Available via license: CC BY 4.0
Content may be subject to copyright.
"I think this is the most disruptive technology"
Exploring Sentiments of ChatGPT Early Adopters
using Twitter Data
Mubin Ul Haque, Isuru Dharmadasa, Zarrin Tasnim Sworna, Roshan Namal Rajapakse, Hussain Ahmad*
{mubinul.haque, isuru.mahaganiarachchige, zarrintasnim.sworna, roshan.rajapakse, hussain.ahmad}@adelaide.edu.au
School of Computer Science, University of Adelaide, Australia
Abstract—Large language models have recently attracted sig-
nificant attention due to their impressive performance on a
variety of tasks. ChatGPT developed by OpenAI is one such
implementation of a large, pre-trained language model that
has gained immense popularity among early adopters, where
certain users go to the extent of characterizing it as a disruptive
technology in many domains. Understanding such early adopters’
sentiments is important because it can provide insights into the
potential success or failure of the technology, as well as its
strengths and weaknesses. In this paper, we conduct a mixed-
method study using 10,732 tweets from early ChatGPT users.
We first use topic modelling to identify the main topics and then
perform an in-depth qualitative sentiment analysis of each topic.
Our results show that the majority of the early adopters have
expressed overwhelmingly positive sentiments related to topics
such as Disruptions to software development,Entertainment and
exercising creativity. Only a limited percentage of users expressed
concerns about issues such as the potential for misuse of Chat-
GPT, especially regarding topics such as Impact on educational
aspects. We discuss these findings by providing specific examples
for each topic and then detail implications related to addressing
these concerns for both researchers and users.
Index Terms—ChatGPT, Generative Pretrained Transformer,
Early adopters, Twitter, Sentiment Analysis, Topic Modeling
I. INTRODUCTION
ChatGPT is an artificial intelligence (AI) chatbot that under-
stands and generates natural human language with remarkable
sophistication, sensitivity, and usability [1]. ChatGPT is an
application of the latest version of GPT-3 (Generative Pretrained
Transformer 3), a state-of-the-art language processing AI model
developed by the OpenAI
1
foundation, that enables it to
generate human-like text. Unlike traditional chatbots, ChatGPT
remembers what the user said earlier in the conversation for
follow-up questions, rejects inappropriate requests, and chal-
lenges incorrect responses [2]. Moreover, ChatGPT provides
answers, solutions, and descriptions to complex questions,
including potential ways to solve layout problems, write code,
and answer optimization queries [3]. Given the advantages of
ChatGPT over traditional chatbots, ChatGPT has attracted more
than 1 million users in just one week after it was launched,
leaving behind other popular online platforms such as Netflix,
All the authors contributed equally to this paper.
*Corresponding author: hussain.ahmad@adelaide.edu.au
1https://openai.com/
Facebook, and Instagram in terms of adoption rates [4]. In
addition, there is a rising number of commentators who predict
that ChatGPT will replace Google in the near future [5]. Some
early adopters of ChatGPT believe that it will eventually
obsolete several professions related to content creation, such as
programmers, professors, playwrights, and journalists [1]. For
example, it has been demonstrated that ChatGPT is capable
of producing high-quality responses to a variety of challenges,
including solving coding challenges and generating accurate
responses to exam queries [6].
As described above, ChatGPT offers several benefits to
a wide range of users. However, being a new technology,
identifying early adopter sentiments is of high importance due
to several reasons. Firstly, early adopters are usually the most
enthusiastic and influential users of a product, and their opinions
and sentiments can help to shape the broader perception of
new technology. This information can provide critical insights
into the potential success or failure of the product.
Secondly, early adopters are often the first to encounter any
issues or problems with new technology, and their feedback
can help to identify and fix these issues before they become
widespread. Therefore, exploring the early adopter sentiments,
particularly for a disruptive technology such as ChatGPT, would
increase the chances of success of the tool in the market.
