Natural Language Interface Put in Perspective:
Interaction of Search Method and Task Complexity
QianYing Wang, Jiang Hu and Clifford Nass
Department of Communication, Stanford University,
Stanford, California, USA
Abstract. A 2x2 mixed design experiment (N=52) was conducted to examine
the effects of search method and task complexity on users’ information-seeking
performance and affective experience in an e-commerce context. The former
factor had two within-participants conditions: keyword (KW) vs. natural lan-
guage (NL) search; the latter factor had two between-participants conditions:
simple vs. complex tasks. The results show that participants in the complex task
condition were more successful when they used KW search than NL search.
They thought the tasks were less difficult and reported more enjoyment and
confidence with KW search. In the meantime, simple task participants per-
formed better when they used NL rather than KW search. They also perceived
the tasks as easier and more enjoyable, and had higher levels of confidence
with the results, when NL was used. The findings suggest that NL search is not
the panacea for all information retrieval tasks, depending on the complexity of
task. Implications for interface design and directions for future research are dis-
cussed.
1 Introduction
From punch cards to keyboards to graphical user interface (GUI) to voice interface
(VUI), computer interfaces have evolved to allow increasingly intuitive and natural
interactions between users and computers. Among all the breakthroughs and im-
provements, the use of natural language (NL) as a means of input and output during
human-computer interaction (HCI) is one of the most researched areas. The promis-
ing future of NL-based conversational interfaces (especially with the presence of
computer agents) has been widely lauded by visionaries such as Brenda Laurel [1].
Although no one has yet to be able to claim complete success in natural language
generation and processing, progress is being made every day: from text-based soft-
ware agent (e.g., Microsoft
TM
Clippy) to speech-recognition customer services auto-
mation (e.g., United Airlines’ flight information hotline), NL-based technologies have
advanced into our daily life.
With the explosion of information brought by the Internet and computers in differ-
en
t forms, information retrieval has become a hot issue that concerns researchers and
general users alike. Along with the phenomenal increase of data storage capability,
information retrieval interfaces have become a research focus that is gaining more
Wang Q., Hu J. and Nass C. (2005).
Natural Language Interface Put in Perspective: Interaction of Search Method and Task Complexity.
In Proceedings of the 2nd International Workshop on Natural Language Understanding and Cognitive Science, pages 3-12
DOI: 10.5220/0002566000030012
Copyright
c
SciTePress
attention than ever before. In the HCI research community, the literature about seek-
ing and interacting with information is incredibly rich with design principles and
maxims [2-11]. However, most of those usability research projects were focused on
keyword (KW) search. In the meantime, NLP researchers have focused on building
robust knowledge based system [12-16]; little attention has been given to interface
issues.
Natural language search has been made possible in some specialized areas such as
U.S. legal materials [17] as well as in everyday contexts such as askjeeves.com [18].
Users may type full sentences when natural language search is available. Boolean
search, on the other hand, has thus far been the dominant means of information re-
trieval. However, our observation and informal interviews reveal that average users
are not familiar with Boolean search. Most of the time what they use are just key-
words or key phrases without operators such as “+” and “/”. For an average user,
“Boolean search” quite possibly equals to “keyword search.”
The most relevant piece of literature to the present paper is Turtle’s [19] study
comparing natural language and Boolean query. In two experiments comparing in-
formation retrieval performance from collections of full-text legal materials, Turtle
found that searchers using a then current-generation natural language system were
more successful than expert users who relied on Boolean search. The results seem to
herald the proliferation of natural language search systems.
However, there is one major problem with Turtle’s studies: the natural language
queries were not free-will sentences made up by the searchers. Instead, all “natural
language” queries were guided by a set of issue statements made available to the
searchers. It is quite likely that these “trained” queries boomed the performance of
natural language understanding. As a result, the conclusion of Turtle’s study has to be
interpreted carefully.
In this paper, we attempt to compare natural language and keyword search within a
more everyday context: e-commerce customer support, with untrained average users.
We are also interested in how task complexity affects users’ interaction with informa-
tion retrieval interface. Our hypothesis is that increased task complexity will signifi-
cantly undercut users’ performance with natural language system, because it is diffi-
cult for users to turn complicated search requests into concise NL queries that can be
understood by the computer.
2 Method
2.1 The Two Interfaces
The natural language search interface was the AIM chat-room, with a natural lan-
guage agent created by Conversagent, Inc. [2]. Popular search interfaces such as
askjeeves.com encourages users to type keywords or questions on the same webpage.
