Acknowledgments in Human-Computer Interaction 
Karen Ward and Peter A. Heeman 
Center for Spoken Language Understanding 
Oregon Graduate Institute of Science & Technology 
20000 NW Walker Road 
Beaverton, Oregon 97006 
wardk@cse, ogi. edu, heeman@cse, ogi. edu 
Abstract 
Acknowledgments are relatively rare in human- 
computer interaction. Are people unwilling to use 
this human convention when talking to a machine, 
or is their scarcity due to the way that spoken-lan- 
guage interfaces are designed? We found that, 
given a simple spoken- interface that pro- 
vided opportunities for and responded to acknowl- 
edgments, about half of our subjects used 
acknowledgments at least once and nearly 30% 
used them extensively during the interaction. 
1 Introduction 
As our ability to build robust and flexible spoken- 
 interfaces increases, it is worthwhile to 
ask to what extent we should incorporate various 
human-human discourse phenomena into our dia- 
logue models. Many studies have shown that peo- 
ple alter their dialogue techniques when they 
believe that they are talking to a computer (e.g., 
Brennan, 1991), so it is not clear that observations 
of human-human conversation will provide us with 
the guidance we need. At the same time, we cannot 
always look to current systems to determine which 
discourse phenomena should be supported. Cur- 
rent-generation interfaces are still relatively fragile, 
and so designers of spoken- systems go to 
some effort to structure dialogues and create 
prompts that guide the user toward short, content- 
ful, in-vocabulary responses (e.g., Basson et al, 
1996; Cole, et al, 1997; Oviatt et al, 1994). One 
result of this approach is the suppression of meta- 
dialogue acts such as acknowledgement and repeti- 
tion. 
The term "acknowledgment" is from Clark and 
Schaefer (1989), who describe a hierarchy of meth- 
ods by which one conversant may signal that 
another's contribution has been understood well 
enough to allow the conversation to proceed. 
Acknowledgments often appear in American 
English conversation as an "okay" or "uh-huh" that 
signals understanding but not necessarily agree- 
ment. These are also called "back channels" or 
"prompts" (e.g., Chu-Carroll & Brown, 1997), 
Closely related to acknowledgments are repeti- 
tions, in which the conversant provides a stronger 
signal that a contribution has been understood by 
repeating part or all of the other's contribution. 
Repetitions are also referred to as "paraphrases" 
(Traum & Heeman, 1996), "echoing" (Swerts et al, 
1998), and "demonstration" (Clark & Schaefer, 
1989). Repetitions are often seen when one is con- 
veying complex information, such as when one 
copies an address or telephone number. 
Neither acknowledgments nor repetitions con- 
tribute new domain information to the conversa- 
tion, but they serve to assure the speaker that 
information has been conveyed successfully. 
Acknowledgments also play a role in managing 
turn-taking in mixed-initiative dialogue; although 
acknowledgments may preface a new contribution 
by the same speaker (Novick & Sutton, 1994), 
often they occur alone as a single-phrase turn that 
appears to serve the purpose of explicitly declining 
an opportunity to take a turn (Sacks et al, 1974). 
Acknowledgments and repetitions are ubiqui- 
tous in many types of human-human conversation. 
In a corpus of problem-solving spoken dialogues, 
for example, Traum and Heeman (1996) found that 
51% of turns began with or consisted of an explicit 
acknowledgment. Given this, one would expect 
that acknowledgments should be modeled in dia- 
logue models for spoken- systems, and 
280 
indeed some research models are beginning to 
incorporate acknowledgments, e.g., Kita et al 
(1996), Aist, (1998), Iwase & Ward (1998). 
Typical human-computer dialogue models are 
structured in ways that suppress the use of 
acknowledgments. In many systems turn-taking is 
completely controlled by one conversant, e.g., the 
user responds to system prompts, which tends to 
eliminate the need for acknowledgments as a turn- 
taking mechanism. In other systems, the use of 
barge-in defeats the common interpretation of an 
acknowledgment; if the user speaks, the system 
contribution is cut off before the user utterance is 
interpreted. If that utterance was intended to signal 
that the contribution should continue, the effect is 
exactly the opposite of the one desired. 
Thus, current design practices both discourage 
and render meaningless the standard uses of 
acknowledgments. If these impediments were 
removed, would people choose to use acknowledg- 
ments when interacting with a computer interface? 
2 Experiment 
This study was designed as a pilot to our larger 
investigation into the effects of incorporating 
acknowledgement behavior in dialogue models for 
spoken- interfaces. Before we attempted 
to compare interfaces with and without acknowl- 
edgement behavior, we wanted to understand 
whether people are willing to use this sort of meta- 
dialogue behavior when interacting with a com- 
puter. 
2.1 Approach 
In this study we hypothesized that subjects will 
choose to use acknowledgments in human-com- 
