Enhancing User Experience in Chinese Initial Text Conversations with Personalised AI-Powered Assistant

Abstract

In the rapidly evolving landscape of text-based communication, the importance of the initial interaction phase remains paramount. This study investigates the potential benefits that a proposed AI chat assistant equipped with text recommendation and polishing functionalities can bring during initial textual interactions. The system allows the users to personalise the language style, choosing between humorous and respectful. They can also choose between three different levels of AI extraversion to suit their preferences. Results of user evaluations indicate the system received a "good" us-ability rating, affirming its effectiveness. Users reported heightened comfort levels and increased willingness to continue interactions when using the AI chat assistant. The analysis of the results offers insights into harnessing AI to amplify user engagement, especially in the critical initial stage of textual interaction.

Full text

Enhancing User Experience in Chinese Initial Text Conversations
with Personalised AI-Powered Assistant
Jindi Wang
jindi.wang@durham.ac.uk
Durham University
UK
Ioannis Ivrissimtzis
ioannis.ivrissimtzis@durham.ac.uk
Durham University
UK
Zhaoxing Li
zhaoxing.li@soton.ac.uk
University of Southampton
UK
Lei Shi
lei.shi@newcastle.ac.uk
Newcastle University
UK
ABSTRACT
ABSTRACT
In the rapidly evolving landscape of text-based communication,
the importance of the initial interaction phase remains paramount.
This study investigates the potential benets that a proposed AI
chat assistant equipped with text recommendation and polishing
functionalities can bring during initial textual interactions. The
system allows the users to personalise the language style, choosing
between humorous and respectful. They can also choose between
three dierent levels of AI extraversion to suit their preferences.
Results of user evaluations indicate the system received a “good” us-
ability rating, arming its eectiveness. Users reported heightened
comfort levels and increased willingness to continue interactions
when using the AI chat assistant. The analysis of the results oers
insights into harnessing AI to amplify user engagement, especially
in the critical initial stage of textual interaction.
CCS CONCEPTS
•Human-centered computing
→
User studies;Empirical stud-
ies in HCI.
KEYWORDS
Chat assistant, Personalisation, Conversational AI, Computer-Assisted
Human Interaction
ACM Reference Format:
Jindi Wang, Ioannis Ivrissimtzis, Zhaoxing Li, and Lei Shi. 2024. Enhancing
User Experience in Chinese Initial Text Conversations with Personalised
AI-Powered Assistant. In Extended Abstracts of the CHI Conference on Human
Factors in Computing Systems (CHI EA ’24), May 11–16, 2024, Honolulu, HI,
USA. ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/3613905.
3651104
1 INTRODUCTION
Large language models (LLMs) like GPT-4 [
30
], have demonstrated
impressive capabilities in comprehending and generating text in
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
CHI EA ’24, May 11–16, 2024, Honolulu, HI, USA
©2024 Copyright held by the owner/author(s).
ACM ISBN 979-8-4007-0331-7/24/05.
https://doi.org/10.1145/3613905.3651104
ways that mirror closely human communication. These AI mod-
els can be conditioned to mimic specic writing styles, thereby
enabling possibilities of style-specic text generation for rened
digital communications [
20
]. Moreover, AI’s ability to discern sen-
timent from text [
36
] can be utilised to understand the emotional
undertones in online conversations and respond with empathy, a
critical aspect towards an enriched user experience.
AI-powered chat assistants have emerged as a transformative
force in the eld of computer-assisted human interaction [
39
]. This
revolution is not merely about convenience or eciency; it touches
upon the foundational aspects of human interaction, such as com-
fort and willingness to engage. That is, AI-powered conversational
interfaces hold the promise, not only of enhanced utility but also
to facilitate deeper human connection [32].
One critical stage in both digital and face-to-face conversations
is the initial interaction, often termed “icebreaking”. Despite being
a brief phase, it can be delicate and set the tone for the entire
communication, aecting the trajectory of subsequent engagement
[
33
], in both professional and informal environments. Thus, given
the persistent availability of AI chat assistants and their ability to
process vast amounts of information swiftly [
10
], integrating them
into these initial phases can have profound implications for user
experience. In particular, functionality such as text recommendation
and polishing can rene and guide these interactions, making them
more meaningful and user-centric [19].
While a great deal of research on Conversational AI interfaces
has focused on their applications in task-oriented scenarios, such
as customer support and virtual assistance [
5
], their potential to
enhance human-human conversations has been relatively under-
explored. This scarcity of relevant research is even more pronounced
on the more specic topic of AI-assisted initial human-human in-
teractions.
Thus, recognising the pivotal role of initial text conversation
in shaping subsequent interactions, and noting that no currently-
used chat application incorporates a personalised AI assistant, this
study focuses on this specic application domain. It aims to address
the primary research question: In initial textual interaction
scenario, can an AI chat assistant, equipped with text recom-
mendation and polishing functionalities, increase the user’s
comfort level and willingness to continue the conversation?
We believe that understanding the inuence of AI in initial user
interactions is crucial for designing user-centric, AI-driven chat
interfaces.

