Targeted Help for Spoken Dialogue Systems: Intelligent Feedback Improves Naive Users' Performance

8 Pages·2003·0.67 MB·English

by  Hockey, Beth Ann

c
Targeted Help for Spoken Dialogue Systems:
intelligent feedback improves naive users’ perfomance
Beth Ann Hockey Oliver Lemon
Research Institute for Advanced School of Informatics,
Computer Science (RIACS), University of Edinburgh,
NASA Ames Research Center, 2 Buccleugh Place
Moffet Field, CA 94035 Edinburgh EH8 9LW, UK
[email protected] [email protected]
Ellen Campana Laura Hiatt Gregory Aist
Department of Brain Center for the Study of Language RIACS
and Cognitive Sciences and Information (CSLI) NASA Ames Research Center,
University of Rochester Stanford University Moffet Field, CA 94035
Rochester, NY 14627 2 10 Panama St , [email protected]
ecampana@bcs .r ochester . edu Stanford, CA 94305
[email protected]
Jim Hieronymous Alexander Gruenstein John Dowding
RIACS BeVocal, Inc. RIACS
NASA Ames Research Center, 685 Clyde Avenue NASA Ames Research Center,
Moffet Field, CA 94035 Mountain View, CA 94043 Moffet Field, CA 94035
[email protected] [email protected] [email protected]
Abstract periments on a spoken dialogue system
for command and control of a simulated
We present experimental evidence that rob0 tic helicopter.
providing naive users of a spoken dia-
logue system with immediate help mes-
l Introduction
sages related to their out-of-coverage ut-
terances improves their success in using Targeted Help makes use of user utterances ‘that
‘ -
the system. A grammar-based recog- are out-of-coverage of the main dialogue sys-
nizer and a Statistical Language Model tem recognizer to provide the user with immedi-
(SLM) recognizer are run simultane- ate feedback, tailored to what the user said, for
ously. If the grammar-based recognizer cases in which the system was not able to un-
suceeds, the less accurate SLM recog- derstand their utterance. These messages can be
nizer hypothesis is not used. When the much more informative than responding to the
grammar-based recognizer fails and the user with some variant of “Sorry I didn’t under-
SLM recognizer produces a recognition stand”, which is the behaviour of most current di-
hypothesis, this result is used by the Tar- alogue systems. To implement Targeted Help we
geted Help agent to give the user feed- use two recognizers: the Primary Recognizer is
back on what was recognized, a diag- constructed with grammar-based language model
nosis of what was problematic about the and the Secondary Recognizer used by the Tar-
utterance, and a related in-coverage ex- geted Help module is constructed with a Statis-
ample. The in-coverage example is in- tical Language Model (SLM). As part of a spo-
tended to encourage alignment between ken dialogue system, grammar based recogniz-
user inputs and the language model of ers tuned to a domain perfom very well, in fact
the system. We report on controlled ex- better than comparable Statistical Language Mod-
els (SLMs) for in-coverage utterances (Knight et low to interpret ongoing events, which may be re-
al., 2001). However, in practice users will some- ported (via NL generation) to the operator. The
times produce utterances that are out of coverage. robot can carry out various activities such as fly-
Ths is particularly true of non-expert users, who ing to a location, fighting fires, following a ve-
do not understand the limitations and capabilities hicle, and landing. Interaction in WITAS thus
of the system, and consequently produce a much involves joint-activities between an autonomous
lower percentage of in-coverage utteraces than ex- system and a human operator. These are activ-
pert users. The Targeted Help strategy for achiev- ities which the autonomous system cannot com-
ing good performance with a dialogue system is to plete alone, but which require some human inter-
use a grammar-based language model and assist vention (e.g. search for a vehicle). These activi-
users in becoming expert as quickly as possible. ties are specified by the user during dialogue, or
This approach takes advantage of the strengths of can be initiated by the UAV. In any case, a major
both types of language models by using the gram- component of the dialogue, and a way of maintain-
mar based model for in-coverage utterances and ing its coherence, is tracking the state of current
the SLM as part of the Targeted Help system for or planned activities of the robot. This system is
out-of-coverage utterances. sufficiently complex to serve as a good testbed for
In this paper we report on controlled experi- Targeted Help.
ments, testing the effectiveness of an implementa-
tion of Targeted Help in a mixed initiative dialogue 2.2 The Targeted Help Module
system to control a simulated robotic helicopter.