puter interaction if they are given an interface that 
provides opportunities for and responds to 
acknowledgments. 
In designing the study, we assumed that it 
would not immediately occur to subjects that they 
could use acknowledgments to a computer. At the 
same time, we did not want to explicitly instruct or 
require subjects to use acknowledgment behavior, 
as that would tell us nothing about their prefer- 
ences. We therefore decided against a comparison/ 
control-group experimental design for this initial 
study and instead focused on creating a situation in 
which subjects would have a reason to use 
acknowledgments, perhaps even gain an advantage 
from doing so, while still keeping the behavior 
optional. 
We decided to focus on a somewhat narrow use 
of acknowledgments. Conversants are especially 
likely to offer acknowledgments and repetitions 
when complex information is being presented, 
especially when the conversant is copying the 
information. While this is certainly explainable in 
terms of mutuality of understanding, this particular 
use of acknowledgment may be viewed from a 
more mechanical standpoint as regulating the pace 
at which information is presented. This insight sug- 
gested to us that a fruitful task for this study might 
be one in which the subject is asked to write down 
verbally-presented information, as when taking 
messages over the telephone. 
2.2 Task 
We selected the domain of telephone interface to 
email and designed a task in which subjects were 
asked to transcribe items of information from the 
messages. Writing is slow in comparison to speak- 
ing, so we anticipated that subjects would require a 
slower pace of information presentation when they 
were writing. The messages included information 
not asked for on the question list to simulate "unin- 
teresting" material that the subject would want to 
move through at a faster pace. In this way we 
hoped to motivate subjects to try to control the 
pace at which information was presented. 
The email was presented in segments roughly 
corresponding to a long phrase. After each seg- 
ment, the system paused to give the subject time to 
make notes. If the subject said nothing, the system 
would continue by presenting the next message 
segment. Subjects could accept--and perhaps 
make use of--this delay, or they could reduce it by 
acknowledging the contribution, e.g., "okay," or by 
commanding the system to continue, e.g., "go on." 
The system signalled the possibility of controlling 
the delay by prompting the subject "Are you ready 
to go on?" after the first pause. This prompting was 
repeated for every third pause in which the subject 
said nothing. In this way we hoped to suggest to 
the subjects that they could control the wait time if 
desired without explicitly telling them how to do 
SO. 
On the surface, there is no functional differ- 
ence in system behavior between a subject's use of 
281 
a command to move the system onward (e.g., "go 
on" "next", "continue") and the use of an acknowl- 
edgment ("okay," "uh-huh", or a repetition). In 
either case, the system responds by presenting the 
next message segment, and in fact it eventually 
presents the next segment even if the subject says 
nothing at all. Thus, the design allows the subject 
to choose freely between accepting the system's 
pace (system initiative), or commanding the system 
to continue (user initiative), or acknowledging the 
presentations in a fashion more typical of mixed- 
initiative human conversation. In this way, we 
hoped to understand how the subject preferred to 
interact with the computer. 
2.3 Subjects 
Subjects were told that the study's purpose was to 
assess the understandability and usability of the 
interface, and that their task was to find the 
answers to a list of questions. They were given no 
instructions in the use of the program beyond the 
information that they could talk to it using normal, 
everyday speech. 
The 14 volunteers were native speakers of 
North American English, and most were staff at a 
research university. Ten were female, four were 
male. Ages ranged from 13 to 57. All used comput- 
ers, typically office software and games, but none 
had significant programming experience. Each ses- 
sion lasted about 45 minutes total, and each subject 
was paid $10.00. One subject declined payment. 
2.4 Interface 
As mentioned earlier, one difficulty with recogniz- 
ing acknowledgements in spoken- inter- 
faces is that the use of barge-in tends to defeat the 
purpose of acknowledgments when they occur in 
overlapped speech. We used a Wizard of Oz proto- 
col as a simple way to allow the system to respond 
to such utterances and to provide robustness in 
handling repetitions. 
The wizard's interface was constructed using 
the Rapid Application Developer in the Center for 
Spoken Language Understanding Toolkit (Sutton, 
et al, 1998). A simple button panel allowed the 
wizard to select the appropriate response from the 
actions supported by the application. The applica- 
tion functionality was deliberately limited to sug- 
gest realistic abilities for a current spoken- 
 interface. Using messages pre-recorded 
in a synthesized voice, the wizard was able to 