To investigate the sentiments of ChatGPT early adopters,
we analyze data from Twitter, which allows users to read and
share 140-character messages called “tweets". Unlike other
social networking platforms (e.g., Facebook and Instagram),
unregistered users can access and read tweets [7]. Moreover,
Twitter is a famous and large social networking micro-
blogging site [8]. With the burgeoning popularity of Twitter,
researchers and practitioners are increasingly using Twitter data
to get untapped information from potential customers [9]. For
example, Twitter has been explored for investigating the public
perception of "Internet of Things" [9], COVID-19 symptoms
emergence [10], drug-related adverse events detection [11],
work emotion and stress analysis [12], mining public health
data [13], and influenza epidemics detection [14]. Therefore,
in this study, we decided to use Twitter as a data source to
analyze the sentiments of early adopters of ChatGPT.
In this paper, we conduct a mixed-method study on 10,732
tweets from early adopters. We first use topic modelling to
1
arXiv:2212.05856v1 [cs.CL] 12 Dec 2022
LDA
Topic Modeling
Filter non-
English tweets
Keyword-
based search
User Location
User occupation
User Verification
Tweets
Characteristics analysis
of ChatGPT early adopters
Duplication
and Noise
Removal
Stop-word
Removal
Lowercasing Lemmatization
* ( ! ;
adding
added
adds
add
Data pre-processing
Sentiment Analysis
of ChatGPT early adopters
ChatGPTTweet Dataset Construction
RQ1
RQ2
RQ3
Fig. 1: Overview of our research methodology
identify the main topics of discussion. Next, we perform manual
sentiment analysis to qualitatively analyze a selected set of
tweets from each topic.
In summary, our paper makes the following contributions.
•
It provides an overarching analysis of topics discussed
by early adopters of ChatGPT through their tweets. We
describe each discussed topic in detail with specific tweet
examples. Moreover, we show the total number of tweets
against each topic to demonstrate its significance compared
to other topics.
•
It presents a high-level investigation of the sentiments
of early ChatGPT adopters for each identified topic.
The users’ sentiments are categorized based on positive,
negative, and neutral. This analysis shows the users’
perception of each topic.
•
We describe potential future research directions and
areas where ethical and societal implications are present
regarding ChatGPT use.
The rest of the paper is organized as follows. Section II
presents the related work of this study. While Section III
reports the research methodology, Section IV describes the
results. Section V presents the implications of our study for
researchers and users. Section VI describes the threats to the
validity of this study. Lastly, the conclusion of this study is
presented in Section VII.
II. RE LATE D WOR K
As ChatGPT technology is quite new, we have not found
substantial research directly related to ChatGPT. However, there
exist literature on the GPT family of text-generating AIs (e.g.,
GPT-2 and GPT-3). Also, we identified sufficient literature on
Twitter data mining to investigate users’ sentiments. Therefore,
in the following, we report some closely related studies to our
study.
Bian et al. [9] analyzed Twitter data to understand the public
sentiments about Internet of Things (IoT). They performed
sentiment analysis and topic modelling to identify prominent
topics and public attitudes toward each topic. As a result, they
found users are more interested in business and technology
compared to other domains of IoT, and have favourable
sentiments toward IoT. Similarly, Trivedi et al. [15] introduced
aRobustly Optimized BERT Pre-training Approach (RoBERTa)
to investigate public sentiments through their tweets on hybrid
work arrangements. The RoBERTa revealed that the majority
of users have positive sentiments for the hybrid work model.
Another study [16] reported Twitter temporal data mining for
predicting a movie’s success. The authors proposed a rating
prediction model and a temporal product popularity model to
forecast users’ satisfaction and movie popularity among users,
respectively. Similarly, Nawaz et al. [17] proposed a prediction
model to predict the political election results of Pakistan. The
authors showed 98% accuracy and efficiency of their model
in predicting the results through Twitter data as compared to
alternative approaches.
To the best of our knowledge, our work is the first study that
conducted a qualitative analysis of early adopter sentiments
and feedback on ChatGPT. We contribute to the literature by
providing a snapshot of the early public responses to this latest
technology.
III. RESEARCH METHODOLOGY
This section presents our research methodology, where
we discuss our research questions, dataset construction, pre-
processing steps, identification of discussed ChatGPT topics,
and sentiment analysis on these topics. Fig. 1 presents the
overview of our research methodology.