However, we believe that online chat interfaces such as that of AIM is much more
conducive to natural language input than is a small text box on a webpage. Agents
available for AIM users included My eBay Buddy, AgentBaseball, EllegirlBuddy and
SmarterChild. My eBay Buddy answers frequently asked questions about buying and
4
selling on eBay. AIM users can add these agents to their buddy lists so as to chat with
them and acquire information from them.
Fig. 1. Screenshots of AIM chat window with My eBay Buddy, and the eBay Help website
For the keyword search condition, we used the eBay Help website, the parallel
website to My eBay Buddy. Figure 1 shows the two interfaces side by side. Search
results provided by My eBay Buddy were presented in a conversational manner while
eBay Help website returned results in a typical list of ranked links. We are fully
aware of the differences between the two presentation formats. However, these forms
of presentation are essential parts of the two radically different interfaces and should
not be studied separately. Behind the interfaces, the eBay Help website and My eBay
Buddy shared the same database. That is to say, two identical queries with the two
interfaces would get the same information.
2.2 Search Tasks
There were two task categories: complex and simple. Complex search tasks included
eight questions concerning eBay transactions; the number of simple search tasks was
20. Compared to simple tasks, complex tasks were more difficult with either NL or
KW interface. Finishing a complex task often involved integrating information from
different aspects related to the task. For example, one complex task was to find out
how much people must pay eBay if they were selling a $35 merchandise. Users had to
find information about listing fee, final value fee, eBay picture service fee, etc., and
combine the information for a correct answer. Our pre-test indicate that people spend
about the same amount of time to finish eight complex tasks as to finish twenty sim-
ple tasks.
The following are two of the simple tasks used in the present experiment:
- Please find out what the gift icon means.
- Please find out what the different colored stars mean.
In the meantime, complex tasks look like these:
- Please find out if it is legal to place a bid right before the auction closes.
5
- Please find out what you should do if you don’t want a certain person to bid
on your item.
For a full list of search tasks used in the present study, please visit
http://www.stanford.edu/~wangqy/projects/tasks.htm.
2.3 Participants
All native speakers of English, 52 college students (N=52) from an introductory com-
munication class participated in the experiment for course credit. They were ran-
domly assigned to all conditions, with gender balanced across conditions. None of the
participants had ever sold anything or bought more than three items on eBay.com.
Experiences with AIM were also balanced across conditions. All participants signed
informed consent forms upon arrival at the lab and were debriefed upon the comple-
tion of the experiment.
2.4 Procedure
The experiment had a 2x2 mixed design, with search complexity (simple vs. com-
plex) as the between-participants factor and search interface (keyword vs. natural
language search) and the within-participants factor. Participants performed search
tasks in a research laboratory equipped with personal computers. Upon arrival to the
laboratory, each participant was seated and assigned to a computer with both Internet
Explorer and AIM.
After reading instructions on–line, participants were given a list of either simple or
complex tasks about buying or selling items on eBay. For half of the tasks, partici-
pants used keyword search on the eBay Help website (i.e., KW); for the other half,
participants conversed with My eBay Buddy (i.e., NL) to find the answers. Half of
the participants started with KW and the other half started with NL. Participants
typed in an answer when they thought they had found one, indicating whether they
believed that they had found the correct one. Upon finishing each task, participants
responded to several questions about the finished task.
2.5 Measures
Actual performance was computed as the percentage of questions that the user an-
swered correctly. Perceived performance was computed as the percentage of ques-
tions that the users believed that they had answered correctly. Time on task was the
average amount of time completing each search.
Questions concerning perceived task difficulty, pleasure, and confidence in work-
ing with the search interface were asked for each task. Participants used radio but-
tons to indicate their responses for these questions. Each question had an independ-
ent, 10-point Likert scale. Perceived task difficulty was an average across searches of
responses to the questionnaire item, “How difficult did you feel the search was?”
Enjoyment was an average across search questions of responses to the questionnaire
6
item, “How enjoyable did you find the search?” Confidence was an average across
search questions of responses to the questionnaire item, “How confident were you
with your answer?
3 Results
3.1 Manipulation Check
Time to finish complex tasks and simple tasks were recorded during the study. Con-
sistent with our manipulation, complex searches took longer than simple search for
both keyword [F(1, 24)=107.6, p<.001] and natural language [F(1,24)=55.6, p<.001;
see Table 1]. There was no statistical difference for time on task between the keyword
search and natural language search (p>.05).