direct the system to: 
• Read a list of all messages. 
• Begin reading a particular message. 
• Read the next message segment. 
• Repeat the current message segment. 
• Repeat the previous message segment. 
• Ask the subject whether the program should 
continue reading the current message. 
• Ask the subject to what to do next. 
• End the program. 
• Play one of several error and help messages. 
The texts of the email messages were pre- 
sented in phrases of varying lengths, with each 
phrase followed by a pause of about five seconds. 
Preliminary tests showed that the combined 
response time of the wizard and the interface was 
between one and two seconds, and that pauses of 
less than five seconds were not obviously different 
from the normal pace of system response. Five sec- 
onds is a long response time, uncomfortably so for 
human-human conversation, so we hoped that this 
lengthy pause would encourage the subjects to take 
the initiative in controlling the pace of the interac- 
tion. 
The messages were divided into segments by 
hand. The divisions were intended to simulate a 
phrase-level presentation, although some short 
phrases were combined to make the presentation 
less choppy. An example of one message and its 
division into phrases may be seen in Figure 1. 
Synthesized speech from the Festival speech 
synthesizer (Taylor, et al, 1998) was used through- 
out the interface. The message texts were presented 
in a synthesized male voice, while the control por- 
tions of the interface used a synthesized female 
voice. Default pronunciations were used except 
when the default was incorrect, e.g., "read" 
defaulted to the past-tense pronunciation in all con- 
texts. Also, there was minor use of the SABLE 
markup  (Wouters, et al, 1999) to flatten 
the pitch range at the end of phrases in list items; 
the intent was to suggest the prosody of list contin- 
uation rather than the default sentence-final drop. 
282 
Message six is from Jo at teleport dot com, about, please stop by store on your way home. 
I'm going to be late getting home tonight, so would you please stop by the store on your way home? 
We need milk, 
eggs, 
a bunch of spinach, 
fresh ginger, 
green onions, 
maple syrup, 
a pound of coos-coos, 
mild curry powder, 
a pound of coffee, 
and a package of seventy five watt light bulbs. 
Thanks! See you tonight. 
Figure 1. Text of a sample message. 
The subject's list of questions included "What items are you supposed to pick up at the store?" 
To improve the understandability, both voices were 
slowed slightly to 90% of the default speaking rate. 
2.5 Measures 
The central question to be answered is: will the 
subject use acknowledgments in interacting with 
the program? A subject can show one of several 
patterns of response: 
• The subject may make no attempt to control the 
pacing of the interface, instead allowing the 
interaction to proceed via time-outs. 
• The subject may use only commands to control 
the pacing. 
• The subject may use only acknowledgments to 
control the pacing. 
• The subject may use a mixture of commands and 
acknowledgments. 
The determination as to whether a particular utter- 
ance constituted an acknowledgment or a com- 
mand was based primarily on word choice and 
dialogue context; this approach is consistent with 
definitions of this behavior, e.g., Chu-Carroll and 
Brown (1997). For example, "yes" in the context of 
a system inform (a segment of an email message) 
was considered an acknowledgment, but "yes" in 
the context of a system question was not. The 
words "okay," "uh-huh," and "yes" (immediately 
following an inform) were taken as evidence of 
acknowledgments, and phrases such as "go on," 
"continue," "next" following an inform were taken 
as evidence of commands. The interpretation was 
confirmed during the post-experiment interview by 
questioning the subjects about their word choice. 
2.6 Post-Experiment Interview 
A post-experiment interview was conducted to 
gather subject feedback and to answer subjects' 
questions. The experimenter took notes and thus 
could have introduced bias in the record of 
responses° No tape recording was made. 
The subject was first invited to comment on the 
interface and the interaction in an open-ended fash- 
ion. When the subject had finished, the experi- 
menter asked several specific questions to assess 
their understanding of the interface functionality. 
During this time, the experimenter reminded the 
subjects of the words that they had used most fre- 
quently to prompt the system to continue during 
pauses and asked the subjects why they had 
selected those words. 
Finally, the experimenter explained the true 
purpose and hypothesis of the experiment, verified 
that the subject was unaware that they had been 
interacting with a Wizard-of-Oz interface, and 
asked the subject to comment on the notion of 
using acknowledgments when interacting with a 
computer. The responses to this question, espe- 
cially, must be assumed to be somewhat optimistic, 
as it is likely that at least some subjects would be 
reluctant to disagree with the experimenter. 
3 Results 
Results are summarized in Table l.Because the 
