Content uploaded by Daniel Schloß
Author content
All content in this area was uploaded by Daniel Schloß on Aug 08, 2023
Content may be subject to copyright.
18th International Conference on Wirtschaftsinformatik,
September 2023, Paderborn, Germany
Towards Designing a NLU Model Improvement System
for Customer Service Chatbots
Research in Progress
Daniel Schloß1, Ulrich Gnewuch1 and Alexander Maedche1
1 Karlsruhe Institute of Technology, human-centered systems lab, Karlsruhe, Germany
{daniel.schloss,ulrich.gnewuch,alexander.maedche}@kit.edu
Abstract. Current customer service chatbots often struggle to meet customer ex-
pectations. One reason is that despite advances in artificial intelligence (AI), the
natural language understanding (NLU) capabilities of chatbots are often far from
perfect. In order to improve them, chatbot managers need to make informed de-
cisions and continuously adapt the chatbot’s NLU model to the specific topics
and expressions used by customers. Customer-chatbot interaction data is an ex-
cellent source of information for these adjustments because customer messages
contain specific topics and linguistic expressions representing the domain of the
customer service chatbot. However, extracting insights from such data to im-
prove the chatbot’s NLU, its architecture, and ultimately the conversational ex-
perience requires appropriate systems and methods, which are currently lacking.
Therefore, we conduct a design science research project to develop a novel arti-
fact based on chatbot interaction data that supports NLU improvement.
Keywords: Customer Service, Chatbots, Analytics, Language Understanding
1 Introduction
Chatbot technology is now widely used in customer service to answer customer ques-
tions or handle concerns end-to-end (Schuetzler et al., 2021).
In the rather short conversations with customer service chatbots, customers place par-
ticular value on performance (Diederich et al., 2022; Følstad and Skjuve, 2019). How-
ever, as known from research and practice, many customers still have negative experi-
ences with chatbots. This is often due to weaknesses in the chatbots’ natural language
understanding (NLU). In the worst case, customers experience a breakdown (“Sorry, I
did not understand that”) or a mismatch (an incorrect or inappropriate response) (Ben-
ner et al., 2021; Følstad and Taylor, 2021). To avoid this, various NLU improvements
performed by chatbot managers are possible: The language model of a customer service
chatbot can be supplemented with additional topics or training data. Additionally, the
architecture of the NLU can be adapted or linguistic fine tuning can be performed (AI
Multiple, 2022). However, with growing options in NLU training, chatbot managers
must make very thoughtful decisions as these can affect the customer service experi-
ence of possibly thousands of customers. However, these decisions do not have to be
made uninformed. As research as well as practical advice on chatbot and NLU devel-
opment suggest, real-world interaction data is valuable for a “conversation-driven de-
velopment” (Akhtar et al., 2019; Rasa, 2023). By “reviewing conversations on a regular
basis” (Rasa, 2023), improvements, especially with regard to NLU and Dialog Man-
agement (DM), can be discovered and conducted (Følstad and Taylor, 2021). However,
it is still largely unknown how a system providing such insights based on real-world
conversation data needs to be designed to enable the discovery of such improvement
opportunities on a mass large scale and in a more automated manner (Chen and Beaver,
2022). To address this gap, we formulate the following research question:
RQ: How ought a system for NLU model improvement
of customer service chatbots to be designed?
Accordingly, the goal of the project presented in this paper is to generate (instantiated)
design knowledge for a user-centric and usage data-based system for NLU model im-
provement. Regarding the usage data, we aim to provide a high degree of automation
(compared to manual analytical approaches, e.g. Kvale et al., 2020). Regarding the us-
ers of the system, we reveal how and what descriptive as well as prescriptive insights
into conversation data help NLU designers and chatbot managers on an aggregate as
well as detailed level to improve their NLU (and DM). Additionally, to ensure practical
relevance and user-centeredness, this research project is conducted as design science
research (DSR) project in collaboration with an industry partner. The setup is presented
in chapter three. Prior to this, we introduce and build on related research from the fields
of chatbot analytics and communication theory. In chapter four we provide a first in-
sight into our project and system. Chapter five concludes with a summary and outlook.