The Targeted Help Module is a separate compo-
nent that can be added to an appropriately struc-
2 System Description
tured dialogue system with minimal changes to ac-
2.1 The WITAS Dialogue System comodate the specifics of the domain. This mod-
Targeted Help was deployed and tested as part ular design makes it quite portable, and a version
of the WITAS dialogue system’, a command and of this agent is in fact being used in a second com-’
control and mixed-initiative dialogue system for mand and control dialogue system (Hockey et al.,
interacting with a simulated robotic helicopter or 2002; ?). It is argued in (?) that “low-level”
UAV (Unmanned Aerial Vehicle) (?). The dia- processing components such as the Targeted Help
logue system is implemented as a suite of agents module are an important focus for future dialogue
communicating though the SRI Open Agent Ar- system research. Figure 1 shows the structure of
chitecture (0-4A) (Martin et d., 1998). The the Targeted Help component and its relationship
agents include: Nuance Communications Recog- to the rest of the dialogue system.
nizer (Nuance, 2002); the Gemini parser and gen- The goal of the Targeted Help system is to han-
erator (Dowding et al., 1993) (both using a gram- dle utterances that cannot be processed by the
mar designed for the UAV application); Festival usual components of the dialogue system, and to
text-to-speech synthesizer (Systems, 2001); a GUI align the user’s inputs with the coverage of the sys-
which displays a map of the area of operation tem as much as possible. To perform this function
and shows the UAV’s location; the Dialogue Man- the Targeted Help component must be able to de-
ager (?); the Robot Control and Report compo- termine which utterances to handle, and then con-
nent, which translates commands and queries bi- struct help messages related to those utterances,
directionally between the dialogue interface and which are then passed to a speech synthesizer. The
the UAV. The Dialogue Manager interleaves mul- module consists of three parts:
tiple planning and execution dialogue threads (?).
While the helicopter is airborne, an on-board 0 the Secondary Recognizer,
active vision system will interpret the scene be-
the Targeted Help Activator,
0
‘See http://www.ida.liu.se/ext/witas
and http://www-csli.stanford.edu/semlab/
witas the Targeted Help Agent.
Prknary speech
recognizer
Speech Synthesis
Figure 1: Architecture
The Targeted Help Activator takes input from What the system heard: a report of the backup
both the main grammar-based recognizer and the SLM recognition hypothesis.
backup category-based SLM recognizer. It uses
What the problem was: a description of the
this input to determine when the Targeted Help
problem with the user’s utterance (e.g. the
component should produce a message. The Acti-
system doesn’t know a word); and
vator’s behavior is as follows for the four possible
combinations of reco,gizer outcomes:
What you might say instead: A similar in-
coverage example.
1. Both recognizers get a recognition hypothe-
sis: In constructing both the diagnostic of the prob-
Targeted Help remains inactive; normal dia- lem with the utterance, and the in-coverage exam-
logue system proccessing proceeds ple, we are faced with the question of whether the
information from the secondary recognizer is suf-
2. Main recognizer gets a recognition hypothe-
ficient to produce useful help messages. Since this
sis and secondary recognizer rejects:
domain is relatively novel, there is not very much
Targeted Help remains inactive; normal dia-
data for training the SLM and the performance re-
logue system proccessing proceeds
flects this. We have designed a rule based system
3. Main recognizer rejects, secondary recog- that looks for patterns in the recognition hypothe-
nizer gets a recognition hypothesis: sis that seem to be detected adequately even with
Targeted Help is activated incomplete or inaccurate recognition.