subject pool was not balanced for gender, results 
for male and female subjects are reported sepa- 
rately. Due to the small number of male subjects in 
283 
Table 1. Summary of Acknowledgment Behavior 
Behavior 
Subjects 
Female 
10 subjects 
Male. 
4 subjects 
Total 
(14) 
Used acknowledgment/repetition at least once 4 (40%) 4 (100%) 8 (57%) 
Used acknowledgment/repetition more than command 3 (30%) 1 (25%) 4 (29%) 
Used acknowledgment but no commands 1 (i0%) 0 1 (7%) 
Described acknowledgment to computer as strange 2 (20%) 0 2 (14%) 
this pilot study, no tests of statistical significance of 
differences in the rates of acknowledgment behav- 
ior were made. 
Eight of the fourteen subjects used an 
acknowledgment or repetition at least once, and 
four used acknowledgment/repetitions more fre- 
quently than they used commands. Only one sub- 
ject used acknowledgments exclusively, while five 
subjects never used acknowledgments. No subject 
relied exclusively on time-outs to allow the system 
to proceed at its own pace, although one subject 
did use that as her predominant method (42 times, 
while using acknowledgments only six times and 
commands three times). Only one subject used rep- 
etition, and he reported during the interview that he 
was unaware of having done so. 
It is interesting to note that while all of the 
male subjects in this sample exhibited acknowledg- 
ment behavior at least once, only one preferred 
acknowledgment over command. One of the male 
subjects used acknowledgments only three times, 
in all cases as prefaces to commands. Conversely, 
although a lower percentage of women used 
acknowledgments (40%), a higher percentage of 
them (30%) used acknowledgments in preference 
to commands. Because of the small numbers of 
subjects, however, we do not conclude that these 
differences are significant. 
During the post-experiment interview, two sub- 
jects (both female) described the idea of using 
acknowledgments to the computer as strange and 
stated that they didn't feel that they would do this 
unless directed tt>---and even then, they would 
regard it as simply an alternate command. Two 
other subjects, both females who had used 
acknowledgments 2-6 times during the task, each 
reported that she had felt silly when she had caught 
herself saying "please" and "okay" to a computer 
but had been pleased when it had worked. The 
remainder of the subjects either expressed no 
strong opinion (two, both female) or expressed a 
positive attitude toward being able to use acknowl- 
edgments when interacting with a computer. Two 
subjects who had not used acknowledgments com- 
mented that they would probably be more likely to 
use human-like conversation if the synthesized 
voice were more human-like. 
Again, this report of the subjects' attitudes 
should be interpreted with caution; at this point in 
the interview they knew the experimenter's hypoth- 
esis and so may have been reluctant to disagree. 
3.1 Other Dialogue Behaviors 
Although we had not formed any hypothesis about 
other dialogue behaviors, we noticed several inter- 
esting dialogue behaviors that we had not antici- 
pated. 
We were surprised at the number of subjects 
who exhibited politeness behavior toward the inter- 
face, either saying "please" when issuing com- 
mands to the computer or responding to the 
program's "good-bye" at the end of the session. 
One subject used "please" throughout the interac- 
tion, but a more common pattern was to use 
"please" at the beginning of the session and to drop 
the behavior as the interface became more familiar. 
Politeness did not seem to be strongly associated 
with a willingness to use acknowledgments, how- 
ever; four of the nine subjects who exhibited 
politeness did not use any acknowledgments in 
their interaction. 
Despite the deliberately-artificial interface, 
several subjects responded at least once to the mes- 
sage content as if they were talking to the message 
284 
System: I could come to your office now or at any of the following times, one thirty 
SUBJECT: continue 
System: three o clock 
SUBJECT: continue 
System: or five fifteen 
SUBJECT: continue 
System: thank you. I look forward to your prompt reply 
SUBJECT: thank you- uh ((laugh)) continue 
Figure 2. Excerpt of transcript. 
Subject thanks the interface. The system is reading the text of one of the messages. 
sender. In the excerpt shown in Figure 2., for exam- 
ple, the subject replied "I'hank you" to the message 
text's "thank you." This did not appear to be a mat- 
ter of misunderstanding the capabilities of the 
interface; the subject later reported that despite the 
synthesized voices she had briefly forgotten that 
she wasn't talking to her secretary. 
Three subjects also made one or more meta- 
comments, e.g., "ah, there it is" when finding a par- 
ticular piece of information. These may have been 
at least partially an artifact of the "treasure hunt" 
nature of the task. When questioned in the post- 
experiment interview, subjects didn't seem aware 
that they'd made these comments. All but one of 
these instances were followed immediately by a 
command, so the wizard responded to the com- 
mand and ignored the meta-comment. The one 