Table 1. Time on task
(minutes) Using NL Using KW
Simple 1.21 1.33
Complex 3.8 4.17
3.2 Actual vs. Perceived Search Performance
Figure 2 shows results for actual and perceived performance. There was a significant
interaction between task complexity and search interface for both perceived
[F(1,50)=31.4, p<.001] and actual performance [F(1,50)=26.5, p<.001]. Complex
task participants were more successful in finding the needed information with KW
than with NL; they also perceived more success in the KW condition. Conversely,
simple task participants were more successful with NL than with KW, and they per-
ceived more success in the NL condition.
0%
20%
40%
60%
80%
100%
Comp
le
x
/N
L
Co
mp
lex
/
K
W
Si
mp
le/N
L
Si
mp
le/K
W
Per cei ved
Act ual
Fig. 2. Accuracy of search results
7
Simple task participants obtained [M
simple
=89%, M
complex
=62%, F(1,50)=50.1,
p<.01] and believed they obtained [M
simple
=93%, M
complex
=76%, F(1,50)=41.6, p<.01]
more correct answers than complex task participants. Not surprisingly, there was a
tendency for participants to over-estimate their actual performance. Though there was
no significant difference between the two search methods for actual success
[M
KW
=79%, M
NL
=73%, F(1,51)=3.24, p>.05], participants thought they obtained
more correct answers from keyword search than from natural language queries
[M
KW
=80%, M
NL
=89%, F(1,51)=7.61, p<.01].
3.3 Perception of Task Difficulty, Task Enjoyment, and Confidence with
Answers
Perceived Task Difficulty. We first analyzed participants’ perception of difficulty
without including performance results as covariates. There was a significant
interaction effect between task complexity and search interface on participants’
perception of difficulty [F(1,50)=105.0, p<.001]. Complex task participants thought
KW search was easier to work with than was NL; conversely, simple task participants
thought NL was easier to work with than was KW search (see Figure 3). High
complexity tasks were perceived to be more difficult than were low complexity tasks
(M
simple
=3.20, M
complex
=4.95, F(1,50)=36.4, p<.001). No main effect was found for
search interface [F(1,50)=.917, p>.34].
3. 6
4. 3
2. 09
6. 3
1
2
3
4
5
6
7
Complex Tasks Simple Tasks
Keyword
NL
Fig. 3. Perception of difficulty
We re-examined the perceived difficulty data with actual performance as covari-
ates. The analysis reaffirmed our finding of an interaction effect between task com-
plexity and search interface on perceived difficulty [F(1,50)=28.95, p<.001]. That is,
even after adjusting the actual performance score, keyword search was preferred for
complex tasks and natural language was preferred for simple tasks.
8
Task Enjoyment and Confidence with Answers. Participants’ perception of
enjoyment and confidence with the search interfaces were analyzed. Table 2 presents
the means.
Table 2. Perception of enjoyment and confidence
Complex Task Simple Task
NL KW NL KW
Enjoyment 4.07 5.43 6.9 4.84
Confidence 5.84 8.03 8.71 7.65
There was an interaction effect between interface and task complexity with respect
to task enjoyment [F(1,50)=47.6, p<.001]. Participants found the tasks to be least
enjoyable while working on complex tasks using NL and most enjoyable while work-
ing on simples tasks using NL. The effect remained after controlling for actual per-
formance [F(1,50)=19.9, p<.001]. No main effect was found for search interface on
participants’ enjoyment [F(1,50)=2.01, p>.05]. Overall, complex task participants
found the tasks to be less pleasant than simple task participants [M
simple
=5.87, M
com-
plex
=4.75, F(1,50)=5.61, p<0.05].
Consistent with participants’ perception of enjoyment, there was an interaction for
confidence with answers (even after controlling for actual performance): Participants
were least confident while working with complex tasks using NL and most confident
while working with simple tasks using NL [F(1,50)=70.9, p<.001; control:
F(1,50)=13.1, p<.001].
Participant were more confident while working with KW search than with NL que-
ries [F(1,50)=8.56, p<.01]. Interviews with participants during debrief sessions con-
cerning confidence are discussed in the following section. Not surprisingly, complex
tasks made user feel less confident than simple tasks [F(1,50)=15.4, p<.001].