CHI EA ’24, May 11–16, 2024, Honolulu, HI, USA Wang, et al.
2 RELATED WORK
2.1 “Breaking the Ice” in Digital
Communication
The act of “breaking the ice” in digital communication has been
a topic of interest in the eld of Computer-Human Interaction
(CHI) and related disciplines. The transition from face-to-face to
digital communication has brought unique challenges and oppor-
tunities in establishing initial rapport. Hancock et al. [
15
] noted
that the absence of non-verbal cues in text-based communication
can both hinder and facilitate self-disclosure, depending on the
context. Tidwell and Walther [
35
] found that, in online settings,
users often employ strategic self-presentation techniques to “break
the ice” eectively. These strategies might involve sharing more
personal information earlier in the conversation than in face-to-
face interactions. Another variable is the role of the interface design
in easing initial interactions. Vlahovic et al. [
37
] highlighted that
platforms incorporating gamied elements or icebreaker questions
can signicantly reduce the initial tension and promote organic
conversations. As digital communication platforms evolve, under-
standing and enhancing the initial text conversation phase remains
pivotal for fostering meaningful online connections.
2.2 Text-based Communication Language Style
The language style in text-based communication has been thor-
oughly studied from various angles, including sociolinguistics, com-
putational linguistics, and computer science.
2.2.1 Sociolinguistic Perspectives.Text-based communication
merges speech and writing elements, creating a unique form that
combines the informality of spoken language with the permanence
of written text [
3
,
17
]. The format demands that users possess “com-
municative competence” [
21
], which is the ability to communicate
eectively within specic contexts. Core to its style are stylistic
variations, such as abbreviations, emojis, and non-standard punctu-
ation [
7
,
34
], which often convey aspects of the user’s identity like
age and dialect [
18
]. Here, however, our chat assistant maintains
a standardised format by not including these stylistic elements.
Emotional expressions, ranging from capitalisations for emphasis
to emojis for nuanced feelings, enrich this communication form
[
22
]. Herring et al.’s categorisation for Computer-Mediated Dis-
course [
16
]—comprising system, participation, structure, and tone
dimensions—aptly captures text communication’s essence. The tone,
varying from formal to playful, signicantly inuences the user
experience and communication eciency.
2.2.2 Computer Science Perspectives.Text-based communi-
cation on digital platforms, characterised by its distinct linguistic
attributes such as brevity and informality, poses both challenges
and opportunities for the elds of NLP, HCI, and dialogue systems.
Traditional NLP techniques often struggle with non-standard lan-
guage forms [
8
], prompting researchers to adapt techniques specic
to platforms, as exemplied by the need for tailored POS tags in
Twitter [13].
Out-of-vocabulary (OOV) words are prevalent in social media
text, and they pose signicant challenges [
14
]. Furthermore, the
evolving nature of online language necessitates periodic model
updates [
6
]. Emoticons and unconventional punctuation, while
complex, oer opportunities for nuanced language models to excel
[27].
Demographic factors, including age [
18
], gender [
1
], and location
[
9
], have been identied as inuential factors in online linguistic
styles. Recent advancements in the domain focus on style adapta-
tion in text generation, with methods ranging from unsupervised
techniques [
24
] to back-translation approaches [
29
]. The intersec-
tion of personalisation and style is emerging as a critical area of
study, highlighting the signicance of individual stylistic prefer-
ences in enhancing user-system interactions [25, 28].
2.3 Personalisation and Style Adjustment in
Chat Assistants
2.3.1 Personalised Response Generation.Recent research has
adopted diversied approaches to personalising dialogue systems.
Li et al. [
26
] used reinforcement learning, processing user feedback
for tailored responses. Zhang et al. [
42
] enhanced engagement by
integrating user personas into the dialogue, while another work by
Zhang et al. [
41
] utilised data from users’ social media proles for
personalisation. Ghandeharioun et al. [
12
] focused on rening user
experience by adapting to their language style and emotions. De-
spite these advances, challenges persist: the robustness of feedback-
based reinforcement learning, the broader application of persona
models, privacy concerns in social media-based insights, and the
full implications of real-time personalisation methods, all require
further exploration.
2.3.2 Style Adjustment in Dialogue Systems.Recent advance-
ments in dialogue systems emphasise the importance of stylistic
adjustments for enhanced user engagement. Wu et al. [
40
] proposed
the “prototype-then-edit” methodology, highlighting the potential
of deliberate stylistic modications. Lample et al. [
24
] introduced
an unsupervised style transfer model that maintains semantic con-
tent while altering style. Additionally, Keskar et al. [
23
] showcased
rened stylistic control in large language models. Together, these
contributions highlight the value of stylistic nuances in advancing
AI-driven communication.
3 KEY FEATURES OF PERSONALISED CHAT
ASSISTANT AND IMPLEMENTATION
Our user interface introduces an AI chat assistant designed to en-
hance user communication experiences. It has two primary func-
tionalities as shown in Fig. 1.
Text Recommendation: This feature is useful when users are
unsure about how to respond during conversations. Based on a
user’s preferences, the AI can suggest potential replies that align
with a predetermined AI extraversion level and linguistic style.