stand-alone meta-comment was treated as an 
unrecognized command (an error message was 
played). 
4 Discussion 
Subjects were provided with three methods for 
controlling the pace at which information was pre- 
sented: silence, command, or acknowledgment/ 
repetition. The majority of the subjects used com- 
mands more than they used acknowledgments, but 
over one half used an acknowledgment or repeti- 
tion at least once during their interaction and 
nearly 30% used acknowledgments in preference 
to commands. This occurred despite the fact that 
subjects were given no reason to think that this 
behavior would be effective: the interface was 
deliberately limited in functionality, and voice syn- 
thesis was used instead of recorded voice to 
emphasize the artificial nature of the interaction. 
Furthermore, the interface did not offer acknowl- 
edgments to the subjects, and the subjects were 
given no instructions suggesting that the interface 
understood acknowledgments. In fact two subjects 
who did use acknowledgments expressed surprise 
that they had worked, and two who had not used 
acknowledgments reported that they would proba- 
bly have used them if they had known it would 
work. 
It is interesting to consider these results in light 
of those reported by Okato et al (1998). They 
describe a Japanese- Wizard-of-Oz study 
in which the subjects were given some instruction 
on using the system, and in which the system both 
presented and accepted back-channel feedback. 
They found that even when the interface offered 
back channels itself the rate of subject back-chan- 
nels was somewhat lower in human-computer 
interaction than in comparable human-human con- 
versation. This makes the fact that our interface 
elicited acknowledgments without offering them 
even more encouraging. Clearly, some people are 
willing to utilize this human conversational con- 
vention in human-computer dialogue. Our post- 
experiment interviews suggest, however, that some 
people will find the use of acknowledgements 
strange or uncomfortable in human-computer inter- 
action. While self-reports of attitudes toward hypo- 
thetical situations must be treated with some 
caution, it seems reasonable to assume that even 
when such interfaces become available there will 
be users who will prefer to interact with computers 
using commands. 
Will attitudes and conversational behavior 
change as people gain experience with more 
advanced spoken- interfaces? Despite the 
285 
relatively short duration of this test--most subjects 
completed the task itself in 15-20 minutes--some 
changes in behavior could be observed over the 
course of the dialogue. In particular, politeness 
behaviors were likely to be seen early in the dia- 
logues and then diminish as the subjects became 
more comfortable with their interaction. We specu- 
late that the use of politeness words did not reflect 
a strong underlying politeness toward the computer 
so much as a falling back on human conventions 
when faced with an unfamiliar dialogue situation. 
One subject who had used "please" 21 times dur- 
ing the interaction, for example, simply hung up 
without warning when she had finished. This con- 
trasts, however, with the findings of Nass et al 
(1999) that people do offer socially-desirable 
behavior to computers. 
Would a better voice increase the incidence of 
acknowledgment behavior? Several subjects 
thought it would, and even with the current synthe- 
sized voices we saw several examples of subjects 
seemingly forgetting briefly that they were not 
talking to a human. We plan to explore this ques- 
tion in future work. 
4.1 Conclusions and Future Work 
We conducted a preliminary study to examine the 
willingness of subjects to use a particular dialogue 
act, acknowledgment, in human-computer interac- 
tion. Although the number of subjects was small, 
we saw that about half of our subjects used 
acknowledgements or repetition at least occasion- 
ally to control the pace at which information was 
presented, and about 29% used acknowledgments 
more frequently than they used commands for that 
purpose. 
Our immediate plans include extending this 
study to a larger and gender-balanced group of sub- 
jects so that we can draw firmer quantitative con- 
clusions about the percentage of people who are 
likely to prefer this style of interaction. In particu- 
lar, we cannot conclude from the current study's 
small sample how strong the preference for using 
acknowledgment might be, especially among male 
subjects. Also, in our current study the subject 
achieved no functional benefit in using acknowl- 
edgments. With better estimates of subject prefer- 
ences, we can then proceed to our larger goal of 
comparing the usefulness and user acceptability of 
spoken  dialogue models with and without 
acknowledgment behavior (c.f. Walker, 1993). We 
also plan to explore the effect of the quality of the 
synthesized voice on the incidence of acknowledg- 
ment behavior. 
Acknowledgments 
This work was partially supported by a grant from 
Intel Research Council. The authors gratefully 
acknowledge and thank David G. Novick and the 
anonymous reviewers for their helpful comments 
and suggestions. 