4 Discussion
Results of the present study suggest that the selection of search interface has impor-
tant and systematic effects on user performance and perceptions. However, these
effects are conditioned by the complexity of search task and the nature of queries.
The general pattern we found was that NL was better (in both reality and perception)
than KW search for simple search tasks, while KW search was better (in both reality
and perception) than NL for complex search tasks. The type of search interface had
no direct effect on either perceived task difficulty or enjoyment, although participants
were more confident with KW than with NL search. During debriefing, several par-
ticipants said that they felt more confident with KW because they had browsed pages
or links that gave them peripheral affirmation. It is important to note that although
there was no difference in terms of actual search performance, participants believed
that they obtained more correct answers by using the KW interface than the NL inter-
face.
9
4.1 Task Complexity Effect and Implications
One clear problem with NL search is that even good NL systems have problems un-
derstanding users. This was true for almost all dialogs between complex task partici-
pants and the Buddy. When the understanding failed, My eBay Buddy revealed its
inability to understand the participant by asking questions back to the participant, thus
giving the participant an impression that he/she was engaged in a conventional con-
versation with the Buddy as if the Buddy was another person.
In a conventional (i.e., human-human) conversation, people adjust their communi-
cation strategies by asking and responding to questions in order to establish some
“common ground” so as to enhance understanding [21]. When a computer agent tries
to talk like a real person, the user would subsequently raise his/her expectations of the
agent’s ability to understand. Some previous research [22, 23] has suggested that
users may say things more freely when they have high expectations of the agent than
when they have low expectations. The freer utterances from the user consequently
impose extra burden on natural language processing system. The gap between uncon-
strained user utterances and the limited understanding capability of NL agent thus
frustrates the user and lowers user satisfaction.
When using the Buddy in the present study, participants with complex tasks in
general found it difficult to form concise questions. As a result, they had to ask more
questions than did simple task participants. Given the limited natural language proc-
essing capability of the Buddy, the more questions participants asked, the more likely
the Buddy would fail to understand. Also, the more unconstrained the questions were,
the more likely the Buddy would appear ignorant. It would be a great idea if the agent
is able to constrain user utterances by leveraging the “alignment” phenomenon ob-
served in human-human dialogs [24].
On the other hand, under the simple task condition where the NL agent was able to
understand and to offer relevant responses, user satisfaction was high. In fact, during
our debrief sections, several simple task participants mentioned that they felt the
Buddy was really smart when it asked questions back to confirm its understanding of
participants’ queries. Under such circumstances, it seemed to be acceptable for the
agent to behave like a human being.
The above findings suggest that a natural language agent should adapt its response
style according to its confidence level in understanding user input. A low confidence
level means that the agent should reduce sentential responses because large amounts
of sentential responses may mislead the user to expect the same intelligent output
from the agent as from a human conversant. The user’s increased level of expectation
may lead to a decrease in satisfaction with the agent if the agent performs poorly.
4.2 Future Research Directions
In the present study, participants were not given the freedom to choose between the
two interfaces; failure rates on complex tasks were high for both NL and KW search.
One potential direction for future research is to investigate the interchange between
KW and NL interfaces. Task performance and perceptions are likely to be different
when users are able to switch from one interface to the other if they think that the
10
initial interface is ineffective. The research question here is whether or not combining
two search interfaces/methods could provide better user experience with information
retrieval systems.
As noted earlier, it is important to understand how the response style of the agent
influences user behaviors and attitudes. Some earlier studies have demonstrated that
linguistic variations such as sentence length may affect users’ input by soliciting
alignment (i.e., mirroring) behaviors from them [22, 25]. However, linguistic align-
ment is a two-way process. How users would evaluate an aligning agent versus a non-
aligning agent is still to be explored. On top of that, researchers may further deter-
mine when and how a computer agent should align with the user to achieve or im-
prove user satisfaction.
4.3 Final Words
The design of search methodology requires an understanding of the complex interac-
tion between technology, psychology, and context. While one could have a reason-
able argument over whether KW search is better than NL-based information retrieval,
or vice versa, the present study shows that the “KW vs. NL” argument is inadequate
without reference to the particular task and the model of search that the user brings to
the task.