This grants users the exibility to tailor the assistant’s responses
in terms of 1) Extroversion Level: Three levels are available -
introverted, average, and extroverted inspired by [
38
]. This ensures
that the AI’s recommended responses resonate with the user’s de-
sired interaction style. 2) Linguistic Style: Users can set the AI
to generate text with a specic tone, choosing between “humor-
ous” and “respectful”, according to results in the preliminary user
survey of the appendix. This oers a personalised touch, making

Enhancing User Experience in Chinese Initial Text Conversations with Personalised AI-Powered Assistant CHI EA ’24, May 11–16, 2024, Honolulu, HI, USA
Figure 1: Interaction between two users of the AI chat assistant. The shown example is in English; however, all participants in
the study consented to converse in Chinese.
the conversation more engaging or formal depending on the user’s
preferences and the perceived context.
Text Polishing: Beyond merely suggesting replies, the assistant
can rene user-composed messages. Similar to the recommendation
feature, the text polishing functionality is also customised based
on the above two settings. This ensures clarity, coherence, and
alignment with the chosen extroversion levels and linguistic styles.
Our Personalised AI Chat Assistant is underpinned by OpenAI’s
GPT-3.5 API
1
. In text recommendation mode, the system analyzes
incoming chat content to generate contextually apt replies, accord-
ing to user-specied stylistic preferences. In the text polishing
mode, the system again analyses the incoming chat and processes
the user’s intended replies to produce rened text outputs. This
adaptability, paired with an interface inspired by WhatsApp’s web
design and backed by an SQLite database for ecient data stor-
age, led to a system suitable for the study of user-centric digital
communication.
4 USER STUDY
4.1 Participants
We conducted an experiment involving a group of 28 Chinese par-
ticipants. After a detailed explanation of the study’s goals, these
participants provided their informed consent by signing a form,
thereby agreeing to participate in the experiment. Detailed informa-
tion about each participant’s characteristics is provided in Table 1
in supplementary material. The brief of the participants was that
they would engage in conversations with people they had not met
before, with a primary focus on the subject of travel. This setting
was adopted to maintain consistency across all trials, in both the
experimental condition of engaging with an unfamiliar person, and
the topic of discussion.
1https://platform.openai.com/docs/api-reference
4.2 Study Setup
•
Participants Pairing: We employed a random pairing strat-
egy for our 28 participants, ensuring that participants in the
control and experimental sessions did not interact with the
same individuals, as shown in Table 1 in the supplementary
material.
•
Experimental Process: Each participant was involved in
two distinct sessions as illustrated in Fig. 2. Control Session:
First, participants engaged in a textual conversation with a
person they were unfamiliar with, without the aid of the AI
chat assistant. Upon conclusion, they rated their chat com-
fort level and their willingness to continue the conversation.
Experimental Session: In the experimental phase, partic-
ipants engaged in a distinct second text conversation with
a new acquaintance, other than the person they engaged
in the control session. In this session, the AI chat assistant
was used, in the form of either AI-generated recommenda-
tions or by using the text polishing functionality. The limited
range of supported language styles (two), and extraversion
levels (three), allowed us to maintain a controlled linguistic
environment, while oering the participants some mean-
ingful choices, thereby catering to individual interaction
preferences. Each session was timed to last approximately 20
minutes, ensuring uniformity in the duration of interactions
across all trials. This structured approach was crucial for
obtaining comparable data while allowing the AI’s impact
on the conversation dynamics to be observed eectively.
4.3 Metrics
System Usability Scale (SUS) serves as a robust instrument for
evaluating the usability of systems such as AI chat assistants. It
primarily assesses the ease and comfort level experienced by users
during system interaction. An elevated SUS score is indicative of

CHI EA ’24, May 11–16, 2024, Honolulu, HI, USA Wang, et al.
Figure 2: Two-phase experimental procedure: initial chat without AI followed by AI-assisted chat, with Comfort Level (CL),
Continuation Willingness (CW) assessments, and post-experiment User Survey.
user-friendliness, implying that the AI assistant is perceived as
more intuitive and manageable.
User Experience (UX) encapsulates the overall impression and
response of users towards engaging with the AI chat assistant. This
metric critically assesses the system’s eectiveness, eciency, and
satisfaction quotient. A high UX score denotes that the AI assistant
meets or exceeds user expectations in terms of functionality and
responsiveness.
Comfort Level (CL) is used as a measure of the user’s ease during
initial textual interactions with new online contacts. Rated on a
scale from 1 (“very uncomfortable”) to 5 (“very comfortable”), with a
neutral midpoint at 3, CL is instrumental in collecting user feedback
for rening chat interfaces and enhancing the experience of online
interactions [4].
Continuation Willingness (CW) quanties a user’s propensity
to extend textual interactions following their initial engagement
with new online contacts. Scored from 1 (“low willingness”) to 5
(“high eagerness”), this metric is crucial for understanding user en-
gagement and the probability of a subsequent sustained interaction
in that digital environment, aiming at fostering a deeper connection
[11].