2 Related Work: Chatbots in Customer Service
In customer service, task-focused chatbots are used (Schuetzler et al., 2021). With
these, customers typically have short conversations regarding specific concerns where
they expect to be served fast and effectively (Diederich et al., 2022; Følstad and Skjuve,
2019). They answer questions, e.g. via texts or links or start guided dialogs in which
multiple conversational “turns” takes place, for example, when asking prescripted ques-
tions to identify a user to ultimately process a business transaction.
While open domain chatbots like ChatGPT or Replika build on non-deterministic Nat-
ural Language Generation (NLG), the architecture of customer service chatbots is in-
tent-based at its core (OpenAI, 2023; Replika, 2023; Schuetzler et al., 2021). As Fig-
ure 1 shows, customer utterances are classified to an intent with a probability by a
Natural Language Understanding (NLU) system. Next, the dialog management (DM)
controls the chatbots’ actions based on the NLU information and defined logical rules.
For customer service chatbots, the responses are then typically retrieved from a
knowledge base (KB) (Kucherbaev et al., 2018). This deterministic approach has the
advantage that the responses, which only need to cover certain topics, are fully control-
lable for service managers as opposed to probalistic NLG responses (Galitsky, 2019).
customer
utterance
chatbot
response
In Scope Topics:
Out-Of-Scope Topics
NLU information,
e.g. intent scores
interface
application
Figure 1. Typical Customer Service Chatbot Architecture (Adamopoulou et al., 2020)
However, problems can arise with this language-based classification problem. In chat-
bot research, these are summarized as conversational or chatbot “breakdowns,” i.e. sit-
uations in which the chatbot responds with “Sorry, I can't help you.” (Benner et al.
2021). A large body of research has been devoted to the moderation and resolution of
these situations within a conversation (e.g., Ashktorab et al., 2019), as they cause cus-
tomer dissatisfaction and poor reputation for the technology and the company (Schuetz-
ler et al., 2021; van der Goot et al., 2021). A breakdown is caused by the failure to
succesfully detect a correct intent. In addition, an incorrect response, a so-called (intent)
mismatch can occur as well (or false positive, Følstad and Taylor, 2021). Both types of
errors must be corrected by a chatbot manager. This often requires manual review of
conversations, as well as small adjustments within a large and difficult-to-manage set
of intents, training data, and features (Beaver and Mueen, 2021). Large NLU systems,
e.g., IBM Watson or Microsoft LUIS, do provide rudimentary reporting for NLU revi-
sion (Diederich et al., 2019). However, NLU improvement support is often based only
on internal training data omitting detailed analysis of the real-world conversation data
and specific interpretation capabilities. For this reason, user-friendly chatbot analytics
systems can help to better understand conversational, specifically NLU, problems and
suggest appropriate actions (Beaver and Mueen, 2020; Yaeli and Zeltyn, 2021).
3 Research Method
To address the problem of detecting and adequately resolving NLU problems in cus-
tomer service chatbots, we are designing a data-driven NLU model improvement sys-
tem following the design science research (DSR) paradigm (Gregor and Jones, 2007).
To ensure the relevance of the project besides scientific rigor we collaborate with an
industry partner (Hevner, 2007). Our industry partner is an IT and service provider for
the utility industry, operating and maintaining customer service chatbots for over 25
small to very large utilities. The DSR project, shown in Table 1, follows the established
structure of Kuechler and Vaishnavi (2008) and consists of two design cycles. The first
design cycle (DC1) is dedicated to the descriptive analysis of chatbot log data (user
messages and internal metrics). To this end, we first explored the problem space build-
ing on research on customer service chatbots and NLU. In the suggestion phase, we
explored insights from (chatbot) analytics and communication research to outline meth-
ods for log data analysis for NLU optimization. Moreover, we had access to more
NLU
I1
I
2
I
3
DM KB
600.000 user messages and NLU log data. Based on this, we have drafted a first system
and are currently in the development phase. Currently, it mainly contains descriptive
exploration of chat data for NLU improvement, e.g. filter functions (e.g. according to
intent scores) and corresponding visualizations. In the second step (DC2), we plan to
upgrade the system for prescriptive analysis. By connecting the support system not only
to the log data source but to the NLU system, we finally want to provide usage data-
driven decision support and recommendations for the NLU, e.g. regarding training data.