Diagnostics are of three major types:
4. Both recognizers reject:
Targeted Help is not activated, default system 0 endpointing errors,
failure message is produced
unknown vocabulary,
0
Once Targeted Help is activated, the Targeted
subcategorization mistakes.
Help Agent constructs a message based on the 0
recognition hypothesis from the secondary SLM We found from an analysis of transcripts that
recognizer. These messages are composed of one these three types of errors accounted for the ma-
or more of the following pieces: jority of failed utterances. Endpointing errors are
cases of one or the other end of an utterance being ity that we suspect is important is matching the
cut off. For example, when the user says “search utterance dalogue-move type (e.g. wh-question,
for the red car” but the system hears “for the red yesho-question, command) otherwise the user is
car”. We use information from the dialogue sys- likely to be misled into thinking that a particular
tem’s parsing grammar (which has identical cover- type of dialogue-move is impossible in the system.
age to its speech recognizer) to determine whether Looking for in-coverage words is fairly robust.
the initial word recognized for an utterance is a Even when the user produces an out-of-coverage
valid initial word. If not, the utterance is diag- utterance they are likely to produce some in-
nosed as a case of the user pressing the push-to- coverage words. The Targeted Help agent looks
talk button too late and the system reports that to for within-domain words in the recognition hy-
the user. Out-of-vocabulary items that can be iden- pothesis from the secondary SLM recognizer. This
tified by Targeted Help are those that are in the gives us a set of target words from which to
SLM’s vocabulary but are out of coverage for the match the example to the dialogue-move type of
grammar based recognizer and so cannot be pro- the user’s utterance: wh-question, yn-question, an-
cessed by the dialogue system. For these items swer, or command.
Targeted Help produces a message of the form Furthermore, for commands (which are a large
“the system doesn’t understand the word X’. percentage of the utterances) we use the in-
Saying “Zoom in on the red car” when the sys- coverage words to produce a targeted in-coverage
tem only has intransitive “zoom in” is an exam- example that is interpretable by the system. These
ple of a subcategorization error. In these cases the examples are intended to demonstrate how in-
word is in-vocabulary hit has been used in a way vocabulary words from the backup recognizer hy-
that is out-of-grammar. To diagnose subcatego- pothesis could be successfully used in commu-
rization errors we consult the recognitiodparsing nicating with the system. For example, if the
grammar for subcategorization information on in- user says something like “fly over to the hospi-
vocabulary verbs in the secondary recognizer hy- tal”, where “over” is out-of-coverage, and the fall-
pothesis, then check what else was recognized to back recognizer detected the words “fly” and “hos-
detennine if the right arguments are there. For pital”, the Targeted Help agent could provide an
these types of errors the system produces a mes- in-coverage example like “fly to the hospital”. For
sage such as “the system doesn’t understand the the other less frequent utterance types we have one
word X used with the red car”. These diagnostics in coverage example per type. The system cur-
are one significant difference from the approach rently uses a look-up table but we hope to incor-
used in (Gorrell et al., 2002). The simple classi- porate generation work which would support gen-
fier approach used in that work to select example eration of these examples on the fly from a list of
sentences would not support these types of diag- in-coverage words (?).
nostics.