5 RESULTS
5.1 Usability and User Experience
To understand users’ perceptions of their interactions with our
personalised AI chat assistant, we used the System Usability Scale
(SUS). This measure oered insights into the system’s perceived
usability. Participants reported an average SUS score of 72.86 with
a standard deviation of 8.04, which represents “good” usability [
2
].
Although the majority of scores clustered between 65 and 75, our
Shapiro-Wilk test result (p = 0.0014) revealed that they did not
follow a normal distribution. However, a median score of 70.00, an
interquartile range [67.50, 75.63], and a bootstrap estimated 95%
condence interval for the median score [67.50, 73.75], indicate that
most users found the AI chat assistant’s usability to be satisfactory.
We employed the questionnaire in Schrepp et al. [
31
], to assess
user experience across the six pivotal dimensions they proposed,
by benchmarking our results against the standards they established.
As illustrated in Fig. 3, the AI chat assistant excelled in the Nov-
elty dimension, achieving a “Good” rating. This underscores users’
perception of the AI chat assistant as a refreshing and innovative
interaction tool. Furthermore, both Perspicuity and Stimulation
received an “Above average” rating, highlighting the system’s clear
and intuitive design and its ability to invigorate user engagement.
However, the metrics for the remaining three dimensions fell
within the “Below average” bracket, indicating areas with room for
enhancement, including rening the aesthetics, streamlining the
dynamic adjustments for increased eciency, and bolstering the
language model’s robustness.
Additionally, the internal consistency of our evaluation, gauged
via Cronbach’s Alpha, revealed robust reliability for several dimen-
sions. Specically, Attractiveness (
𝛼
= 0.77), Dependability (
𝛼
= 0.76), and Stimulation (
𝛼
= 0.80) all surpassed the empirically
established threshold of 0.7. Other dimensions, namely Perspicuity
(
𝛼
= 0.65), Eciency (
𝛼
= 0.63), and Novelty (
𝛼
= 0.65), hovered
close to this benchmark, further supporting the credibility of our
user experience ndings.
5.2 Comfort Level and Continuation
Willingness
Fig. 4 depicts the distribution of user ratings for comfort level
and willingness to continue the conversation across the two ex-
perimental conditions. Descriptive statistics reveal a noticeable
improvement in user responses when the AI chat assistant is active.
Specically, for comfort level (CL), we observed a mean score of
𝑀=
2
.
86 with a standard deviation of
𝑆𝐷 =
0
.
69 in the absence
of AI, which increased to
𝑀=
3
.
61
, 𝑆𝐷 =
0
.
94 with AI assistance,
corresponding to a 26% improvement. For the continuation willing-
ness metric (CW ), the respective values were
𝑀=
2
.
61
, 𝑆𝐷 =
0
.
72
without and
𝑀=
3
.
79
, 𝑆𝐷 =
0
.
82 with AI, corresponding to an even
bigger improvement of a 45.21%.
Normality tests with the Shapiro-Wilk method showed non-
normal distributions for both CL (
𝑝<
0
.
001) and CW (
𝑝<
0
.
001).
Given this, eect sizes were determined using Cli’s delta. The com-
puted values indicated a medium eect for comfort level (
𝛿CL-AI =
0
.
43), suggesting a moderate improvement in comfort. In contrast,

Enhancing User Experience in Chinese Initial Text Conversations with Personalised AI-Powered Assistant CHI EA ’24, May 11–16, 2024, Honolulu, HI, USA
Attractiveness Perspicuity Efficiency Dependability Stimulation Novelty
0.0
0.5
1.0
1.5
2.0
2.5
Excellent
Good
Above Average
Below Average
Bad
Mean
Figure 3: Distribution of scores across six key dimensions of user experience.
 










Figure 4: Boxplot outcomes of user Comfort Level (CL) and
Continue Willingness (CW) in the presence vs. absence of AI
intervention.
CW exhibited a large eect size (
𝛿CW-AI =
0
.
68), indicating a sig-
nicant increase in users’ willingness to continue the conversation
with AI assistance. The empirical ndings underscore the positive
eect of AI assistance in conversational interfaces. The enhanced
comfort level suggests that users nd the AI-enabled environment
more conducive to interaction. Moreover, the pronounced increase
in continuation willingness underscores greater user engagement
and a generally positive user attitude towards conversational AI
systems.
In summary, the integration of AI in chat systems evidently
contributes to a more engaging and comfortable user experience.
This nding holds a substantial promise for the design of future user-
centric conversational agents and advocates for a human-centred
approach in AI development.
5.3 Discussion
5.3.1 Usability and User Experience Reflection.Our nd-
ings reveal that the AI chat assistant signicantly enhances text-
based interactions in initial textual conversation scenarios. The
system’s high usability score signies that users nd it intuitive and
user-friendly, a critical aspect in encouraging the adoption of new
technology. The ease of use, minimal learning requirements, and
straightforward interaction contribute to its favourable reception.