References 

Gregory Aist. 1998. Expanding a Time-Sensitive 
Conversational Architecture for Turn-Taking 
to Handle Content-Driven Interruption. In Proceedings of lCSLP 98 Fifth International Conference on Spoken Language Processing, pages 
413-417. 

Sara Basson, Stephen Springer, Cynthia Fong, 
Hong Leung, Ed Man, Michele Olson, John 
Pitrelli, Ranvir Singh, and Suk Wong. 1996. 
User Participation and Compliance in Speech 
Automated Telecommunications Applications. 
In Proceedings of ICSLP 96 Fourth International Conference on Spoken Language Processing, pages 1676-1679. 

Susan E. Brennan. 1991. Conversation With and 
Through Computers. User Modeling and User-Adapted Interaction. 1:67-86. 

Jennifer Chu-Carroll and Michael K. Brown. 1997. 
Tracking Initiative in Collaborative Dialogue 
Interactions. In Proceedings of the 35th Annual 
Meeting of the Association for Computational 
Linguistics, pages 262-270. 

Herbert H. Clark and Edward F. Schaefer. 1989. 
Contributing to Discourse. Cognitive Science, 
13:259-294. 

Ronald A. Cole, David G. Novick, Pieter J. E. Vermeulen, Stephen Sutton, Mark Fanty, L. F. A. 
Wessels, Jacques Ho de Villiers, Johan Schalkwyk, Brian Hansen and Daniel Bumett. 1997. 
Experiments with a Spoken Dialogue System 
for Taking the U.S. Census. Speech Communications, Vol. 23. 