3 Design of Experiments
In constructing examples that are similar to the
user’s utterance one issue is in what sense they In order to assess the effectiveness of the targeted
should be similar. One aspect we have looked help provided by our system, we compared the
at is using in-coverage words from the user’s ut- performance of two groups of users, one that re-
terance. It is likely to help naive users learn ceived targeted help, and one that did not. Twenty
the coverage of the system if the examples give members of the Stanford University community
them valid uses of in-coverage words they pro- were randomly assigned to one of the two groups.
duced in their utterance. By using words from the There were both male and female subjects, the ma-
user’s utterance the system provides both confir- jority of subjects were in their twenties and none
mation that those words are in coverage and an in- of the subjects had prior experience with spoken
coverage pattern to imitate. We believe that this dialogue systems. The structure of the interaction
leads to greater linguistic alignment between the with the system was the same for both groups.
user and the system. Another aspect of similar- They were given minimal written instruction on
how to use the system before the interaction be- marked in the first task than in the fifth task
gan. They were then asked to use the system to (LARGER EARLY EFFECT).
complete five tasks, in whch they directed a heli-
4 Experimental Results
copter to move within a city environment to com-
plete various task oriented goals whch were dif-
We found clear evidence that targeted help im-
ferent for four of the five tasks. For each task the
proves performance in this environment, as mea-
goals were given immediately prior to the start of
sured by both the frequency with which th- user
p
the interaction, in language the system could not
simply explicitly gave up on a task, and the time
process to prevent users from simply reading the
to complete the remaining tasks. In this section we
goal aloud to the system. A given task ended when
present the statistical analyses of the experiment.
one of the following criteria was met:
For the following analyses two subjects, both in
the No Help condition, were excluded from the
1. the task was accurately completed and the
analyses because they gave up on every task, leav-
user indicated to the system that he or she had
ing 9 users in each of the two help conditions. Ex-
finished,
ceptions are noted.
We begin by examining the percentage of trials
2. the user believed that the task was completed
in which users explicitly gave up on a task before
and indicated this to the system when in fact
it was completed. We compared the percentage of
the task was not accurately completed, or
trials in which tie user ciicked the “give up” but-
3. the user gave up. ton in both tasks for users in both help conditions.
As predicted, a i-within (Task), i-between (Help
The first and last of the sequence of five tasks condition) subjects ANOVA revealed a main effect
were the critical trials that were used to assess per- of the help condition (F1(1,16)=6.000, p<.05).
formance. Both of the tasks had goals of the form Users who received targeted help were less likely
“locate an x and then land at the y” The experi- to give up than those who did not receive help, par-
ment was conducted in a single session. An exper- ticularly during the first task (11% vs. 27%). If
imenter was present throughout, but when asked we include the two subjects in the No Help con-
she refused to provide any feedback or hints about dition who gave up on every task the difference is
how to interact with the system. even more striking. For the first task only 1I % of
As stated above, the critical difference between the users who received help gave up, compared to
the two groups of users was the feedback they re- 45% of the users who did not receive help. The
ceived during interaction with the system. When pattern holds up even if we include the three in-
the users in the No Help condition produced out- tervening filler trials along with the experimental
of-coverage utterances the system responded only trials, as demonstrated by a paired t-test item anal-
with a text display of the message “not recog- ysis (t(4) = 7.330, p<.05). Those who received
nized”. In contrast, when users in the Help condi- help were less likely to explicitly give up even on
tion produced ogt-of-coverage utterances they re- this wider variety of tasks.