In the realms of Novelty and Perspicuity, the AI chat assistant
excelled, showcasing innovative characteristics and a clear and in-
tuitive interface design. These attributes are essential in ensuring
user engagement and satisfaction. However, the lower scores in
other user experience dimensions, such as Attractiveness and E-
ciency, point towards necessary enhancements. Addressing these
areas could involve enhancing the visual appeal of the interface,
expanding the range of dynamic adjustments for accommodating
user preferences, and improving the sophistication of the language
model. These improvements could signicantly bolster the overall
user experience, making the AI chat assistant not only a functional
tool but also an aesthetically pleasing and ecient one.
5.3.2 Comfort Levels and Continuation Willingness.The im-
provement in users’ comfort levels and willingness to continue
interactions when assisted by the AI is particularly noteworthy.
This aspect of our ndings highlights the AI’s role in creating a
supportive environment for conversation, particularly in the ini-
tial, often awkward stages of unstructured interactions. The AI’s
ability to facilitate a more comfortable and engaging conversation
environment is pivotal in contexts where establishing rapport and
sustaining engagement are crucial.
5.3.3 Limitations and Future Directions.While our study of-
fers some valuable insights, it is important to acknowledge its limi-
tations. The focus on a specic ethnic demographic limits the gener-
alisability of our ndings, emphasising the need for future research
to encompass a more diverse participant pool. Such an extension of
our research could provide a more comprehensive understanding
of the AI chat assistant’s applicability across various cultural con-
texts. Variations in AI behaviour, even within the same extraversion
level, could inuence user perceptions. These variations, though
minor, underscore the importance of consistency in AI interac-
tions. Investigating these nuances can provide deeper insights into
user-AI interaction dynamics. The potential over-reliance on the
text polishing feature of the AI raises questions about long-term
user dependency. Future studies should explore the implications of

CHI EA ’24, May 11–16, 2024, Honolulu, HI, USA Wang, et al.
such reliance, particularly in terms of users’ communication skills
development and their interaction with AI technology over time.
Moreover, dierent users may interpret the same AI responses in
various ways, indicating a need for more standardised measures or
a broader range of qualitative data to capture the full spectrum of
user experiences. In addition, the use of xed AI personality proles
in our study presents a limitation. Future research should explore
adaptable AI personalities that can dynamically adjust to match the
user’s own personality, potentially enhancing the user experience
and engagement further. This approach could address the needs of
users with uid personality traits, oering a more personalized and
responsive interaction. Lastly, while our study sheds light on the
potential of AI in enhancing initial text-based interactions, it also
opens several avenues for future research on other types of human-
human textual interactions. Exploring these will not only address
the identied limitations but also expand our understanding of the
complexities and potential of AI-assisted communication.
6 CONCLUSION
In the digital era, the generation of AI-powered, high-quality, text-
based initial interactions is a very relevant yet understudied prob-
lem domain. Our results highlight the tangible benets of the use
of an AI chat assistant in elevating user experience. Users found
the system highly usable and appreciated the heightened comfort
it oered. Most notably, achieving consistent AI responses and ac-
commodating diverse user personalities stand as promising areas
for future renement.
In summary, our research underscores the promising potential
of AI in enhancing initial text-based exchanges. As technology
advances, it is crucial to anchor developments in user-centricity.
With sustained innovation, AI chat assistants have the potential
to reshape digital communication, ensuring enriched and smooth
human interactions.
REFERENCES
[1]
David Bamman, Jacob Eisenstein, and Tyler Schnoebelen. 2014. Gender identity
and lexical variation in social media. Journal of Sociolinguistics 18, 2 (2014),
135–160.
[2]
Aaron Bangor, Philip Kortum, and James Miller. 2009. Determining what indi-
vidual SUS scores mean: Adding an adjective rating scale. Journal of usability
studies 4, 3 (2009), 114–123.
[3]
Naomi S Baron. 2004. See you online: Gender issues in college student use
of instant messaging. Journal of language and social psychology 23, 4 (2004),
397–423.
[4]
Susan Bergin and Ronan Reilly. 2005. The inuence of motivation and comfort-
level on learning to program. In Proceedings of the 17th Workshop of the Psychology
of Programming Interest Group. Psychology of Programming Interest Group,
Brighton, UK, 293–304. https://mural.maynoothuniversity.ie/8685/
[5]
Hongshen Chen, Xiaorui Liu, Dawei Yin, and Jiliang Tang. 2017. A survey on
dialogue systems: Recent advances and new frontiers. Acm Sigkdd Explorations
Newsletter 19, 2 (2017), 25–35.
[6]
Cristian Danescu-Niculescu-Mizil, Robert West, Dan Jurafsky, Jure Leskovec,
and Christopher Potts. 2013. No country for old members: user lifecycle and
linguistic change in online communities. In Proceedings of the 22nd International
Conference on World Wide Web (Rio de Janeiro, Brazil) (WWW ’13). Association
for Computing Machinery, New York, NY, USA, 307–318. https://doi.org/10.
1145/2488388.2488416
[7]
Marcel Danesi. 2017. The semiotics of emoji: The rise of visual language in the age
of the internet. Bloomsbury Publishing, London, UK.