This will be guided by two evaluations we plan in both design cycles, first a think aloud
study with experts of the industry partner, in DC2 an evaluation including external ex-
perts such as researchers and a case study where we evaluate the NLU performance,
e.g. classification, ex ante and ex post implementing changes suggested by our system.
Table 1. Overview of the Design Science Research Project
DC1: Log Data Analysis DC2: Log – Training Data comparison
1) Problem Awareness
Chatbot
and NLU
Research
2) Suggestion
Theory
and Sample Dataset
Concept
for
Dat
a Integration
of
System
3) Developement
System Prototype
Update: NLU/KB
Dat
a
Integration
4) Evaluation
Expert Think
-
Aloud Sessions
Expert Assessmen
t of System Guidance
5) Conclusion
Refinement and
Planning DC2
Summary of
Design Knowledge
4 NLU Model Improvement System
4.1 Problem Awareness
If the NLU of a chatbot does not understand the customer's concern and the chatbot
does not provide a relevant answer, it creates customer frustration and runs counter to
the goal of automating customer service (Schuetzler et al., 2021; van der Goot et al.,
2021). Therefore, chatbot managers need to continuously improve the NLU. But just
like Large Language Models, intent-based NLU systems are often referred to as a
“black box”. Chatbot managers have only limited insight into the statistical operations
of the NLU offered by the chatbot platforms (Diederich et al., 2019; Schuurmans and
Frasincar, 2019). What they can control, is the selection and partitioning of training
data, as well as the use of specific NLU features (Ruane et al., 2020). Since language
understanding in intent-based (retrieval) chatbots is based on a classification problem,
the performance of the model is often evaluated by means of precision, recall or F1
score, testing training utterances internally in the NLU (Rasa, 2023). Against this back-
ground, there are different reasons why breakdowns (or mismatches) occur. First, it
may be that a customer has simply entered (1) semantic nonsense (e.g., “Colorless green
ideas sleep furiously,” Chomsky and Lightfoot (2007)) or (2) grammatical nonsense
(e.g., “sadsasdsde”) the NLU does not need cover. However, it may also be that he or
she expresses appropriately, but the topic addressed is (3) out-of-scope of the chatbot
(Følstad and Taylor, 2021). In all cases, the actual tokens (e.g. “sleep”) or the semantic
meaning will intentionally be underrepresented in the NLU training data (Debortoli et
al., 2016). It is also possible, as we observed in the provided log data, that customers
use familiar words, but (4) formulate too short or verbose (Beaver and Mueen, 2020;
Reinkemeier and Gnewuch, 2022). If all that is not the case and a breakdown occurred,
it might be that (5a) the training data of the NLU may not be grammatically (i.e. syn-
tactically) sufficient (e.g., “I want to cancel” vs. “Cancellation please) or (5b) lacks
synonyms for equal semantics (e.g., “I want to buy a product” vs. “I want to purchase
a product”) (Jurafsky and Martin, 2009; Sperber and Wilson, 1986). In this case, chat-
bot managers need to add on the training data. Last, the NLU itself may have (6) super-
vised learning performance issues such as imbalances (over-/underfitting) or poorly
weighted linguistic features (Rasa, 2023). All of these causes ultimately lead to poor
user experiences, but can be mitigated incrementally through informed NLU improve-
ments, particularly based on real-world usage data (Diederich et al., 2019).
4.2 Suggestion
What can chatbot managers do to improve their language model to meet customer
needs? To inform the design of our improvement system, we draw on the communica-
tion theories “Grounding in Communication” and “Communication Accommodation
Theory” (Clark, 1996; Giles and Ogay, 2007). These theories propose the main com-
ponents of communication:
Semantical: First, successful communication is characterized by all parties having
a “common ground”, i.e. a shared level of knowledge of the world (Clark, 1996). With
regard to customer service chatbots, customers expect a broad knowledge base on all
their concerns. Therefore an important component of the NLU Model Improvement
System is the discovery of the situations where the chatbot violates the assumed com-
mon ground, see (3). These situations can be used as opportunities to create new intents
and training data (Chen and Beaver, 2022). We have addressed this in the prototype
shown on the next page by means of filter functions for unrecognized utterances and
are also working on fast topic modeling for an easier overview of the unknown topics.