ceived in-depth feedback such as: “The system We next examine the time it took users to com-
heardjy betweerr the hospital and the sclzool, ~ n - plete the individual tasks. Here it is necessary to
fortunately it doesn’t understand JEy when used be clear about what is meant by “completion.” It
with the words between the hospital and the is more ambiguous than it may seem. Each task
school. You could try sayingJEy to the hospital.” had several sub-goals, and it was even difficult
We hypothesized that: 1) providing Targeted to objectively evaluate whether a single sub goal
Help would improve users’ ability to complete had been met. For instance, the goal of the first
tasks (HIGHER TASK COMPLETION); and 2) time task was to find a red car near the warehouse and
to complete tasks would be reduced for users re- then land the helicopter. Users tended to indicate
ceiving Targeted Help (REDUCEDT IME). We that they had finished as soon as they saw the red
also anticipated that both effects would be more car, failing to land the helicopter as the instruc-
tions specified. Another common source of ambi-
Lenient Criterion Analysis
guity was when the user saw the car on the map
but never brought it up in the dialogue, simply 3IQ
landmg the helicopter and clicking “finished.” The
problem with this is that there is no way of know-
ing whether the user actually saw the car before
cliclung finish, and there was no explicit record
that they were aware of its presence. For all tri-
t :
als the experimenter evaluated the task comple-
tion, recording what was done and what was left
undone. According to the experimenter, in most
cases of potential ambiguity the basic goal was
Figure 2: Time to complete task under Lenient
completed. In a few instances, however, the user
Criterion for completion
indicated belief that the task had been completed
when it obviously had not. An example of this is
the following: The goal specified was to find a red than the last one (365.5 seconds vs. 220.4 sec-
car near the warehouse and then land. The user onds), and the difference between the help and no
flew the helicopter to the police station, and then help conditions was more marked on the first task
clicked “finished,” ending the task. Wc dealt with than on the iast one (150.2 seconds vs. 94 sec-
the ambiguity problem by analyzing the time to onds). Figure 2 shows these results.
complehr, dat2 separately acccrdicg to tvm dif- One criticism of this analysis is that it may in-
ferent inclusion criteria. In both cases the pattern
clude trials in which the task objectives were not
was the same: Users who received help took less accurately completed before the subject clicked
time to complete tasks than those who did not, the
“finished”. We wished to avoid experimenter sub-
first task took longer to complete than the last one,
jectivity with respect to task completion, so we
and the difference between the help and no help
conducted another analysis using the strictest in-
conditions was more marked on the first task than
clusion criterion the experimental design allowed.
on the last one.
In this analysis we included only those trials in
In the first analysis we included all trials in which all task objectives were completed and
which the user clicked the “finished” button, re- could be verified using the transcripts. This meant
gardless of their actual performance. Subjects who that for all of the trials we included, the goal entity
failed to complete one of the two critical tasks was explicitly mentioned ir, the dialogue. Accord-
(tasks 1 and 5) were excluded from the analysis. ing to this criterion only 44% of users in the Help
We used a 1-within (Task), 1-between (Help con- condition and 18% of users in the No Help con-
dition) subjects ANOVA. For task 1, 89% of the dition completed the first task. Similarly, 89% of
trials in the Help condition and 55% of the trials users in the Help condition and 40% of users in the
in the No Help were considered ”completed.” For No Help condition accurately completed the task.
task 5, 100% of the trials in the Help condition and Although this analysis is conducted on sparse data,
80% of the trials in the No Eieip condition were it provides strong supporting evidence for the data
considered “completed.” The analysis revealed a pattern observed in the more lenient analysis.
marginally significant main effect of the help con- We examined the time it took to complete tasks
dition (Fl(1,ll) = 3.809, p<.l), a main effect of according to the strict criterion, excluding all other
task (F1,~1=62.545p, <.001) and a help condition trials. The ANOVA analysis was identical to the
by task interaction (F1(1,11)=10.203, p < .OS). previous one. It, too, revealed a main effect of
The effects were in the predicted direction. Users help condition (F1(1,3) = 15.438, p<.05), a main
who received help took less time to complete tasks effect of task (F1,~=83.512,p < .Ol), and a help
than those who did not (290.4 seconds vs. 440.6 condition by task interaction (F1( 1,3)=20.335, p
seconds), the first task took longer to complete < .OS). Again the effects were in the predicted di-
.
These findings are remarkable because they
Strict Criterion Analysis demonstrate that it is possible to construct ef-
51)s ---__ --____-- fective Targeted Help messages even from fairly
1511 1 low quality secondary recognition. Moreover, the
, I
400 study suggests that such an approach can improve
- .
1
the speed of training for naive users, and may re-
I m{iP
DXir llrlp sult in lasting improvements in the quality of their
understanding.