[8]
Jacob Eisenstein. 2013. What to do about bad language on the internet. In Pro-
ceedings of the 2013 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, Lucy Vanderwende,
Hal Daumé III, and Katrin Kirchho (Eds.). Association for Computational Lin-
guistics, Atlanta, Georgia, 359–369. https://aclanthology.org/N13-1037
[9]
Jacob Eisenstein, Brendan O’Connor, Noah A. Smith, and Eric P. Xing. 2010. A
latent variable model for geographic lexical variation. In Proceedings of the 2010
Conference on Empirical Methods in Natural Language Processing (Cambridge,
Massachusetts) (EMNLP ’10). Association for Computational Linguistics, USA,
1277–1287.
[10]
Tira Nur Fitria. 2023. Articial intelligence (AI) technology in OpenAI ChatGPT
application: A review of ChatGPT in writing English essay. ELT Forum: Journal of
English Language Teaching 12, 1 (March 2023), 44–58. https://doi.org/10.15294/
elt.v12i1.64069 Number: 1.
[11]
Mark Freiermuth and Douglas Jarrell. 2006. Willingness to communicate: can
online chat help? 1. International journal of applied linguistics 16, 2 (2006), 189–
212.
[12]
Asma Ghandeharioun, Judy Hanwen Shen, Natasha Jaques, Craig Ferguson, Noah
Jones, Agata Lapedriza, and Rosalind Picard. 2019. Approximating interactive
human evaluation with self-play for open-domain dialog systems. Advances in
Neural Information Processing Systems 32 (2019), 13658–13669.
[13]
Kevin Gimpel, Nathan Schneider, Brendan O’Connor, Dipanjan Das, Daniel
Mills, Jacob Eisenstein, Michael Heilman, Dani Yogatama, Jerey Flanigan, and
Noah A. Smith. 2011. Part-of-Speech Tagging for Twitter: Annotation, Features,
and Experiments. In Proceedings of the 49th Annual Meeting of the Association
for Computational Linguistics: Human Language Technologies, Dekang Lin, Yuji
Matsumoto, and Rada Mihalcea (Eds.). Association for Computational Linguistics,
Portland, Oregon, USA, 42–47. https://aclanthology.org/P11-2008
[14]
Bo Han and Timothy Baldwin. 2011. Lexical normalisation of short text messages:
Makn sens a# twitter. In Proceedings of the 49th annual meeting of the associa-
tion for computational linguistics: Human language technologies. Association for
Computational Linguistics, Portland, OR, USA, 368–378.
[15]
Jerey T. Hancock, Catalina Toma, and Nicole Ellison. 2007. The truth
about lying in online dating proles. In Proceedings of the SIGCHI Conference
on Human Factors in Computing Systems (<conf-loc>, <city>San Jose</city>,
<state>California</state>, <country>USA</country>, </conf-loc>) (CHI ’07).
Association for Computing Machinery, New York, NY, USA, 449–452. https:
//doi.org/10.1145/1240624.1240697
[16]
Susan C Herring. 2007. A faceted classication scheme for computer-mediated
discourse. Language @ internet 4, 1 (2007), 37 pages.
[17]
Susan C. Herring. 2008. Language and the Internet. In The International Ency-
clopedia of Communication. John Wiley & Sons, Ltd, Hoboken, New Jersey, USA.
https://doi.org/10.1002/9781405186407.wbiecl005
[18]
Dirk Hovy and Anders Søgaard. 2015. Tagging Performance Correlates with
Author Age. In Proceedings of the 53rd Annual Meeting of the Association for
Computational Linguistics and the 7th International Joint Conference on Natural
Language Processing (Volume 2: Short Papers), Chengqing Zong and Michael
Strube (Eds.). Association for Computational Linguistics, Beijing, China, 483–488.
https://doi.org/10.3115/v1/P15-2079
[19]
Zhiting Hu, Haoran Shi, Bowen Tan, Wentao Wang, Zichao Yang, Tiancheng
Zhao, Junxian He, Lianhui Qin, Di Wang, Xuezhe Ma, Zhengzhong Liu, Xi-
aodan Liang, Wangrong Zhu, Devendra Singh Sachan, and Eric P. Xing. 2019.
Texar: A Modularized, Versatile, and Extensible Toolkit for Text Generation.
arXiv:1809.00794 [cs.CL]
[20]
Zhiting Hu, Zichao Yang, Xiaodan Liang, Ruslan Salakhutdinov, and Eric P. Xing.
2017. Toward Controlled Generation of Text. In Proceedings of the 34th Interna-
tional Conference on Machine Learning (Proceedings of Machine Learning Research,
Vol. 70), Doina Precup and Yee Whye Teh (Eds.). PMLR, Sydney, Australia, 1587–
1596. https://proceedings.mlr.press/v70/hu17e.html
[21]
Dell Hymes. 2001. Foundations in sociolinguistics: An ethnographic approach.
Routledge, London, UK.