Peter A. Heeman, Michael Johnston, Justin Denney and Edward Kaiser. 1998. Beyond Structured Dialogues: Factoring out Grounding. In 
Proceedings of ICSLP 98 Fifth International 
Conference on Spoken Language Processing, 
pages 863-867. 

Tatsuya Iwase, and Nigel Ward. 1998. Pacing Spoken Directions to Suit the Listener. In Proceedings of ICSLP 98 Fifth International 
Conference on Spoken Language Processing, 
Vol. 4, page 1203-1207. 

K. Kita, Y. Fukui, M. Jagata and T. Morimoto. 
1996. Automatic Acquisition of Probabilistic 
Dialogue Models. In Proceedings of lCSLP 96 
Fourth International Conference on Spoken 
Language Processing, pages 196-199. 

Clifford Nass, Youngme Moon and Paul Carney. 
1999. Are Respondents Polite to Computers? 
Social Desirability and Direct Responses to 
Computers. Journal of Applied Social Psychology, 29:5, pages 1093-1110. 

H. Noguchi and Yasuharu Den. 1998. Prosody-Based Detection of the Context of Back channel Responses. In Proceedings of ICSLP 98 
Fifth International Conference on Spoken Language Processing, Vol. 2, pages 487-490. 

David G. Novick and S. Sutton. 1994. An Empirical Model of Acknowledgment for Spoken-Language Systems. In Proceedings of the 32nd 
Annual Meeting of the Association for Computational Linguistics, pages 96-101. 

Yohei Okato, Keiji Kato, Mikio Yamamoto and 
Shuichi Itahashi. 1998. System-User Interaction and Response Strategy in Spoken Dialogue System. Proceedings of lCSLP 98 Fifth 
International Conference on Spoken Language Processing, Vol. 2, pages 495-498. 

S. L. Oviatt and P. R. Cohen and M. Wang. 1994. 
Toward Interface Design for Human Language 
Technology: Modality and Structure as Determinants of Linguistic Complexity. In Speech 
Communication, 15:3-4, pages 283-300. 

H. Sacks, E. Schegloff and G. Jefferson. 1974. A 
Simplest Systematics for the Organization of 
Turn-Taking in Conversation. Language, 
50:696-735. 

Stephen Sutton, Ronald Cole, Jacques de Villiers, 
Johan Schalkwyk, Pieter Vermeulen, Mike 
Macon, Yonghong Yan, Ed Kaiser, Brian Rundie, Khaldoun Shobaki, Paul Hosom, Alex Kain, Johan Wouters, Dominic Massaro and 
Michael Cohen. 1998. Universal Speech Tools: 
the CSLU Toolkit. In Proceedings of the International Conference on Spoken Language Processing, pages 3221-3224. 

Marc Swerts, Hanae Koiso, Atsushi Shimojima 
and Yasuhiro Katagiri. 1998. On Different 
Functions of Repetitive Utterances. In Proceedings of lCSLP 98 Fifth International Conference on Spoken Language Processing, Vol. 
2, pages 483-487. 

Paul Taylor, Alan W. Black and Richard Caley. 
1998. The Architecture of the Festival Speech 
Synthesis System. In The Third ESCA/ 
COCOSDA Workshop on Speech Synthesis, 
pages 147-151. 

David R. Traum and Peter A. Heeman. 1996. 
Utterance Units and Grounding in Spoken Dialogue. In Proceedings of ICSLP 96 Fourth 
International Conference on Spoken Language Processing, pages 1884-1887. 

Marilyn A. Walker. 1993. Informational Redundancy and Resource Bounds in Dialogue. Doctoral dissertation, University of Pennsylvania. 

Johan Wouters, Brian Rundle and Michael W. 
Macon. 1996. Authoring Tools for Speech 
Synthesis using the Sable Markup Standard. In 
Proceedings of Eurospeech '99. 

Nigel Ward. 1996. Using Prosodic Clues to Decide 
When to Produce Back-Channel Utterances. In 
Proceedings of ICSLP 96 Fourth International 
Conference on Spoken Language Processing, 
page 1724-1727.