Syntactical: Second, according to “communication accommodation” good commu-
nication is characterized by the fact that a convergence of communication parties hap-
pens in a conversation. This can refer to social aspects or gestures, concerning text-
based chatbots as in this stage of the project, this convergence needs to manifest in the
applied language (understanding), see (4), (5a), (5b) (Giles and Ogay, 2007). Customer-
friendly customer service chatbots must therefore not only choose their answers in the
style of the customer, they also need to acquire the customer's understanding of the
language (semantic and grammatical) over time. For our proposed system, this means
that it must provide insight into the patterns of customers, especially with regard to
utterances that are not yet well understood. This is represented by our initial filters, but
also by linguistic metrics such as most frequent words or part-of-speech.
Structural: Third, there are internal technical premises of NLU systems that affect
their performance independently of the specific semantics or syntax of the chosen train-
ing data, see (6). An example of this is a “skewed” distribution of training data leading
to imbalances (Ruane et al., 2020). We will address the analysis of these structural NLU
improvement opportunities in DC2 when the NLU system is connected to our system.
4.3 Development
Figure 2. Screenshot of the NLU Model Improvement System Prototype
The prototype of the NLU improvement system, developed on the basis of python mod-
ules and connected to a MySQL log database, is built in a top-down design. This allows
for generous filtering and mining of many customer utterances to apply detailed analy-
sis for specific utterances (Beaver and Mueen, 2020). First, timestamps can be selected
(due to NLU adaptations), but also intents can be filtered (left). In addition, the tool
provides an overview on current intent utilization, since a too “long tail of intents” can
also be cause of structural NLU performance problems (Greyling, 2022). Moreover,
chatbot managers have the possibility of selecting the most interesting utterances via
intent scores/thresholds or the text length as the first linguistic feature. After that, they
can examine in-depth structural linguistic metrics of their selection or individual utter-
ances (e.g. n-grams or Part-of-Speech, Manning et al., 2014). In DC2, it is planned to
expand these metrics and compare them to NLU training data according to the customer
service chatbot experts’ input and requirements. Finally, the right-hand side of the sys-
tem offers the possibility to apply parameterizable topic mining to an utterance selec-
tion. In this way, new (in-scope) topics can be discovered (Chen and Beaver, 2022).
5 Conclusion and Outlook
The NLU model improvement system we present addresses the current problem of im-
performant customer service chatbots. We are planning the technical completion of the
system in order to evaluate it with customer service experts in the next DSR stage. The
descriptive and exploratory nature of the system, as applied to log data, will then be
augmented with research and practical expertise to provide a stronger guiding function.
References
Akhtar, M., Neidhardt, J., and Werthner, H. (2019). The potential of chatbots:
analysis of chatbot conversations. In 2019 IEEE 21st conference on business
informatics (CBI) (Vol. 1, pp. 397-404). IEEE.
AI Multiple, (2022). In-Depth Guide Into Chatbots Intents Recognition in 2023.
https://research.aimultiple.com/chatbot-intent/, Accessed: 08.03.2023.
Beaver, I., and Mueen, A. (2020). Automated conversation review to surface
virtual assistant misunderstandings: Reducing cost and increasing privacy.
In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 34,
No. 08, pp. 13140-13147).
Chen, X., and Beaver, I. (2022). An Adaptive Deep Clustering Pipeline to In-
form Text Labeling at Scale. Baltimore, Maryland, USA: Proceedings of the
39 th International Conference on Machine Learning.
Chomsky, N., and Lightfoot, D. W. (2002). Syntactic structures. Walter de
Gruyter.
Clark, H. H. (1996). Using language, Cambridge: Cambridge University Press.
Diederich, S., Brendel, A. B., and Kolbe, L. M. (2019). Towards a Taxonomy
of Platforms for Conversational Agent Design. Internationale Tagung
Wirtschaftsinformatik, Siegen, Germany.
Diederich, S., Brendel, A. B., Morana, S., and Kolbe, L. (2022). On the design
of and interaction with conversational agents: An organizing and assessing
review of human-computer interaction research. Journal of the Association
for Information Systems, 23(1), 96-138.