[22]
Linda K Kaye, Helen J Wall, and Stephanie A Malone. 2016. “Turn that frown
upside-down”: A contextual account of emoticon usage on dierent virtual plat-
forms. Computers in Human Behavior 60 (2016), 463–467.
[23]
Nitish Shirish Keskar, Bryan McCann, Lav R. Varshney, Caiming Xiong, and
Richard Socher. 2019. CTRL: A Conditional Transformer Language Model for Con-
trollable Generation. https://doi.org/10.48550/arXiv.1909.05858 arXiv:1909.05858
[cs].
[24]
Guillaume Lample, Sandeep Subramanian, Eric Smith, Ludovic Denoyer,
Marc’Aurelio Ranzato, and Y-Lan Boureau. 2019. Multiple-Attribute Text Rewrit-
ing. In International Conference on Learning Representations. openreview.net, New
Orleans, LA, USA, 20 pages.
[25]
Jiwei Li, Michel Galley, Chris Brockett, Georgios P. Spithourakis, Jianfeng Gao,
and Bill Dolan. 2016. A Persona-Based Neural Conversation Model. https:
//doi.org/10.48550/arXiv.1603.06155 arXiv:1603.06155 [cs].
[26]
Jiwei Li, Will Monroe, Alan Ritter, Michel Galley, Jianfeng Gao, and Dan Jurafsky.
2016. Deep Reinforcement Learning for Dialogue Generation. https://doi.org/
10.48550/arXiv.1606.01541 arXiv:1606.01541 [cs].
[27]
Hannah Miller, Jacob Thebault-Spieker, Shuo Chang, Isaac Johnson, Loren Ter-
veen, and Brent Hecht. 2016. “Blissfully Happy” or “Ready toFight”: Varying
Interpretations of Emoji. Proceedings of the International AAAI Conference on Web
and Social Media 10, 1 (2016), 259–268. https://doi.org/10.1609/icwsm.v10i1.14757
Number: 1.

Enhancing User Experience in Chinese Initial Text Conversations with Personalised AI-Powered Assistant CHI EA ’24, May 11–16, 2024, Honolulu, HI, USA
[28]
Shereen Oraby, Lena Reed, Shubhangi Tandon, T. S. Sharath, Stephanie Lukin,
and Marilyn Walker. 2018. Controlling Personality-Based Stylistic Variation with
Neural Natural Language Generators. https://doi.org/10.48550/arXiv.1805.08352
arXiv:1805.08352 [cs].
[29]
Shrimai Prabhumoye, Yulia Tsvetkov, Ruslan Salakhutdinov, and Alan W. Black.
2018. Style Transfer Through Back-Translation. https://doi.org/10.48550/arXiv.
1804.09000 arXiv:1804.09000 [cs].
[30]
Colin Rael, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang,
Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2020. Exploring the limits of
transfer learning with a unied text-to-text transformer. The Journal of Machine
Learning Research 21, 1 (2020), 5485–5551.
[31]
Martin Schrepp, Andreas Hinderks, and Jörg Thomaschewski. 2017. Construction
of a Benchmark for the User Experience Questionnaire (UEQ). International
Journal of Interactive Multimedia and Articial Intelligence 4 (June 2017), 40–44.
https://doi.org/10.9781/ijimai.2017.445
[32]
Iulian V. Serban, Chinnadhurai Sankar, Mathieu Germain, Saizheng Zhang,
Zhouhan Lin, Sandeep Subramanian, Taesup Kim, Michael Pieper, Sarath Chan-
dar, Nan Rosemary Ke, Sai Rajeshwar, Alexandre de Brebisson, Jose M. R. Sotelo,
Dendi Suhubdy, Vincent Michalski, Alexandre Nguyen, Joelle Pineau, and Yoshua
Bengio. 2017. A Deep Reinforcement Learning Chatbot. arXiv:1709.02349 [cs.CL]
[33]
Bayan Abu Shawar and Eric Atwell. 2007. Chatbots: are they really useful? Journal
for Language Technology and Computational Linguistics 22, 1 (2007), 29–49.
[34]
Crispin Thurlow. 2006. From statistical panic to moral panic: The metadiscursive
construction and popular exaggeration of new media language in the print media.
Journal of computer-mediated communication 11, 3 (2006), 667–701.
[35]
Lisa Collins Tidwell and Joseph B Walther. 2002. Computer-mediated communi-
cation eects on disclosure, impressions, and interpersonal evaluations: Getting
to know one another a bit at a time. Human communication research 28, 3 (2002),
317–348.
[36]
G Vinodhini and RM Chandrasekaran. 2012. Sentiment analysis and opinion
mining: a survey. International Journal 2, 6 (2012), 282–292.
[37]
Tatiana A Vlahovic, Sam Roberts, and Robin Dunbar. 2012. Eects of duration
and laughter on subjective happiness within dierent modes of communication.
Journal of Computer-Mediated Communication 17, 4 (2012), 436–450.