A. Research Questions
The following RQs motivated our empirical study.
•RQ1
. What are the characteristics of ChatGPT early
adopters?
•RQ2
. ChatGPT Topics - What are the main topics that
are being discussed about ChatGPT on Twitter?
•RQ3
. ChatGPT Sentiments - What are the sentiments that
are being expressed about ChatGPT topics on Twitter?
2
Fig. 2: Qualitative analysis of tweets: Initial version of the code sheet for T1
B. ChatGPT Tweet Dataset Construction
To assess public sentiments on the early adoption of Chat-
GPT, we collected social media data, specifically from Twitter.
We collected tweets from December 5, 2022 to December 7,
2022. While collecting the tweets, we only considered the
tweets that included the keyword "ChatGPT". Besides, we
only collected the tweets that were written in the English
language. We used Python and Twitter API to extract Twitter
data. In our
ChatGP T T w eet
dataset, we have 18K tweets,
where we collected the text, user location, user occupation,
user verification, posted date, and hashtags for each tweet. To
answer RQ1, we analyzed user location, user occupation, and
user verification status information.
C. Data Pre-processing
We pre-processed our
ChatGP T T w eet
dataset using the
following steps:
1) Duplication Removal: We removed retweets and duplicate
tweets where a user repeated a tweet of another user. After
duplicate tweet removal, our ChatGPTTweet dataset includes
10,732 tweets that we used in our analysis.
2) Lowercasing: We lowercased the tweets that represent
words in different cases (e.g., StackOverflow and Stackover-
flow) to the same lower-case form (e.g., StackOverflow).
3) Noise Removal: We removed noises (e.g., punctuation
marks) to retain only the alphanumerical data for cleaning our
ChatGP T T w eet
dataset. In the noise removal step, we also
removed URLs, Emojis, and Twitter handles.
4) Stop words Removal: We removed stop-words that appear
frequently (e.g., this, are, and a) but do not help to distinguish
one tweet from another. We removed stop-words using the
NLTK [18] English stop-word list. Besides, we removed the
top three frequently appearing key domain specific words, such
as ChatGPT, OpenAI, and AI.
5) Lemmatization: We performed WordNet-based lemmati-
zation using NLTK [18]. We used lemmatization to represent
a word’s inflected forms (e.g., getting, gets) to its dictionary-
based root form (e.g., get), which is a common practice in the
existing literature [19].
D. Identification of ChatGPT Topics
To answer RQ2, we identified a set of ChatGPT key
topics using the Latent Dirichlet Allocation (LDA) modelling
technique [20]. LDA is a commonly used technique for
topic modelling in the existing Software Engineering (SE)
literature [21], [22]. LDA is used to group tweets of our
ChatGP T T w eet
dataset into a set of topics using word co-
occurrence and frequency. A set of probabilities are assigned
to each tweet by LDA. Here, the probabilities refer to the
chances of a tweet being related to a specific topic. We used
MALLET [23] implementation of LDA, which is commonly
used in existing literature [24], [25].
Identification of the optimal number of topics
N
, while
implementing LDA is critical as LDA may generate a huge
number of narrow topics for a high value of
N
. In contrast,
LDA may create broad generalized topics for a low value
of
N
. Hence, we executed a broad range of experiments to
analyze the coherence scores, which were achieved by varying
N
values from 5 to 60, with steps=2 and iteration=100, 500,
and 1000 using MALLET. The coherence score depicts the
understandability of LDA topics that is relevant to human
comprehensibility [26]. We obtained a comparatively high
coherence score by running MALLET when the number of
topics ranged from 9 to 12.
To ensure that we select the optimal number of topics, we
examined randomly selected 15 tweets from each topic for 9
≤N≤12. Based on our examination, we found that N= 9
(i.e, the optimal number of topics is 9) serves the purpose of
balancing the comprehensibility of our dataset. This approach
is followed in similar studies to find the optimal number of
3
Fig. 3: Geographic distribution of the early adopters of ChatGPT based on tweets
topics [22]. We then ran LDA using MALLET with N=9 and
generated a CSV file for each topic. Each CSV file was sorted
based on the highly relevant documents of the topic from LDA
topic modeling method.