Debortoli, S., Müller, O., Junglas, I., and Vom Brocke, J. (2016). Text mining
for information systems researchers: An annotated topic modeling tutorial.
Communications of the Association for Information Systems, 39(1), 7.
Følstad, A., & Skjuve, M. (2019, August). Chatbots for customer service: user
experience and motivation. In Proceedings of the 1st international confer-
ence on conversational user interfaces (pp. 1-9).
Følstad, A., and Taylor, C. (2021). Investigating the user experience of cus-
tomer service chatbot interaction: a framework for qualitative analysis of
chatbot dialogues. Quality and User Experience, 6(1), 6.
Galitsky, B. (2019). Developing enterprise chatbots. New York: Springer In-
ternational Publishing.
Giles, H., and Ogay, T. (2007). Communication accommodation theory.
Gregor, S., and Jones, D. (2007). The Anatomy of a Design Theory. Journal of
the Association for Information Systems, 8(5), 312–335.
Greyling, C. (2022). Solving for the long tail of intent distribution. https://co-
busgreyling.medium.com/solving-for-the-long-tail-of-intent-distribution-
24daa372fcc, Accessed: 08.03.2023
Greyling, C. (2023). NLU & NLG Should Go Hand-in-Hand. https://co-
busgreyling.medium.com/nlu-nlg-should-go-hand-in-hand-8cc7952ebab8,
Accessed: 08.03.2023.
Hevner, A. R. (2007). A three cycle view of design science research. Scandina-
vian journal of information systems, 19(2), 4.
Jurafsky, D. and Martin, J. H. (2009). Speech and Language Processing, 2nd
Edition, New Jersey: Prentice-Hall.
Kuechler, B., & Vaishnavi, V. (2008). On theory development in design science
research: anatomy of a research project. European Journal of Information
Systems, 17(5), 489-504.
Kvale, K., Sell, O. A., Hodnebrog, S., & Følstad, A. (2020). Improving conver-
sations: lessons learnt from manual analysis of chatbot dialogues. In Chatbot
Research and Design: Third International Workshop, CONVERSATIONS
2019, Amsterdam, The Netherlands, November 19–20, 2019, Revised Se-
lected Papers 3 (pp. 187-200). Springer International Publishing.
Manning, C. D., Surdeanu, M., Bauer, J., Finkel, J. R., Bethard, S., &
McClosky, D. (2014, June). The Stanford CoreNLP natural language pro-
cessing toolkit. In Proceedings of 52nd annual meeting of the association
for computational linguistics: system demonstrations, pp. 55-60..
Open AI (2023). ChatGPT. https://chat.openai.com/chat, Accessed:
08.03.2023.
Rasa (2023). Conversation-Driven Development. https://rasa.com/docs/ , Ac-
cessed: 08.03.2023.
Reinkemeier, F., and Gnewuch, U. (2022). Designing effective conversational
repair strategies for chatbots. ECIS 2022 Research Papers
Replika (2023), The AI companion who cares. https://replika.ai/, Accessed:
08.03.2023
Ruane, E., Young, R., & Ventresque, A. (2020, March). Training a chatbot with
Microsoft LUIS: effect of intent imbalance on prediction accuracy. In Pro-
ceedings of the 25th International Conference on Intelligent User Interfaces
Companion (pp. 63-64).
Schuetzler, R. M., Grimes, G. M., Giboney, J. S., and Rosser, H. K. (2021).
“Deciding Whether and How to Deploy Chatbots”, MIS Quarterly Execu-
tive 20 (1), 1-15.
Schuurmans, J., and Frasincar, F. (2019). Intent classification for dialogue ut-
terances. IEEE Intelligent Systems, 35(1), 82-88.
Sperber, D. and Wilson, D. (1986). Relevance: Communication and cognition,
Vol. 142. Cambridge, MA: Harvard University Press.
Van der Goot, M. J., Hafkamp, L., and Dankfort, Z. (2021). Customer service
chatbots: A qualitative interview study into the communication journey of
customers. In Chatbot Research and Design: 4th International Workshop,
CONVERSATIONS 2020, Virtual Event, November 23–24, 2020, Revised
Selected Papers 4 (pp. 190-204). Springer International Publishing.