[38]
Sarah Theres Völkel, Ramona Schoedel, Lale Kaya, and Sven Mayer. 2022. User
Perceptions of Extraversion in Chatbots after Repeated Use. In Proceedings
of the 2022 CHI Conference on Human Factors in Computing Systems (<conf-
loc>, <city>New Orleans</city>, <state>LA</state>, <country>USA</country>,
</conf-loc>) (CHI ’22). Association for Computing Machinery, New York, NY,
USA, Article 253, 18 pages. https://doi.org/10.1145/3491102.3502058
[39]
J Wang, I Ivrissimtzis, Z Li, and L Shi. 2024. Impact of Personalised AI Chat
Assistant on Mediated Human-Human Textual Conversations: Exploring Female-
Male Dierences. In 29th Annual ACM Conference on Intelligent User Interfaces.
Newcastle University, ACM Press, Greenville, SC, USA, 5 pages.
[40]
Yu Wu, Furu Wei, Shaohan Huang, Yunli Wang, Zhoujun Li, and Ming Zhou.
2019. Response Generation by Context-Aware Prototype Editing. Proceedings
of the AAAI Conference on Articial Intelligence 33, 01 (July 2019), 7281–7288.
https://doi.org/10.1609/aaai.v33i01.33017281
[41]
Saizheng Zhang, Emily Dinan, Jack Urbanek, Arthur Szlam, Douwe Kiela, and
Jason Weston. 2018. Personalizing Dialogue Agents: I have a dog, do you have
pets too? https://doi.org/10.48550/arXiv.1801.07243 arXiv:1801.07243 [cs].
[42]
Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang
Gao, Jianfeng Gao, Jingjing Liu, and Bill Dolan. 2020. DialoGPT: Large-Scale
Generative Pre-training for Conversational Response Generation. https://doi.
org/10.48550/arXiv.1911.00536 arXiv:1911.00536 [cs].
A SUPPLEMENTARY MATERIAL
B PRELIMINARY USER SURVEY
Prior to the commencement of an in-depth exploration into the
domain of personalised AI chat assistants in initial text-based inter-
actions, participants were duly informed of the study’s aims and
methodologies. Consent was obtained in alignment with the ethical
guidelines of our institution. A preliminary survey was undertaken
to ascertain the genuine interest and necessity for personalisation
among potential users.
An online questionnaire
2
was formulated and disseminated among
a heterogeneous group of individuals, encompassing an age spec-
trum of 15 to 44 years, along with other diverse demographic at-
tributes. The survey witnessed participation from 162 individuals.
2https://qfreeaccountssjc1.az1.qualtrics.com/jfe/form/SV_0VQb8OE6ZM5R5u6
Table 1: Demographics and control group and experimental
group pairs
Control Group Without AI (Age, Male/Female) Experimental Group With AI
P1 (27, Female) & P2 (28, Male) P1 & P16
P3 (28, Male) & P4 (28, Male) P3 & P18
P5 (33, Female) & P6 (25, Female) P5 & P20
P7 (27, Male) & P8 (28, Male) P7 & P22
P9 (26, Male) & P10 (27, Male) P9 & P24
P11 (28, Male) & P12 (27, Female) P11 & P26
P13 (28, Female) & P14 (21, Female) P13 & P28
P15 (27, Male) & P16 (29, Male) P15 & P2
P17 (27, Male) & P18 (29, Female) P17 & P4
P19 (27, Male) & P20 (27, Male) P19 & P6
P21 (35, Female) & P22 (20, Male) P21 & P8
P23 (35, Male) & P24 (33, Male) P23 & P10
P25 (30, Male) & P26 (31, Female) P25 & P12
P27 (26, Male) & P28 (28, Male) P27 & P14
The questionnaire was initiated with queries aimed at discerning
the general attitude towards chatbots and AI assistants, focusing
on current levels of satisfaction and pinpointing areas necessitating
enhancements.
A distinct section of the questionnaire presented participants
with a range of language style options, such as ‘formal’, ‘humorous’,
‘playful’, ‘serious’, ‘respectful’, and ‘oensive’. Participants were
solicited to rate these styles, contemplating their ideal AI chat
assistant, on a scale ranging from 1 (indicating “not needed at all”) to
5 (signifying “highly needed”). Moreover, a rating bar was provided
for participants to express their perceived need for a personalised
AI chat assistant.
In the analysis of the results, scores exceeding 3 were considered
indicative of a positive inclination. It was observed that a substantial
majority of the respondents (82.67%) articulated a requirement for
a more personalised AI chat experience. Regarding language style
preferences, ‘Respectful’ and ‘Humorous’ were the predominant
choices, preferred by 67.72% and 74.16% of respondents, respectively.
Additionally, ‘Playful’ and ‘Serious’ styles were favoured by 66.93%
and 53.54% of respondents, respectively. Conversely, a mere 33.86%
of respondents exhibited a preference for an ‘Oensive’ tone.
This conrmation of user demand for personalisation, coupled
with the identication of ‘Respectful’ and ‘Humorous’ as the most
favoured language styles, facilitated the renement of our subse-
quent experimental endeavours. This strategic realignment ensured
that our study was deeply anchored in the actual preferences of
users, thereby signicantly enhancing its pertinence and potential
for practical application.