6 Future Work
This work suggests many interesting directions for
Figure 3: Time to complete task under Strict Cn- further research. One area of investigation is the
tenon for completion contribution of various factors in the effectiveness
of the Targeted Help message for example:
rection. Users who received help took less time to
What benefit is due to the online nature of the
complete tasks than those who did not (226.2 sec- 0
help?
onds vs. 377.5 seconds), the first task took longer
to coqlete than the last one (379.9 secoyds vs.
What benefit is due to the information con-
223.75), and the difference between the help and 0
tent?
no helm cocditims was mcre mzked en *e first
task than on the last one (190.4 seconds vs. 112.3
What is the relative contribution of the vari-
0
seconds). These results are shown in Figure 3.
ous parts of the Targeted Help message to the
improvement in user performance.
5 Conclusions
- Is the diagnostic alone more or less ef-
We have shown that users benefit from having on-
fective than the example alone?
line Targeted Help. Naive users who received
Targeted Help messages were less likely to give - How much does getting the back up rec-
up and significantly faster to complete tasks than ognizer hypothesis help the user?
users who did not. Overall, those who did not - What is the most effective combination
receive help gave up on 39% of the trials, while of these components?
those who received our Targeted Help only gave
up on 6% of the trials. With respect to time, Another interesting direction is to look at effec-
when we considered all trials in which the user tiveness across different types of applications. The
indicated that the goal had been completed (re- fact that we found positive results in this domain
gardless of performance), those users who did not and that (Gorrell et al., 2002) also found a variant
receive our Targeted Help took 53% longer than of Targeted Help useful on a quite different do-
those who did. Under stricter inclusion criteria, main suggests that the approach could be generally
whch required the users to explicitly mention the useful for a variety of types of dialogue systems.
goal and accurately complete the task, the differ- We are currently looking at porting our Targeted
ence was even more pronounced. Those users who Help agent to additional domains.
did not receive help took 67.0% longer to com-
plete the tasks than those who received our Tar- Acknowledgements
geted Help. In both help conditions, performance
improved over the course of the experimental ses- This work was partially funded by the Wallenberg
sion. However, the advantage conferred by help Foundation’s WITAS project, Linkoping Univer-
merely diminished and did not disappear during sity, Sweden.
the session.
References
J. Dowding, M. Gawon, D. Appelt, L. Cherny,
R. Moore, and D. Moran. 1993. Gemini: A nat-
ural language system for spoken language under-
standing. In Proceedings of the Thirty-First Annual
Meeting of the Association for Computational Lin-
guistics.
G. Gorrell, I. Lewin, and M. Rayner. 2002. Adding
intelligent help to mixed-initiative spoken dialogue
systems. In Proceedings of the Seventh Intema-
tional Conference on Spoken Language Processing
(ZCSLP),D enver, CO.
B.A. Hockey, G. Aist, J. Dowding, and J. Hieronymus.
2002. Targeted help and dialogue about plans. In
Proceedings of the 40th Annual Meeting of the Asso-
ciation for Computational Linguistics (demo track),
Philadelphia, PA.
S. Knight, G. Gorrell, M. Rayner, D. Milward, R. Koel-
ing, and I. Lewin. 200 1. Comparing grammar-based
and robust approaches to speech understanding: a
case study. In Proceedings of Eurospeech 2001,
pages 1779-1782, Aalborg, Denmark.
D. Martin, A. Cheyer, and D. Moran. 1998. Build-
ing distributed software systems with the open agent
architecture. In Proceedings of the Third Intema-
tional Conference on the Practical Application of In-
telligent Agents and Multi-Agent Technology, Black-
pool, Lancashire, UK.
Nuance, 2002. http:Nwww.nuance.com. As of 1 Feb
2002.
The Festival Speech Synthesis Systems, 2001.
http://www.cstr.ed.ac.uk/projects/festival. As of 28
February 2001.