colleen Richey, Elizabeth Shriberg, Kemal
S¨onmez, Fuliang Weng, and Jing Zheng. 2000. The
SRI March 2000 Hub-5 conversational speech tran-
scription system. In Proceedings of the 2000 Speech
Transcription Workshop, College Park, MD.
Andreas Stolcke. 2002. SRILM – an extensible lan-
guage modeling toolkit. In Proceedings of the 7th
International Conference on Spoken Language Pro-
cessing (ICSLP-02), pages 901–904, Denver, CO.
Sebastian Varges. 2005. Spatial descriptions as refer-
ring expressions in the MapTask domain. In Pro-
ceedings of the 10th European Workshop on Natural
Language Generation.
Marilyn A. Walker, Janet E. Cahn, and Steve J. Whit-
taker. 1997. Improvising linguistic style: So-
cial and affective bases for agent personality. In
J. Lewis and B. Hayes-Roth, editors, Proceedings
of the 1st International Conference on Autonomous
Agents (Agents’97), pages 96–105. ACM Press.
Michael White. 2004. Reining in CCG Chart Re-
alization. In Proceedings of the 3rd International
Conference on Natural Language Generation, pages
182–191.
Michael White. 2006. Efficient Realization of Coor-
dinate Structures in Combinatory Categorial Gram-
mar. Research on Language & Computation, on-
line first, March.