References
1. Laurel, B.: Interface Agents: Metaphors with Character. In Laurel, B. (ed.), The Art of
Human-Computer Interface Design. Addison-Wesley, Reading, MA (1990)
2. Blackmon, M.H., Polson, G.P., Kitajima, M, Lewis C.: Design Methods: Cognitive Walk-
through for the Web. Proceedings of CHI 2002. ACM Press, 463 – 470
3. Borgman, C.L., Belkin, N.J., Croft, W.B., Lesk, M.E., Landauer, T.K.: Retrieval Systems
for the Information Seeker: Can the Role of Intermediary be Automated? Proceedings of
CHI '88 (Washington, D.C.), ACM Press, 51–53
4. Card, S., Pirolli, P., Van Der Wege, M., Morrison, J., Reeder, R., Schraedley, P., Boshart,
J.: Information Scent as a Driver of Web Behavior Graphs: Results of a Protocol Analysis
Method for Web Usability. Proceedings of CHI 2001. ACM Press, 498–505
5. Chi, E., Pirolli, P., Chen, K., Pitkow, J. E.: Using Information Scent to Model User Needs
and Actions on the Web. Proceedings of CHI 2001. ACM Press, 490–497
6. Larson, K., Czerwinski, M.: Web Page Design: Implications of Memory, Structure and
Scent for Information Retrieval. Proceedings of CHI '98.
ACM Press, 25–32
7. Nielsen, J.: Designing Web Usability. New Riders, Indianapolis, IN (1999)
8. Pirolli, P.: Computational Models of Information Scent-Following in a Very Large
Browsable Text Collection. Proceedings of CHI '97. ACM Press, 3–10
9. Pirolli, P., Card, S. K.: Information Foraging. Psychological Review 106(4) (1999), 643–
675
10. Sellen A.J., Murphy R., Shaw K.L.: How Knowledge Workers Use the Web. Proceedings
of CHI 2002. ACM Press, 227–234
11. Spool, J.M., Scanlon, T., Schroeder, W., Snyder, C., DeAngelo, T.: Web Site Usability.
Morgan Kaufman, San Francisco, CA (1999)
11
12. Doszkocs T.: Natural Language Processing in Intelligent Information Retrieval. Proceed-
ings of the 1985 ACM Annual Conference on The Range of Computing: Mid-80's Perspec-
tive. ACM Press, 356–359
13. Guglielmo E.J., Rowe N.C.: Natural-Language Retrieval of Images Based on Descriptive
Captions. ACM Transactions on Information Systems (TOIS), July 1996, Vol. 14, Issue 3,
237–267
14. Jacob, P. S., Rau, L.F.: Natural Language Techniques for Intelligent Information Retrieval.
Proceedings of the Eleventh International Conference on Research & development in In-
formation Retrieval (1988), 85–99
15. Meng, F.: A Natural Language Interface for Information Retrieval from Forms on the
World Wide Web. Proceeding of the 20th International Conference on Information Systems
(1999), 540–545
16. Vivacqua A., Lieberman H.: Agents to Assist in Finding Help. Proceedings of CHI 2000.
ACM Press, 65–72
17. Griffith, C.: WESTLAW’s WIN: Not Only Natural, But New. Information Today, 9-11,
October 1992.
18. Ask Jeeves, Inc. http://www.ask.com
19. Turtle H.: Natural Language vs. Boolean Query Evaluation: A Comparison of Retrieval
Performance. Proceedings of the Seventeenth Annual International ACM-SIGIR Confer-
ence on Research and Development in Information Retrieval (1994), 212–220
20. Conversagent, Inc. http://www.conversagent.com
21. Clark, H.H.: Using Language. Cambridge University Press, New York (1996)
22. Brennan, S.: Conversation with and through Computers. User Modeling and User-Adapted
Interaction, 1 (1991), 67–86
23. Pearson, J., Pickering, M.J. Branigan, H.P., McLean, J.F., Nass, C.I., Hu, J.: The Influence
of Beliefs about an Interlocutor on Lexical and Syntactic Alignment: Evidence from Hu-
man-Computer Dialogues. Proceeding of the 10
th
Annual Architectures and Mechanisms of
Language Processing Conference (2004)
24. Nass, C.I., Hu, J., Pearson, J., Pickering, M.J., Branigan, H.P.: Linguistic Alignment in HCI
vs. CMC: Do Users Mimic Conversation Partners’ Grammar and Word Choices? Unpub-
lished manuscript (2004).
25. Zoltan-Ford, E.: How to Get People to Say and Type What Computers Can Understand.
International Journal of Man-Machine Studies, 34(4) (1991), 527-547
12