E. Sentiment Analysis on ChatGPT Topics
To answer RQ3, we performed sentiment analysis for each
of our identified topics on ChatGPT. For this task, we initially
used Python’s NLTK Library to automatically classify the
tweets per topic. However, upon inspecting the results, we
were not satisfied with some of the automated classifications
of the library. For example, certain tweets that we identified
as neutral (discussing both negative and positive aspects of
ChatGPT) were classified as negative by the library. In addition,
a substantial amount of tweets were accompanied by images
(e.g., Screenshots). We preferred to have a look at the complete
tweet with these images for an overall view of the content
of the tweet. Finally, our previous work, which explored the
most recent trends and emerging solutions of a highly evolving
field, showed that qualitative analysis enables a more in-depth
and nuanced data analysis [27]. Therefore, we chose to do a
manual qualitative analysis of the tweets. We aim to extend
this work with the inclusion of a larger dataset in the future
to assess overall trends with time, where we plan to use an
automated analysis method.
For sentiment analysis, we manually labelled 9 datasets,
where each sample dataset (
Si
, where
i
ranges from 1 to 9)
includes 100 randomly selected tweets for a specific topic
using MS excel. We labelled a tweet with -1, 0, and 1,
where -1 represents negative sentiment, 1 represents positive
sentiment, and 0 represents neutral sentiment. We kept the
neutral sentiment label for tweets that discuss both positive
and negative aspects and tweets that were ambiguous to identify
a specific positive or negative sentiment. We then did open
coding [28] for each tweet to enable us to identify the common
patterns of discussion within a topic. In this task, we identified
keypoints (i.e., summarised points) and then codes (i.e., a phrase
that further summarizes the key point in 2 or 3 words) and
recorded them in the same spreadsheets as an example snapshot
is shown in Fig. 2.
All five authors were involved in the manual labelling and
qualitative analysis of our data. Here, each data set per topic
was labelled by two authors. All disagreements were resolved
through open discussions among the annotators to reach an
agreement for mitigating any possible errors.
IV. RES ULT ANA LYSIS
We describe the results for our RQ1, RQ2, and RQ3 in
Section IV-A, IV-B, and IV-C, respectively.
A. RQ1. What are the characteristics of ChatGPT early
adopters?
Motivation
In this RQ, we aim to analyze the characteristics
of the early adopters of ChatGPT. We analyze the characteristics
in terms of user location, occupation and account verification
status. This analysis will enable us to focus on the character-
istics of the early adopters to understand users from which
professions are interested in ChatGPT and the demographic
distribution of these users.
Approach
We quantitatively analyze the user loca-
tion, user occupation, and user verification status of our
ChatGP T T w eet dataset to answer this RQ.
Result
For user location, Fig. 3 demonstrates that early
adopters are geographically dispersed, where the majority of
the tweets originated from the North American and Asian
4
2%
98%
Verified
Not Verif ied
Fig. 4: Distribution of the verification status of the early
adopters of ChatGPT based on Twitter
16%
10%
8%
6%
6%
6%
6%
5%
37%
Software Practitioner
Academics/Researcher
Student
Data Scientist
Investor
Business Analyst
Entertainer
Journalist
Other
Fig. 5: Occupation distribution of the early adopters of ChatGPT
based on tweets
regions, and a similar number of tweets from Europe, Australia
and South America. In our analysis, we found USA, India,
UK, Canada, and Germany are the top-5 countries in terms of
expressing their opinion while adopting ChatGPT.
Besides, our analysis identifies that only 2% of the early
adopters of ChatGPT are verified Twitter users as shown in
Fig. 4. It implies that ChatGPT is not confined only to verified
users; rather, it is widely adopted across all Twitter users.
For user occupation, we identified a broad, wide, and
diverse range of early adopter communities for ChatGPT.
In our analysis, we also identified Twitter users with no
specific user job description, which we showed as other in
our occupation distribution that is depicted in Fig. 5. The
top 3 user occupations of the early adopters of ChatGPT
are software practitioners, academics, and students. Other
occupations include data scientists, investors, business analysts,
entertainers (e.g., artists, singers, actors), and journalists, as
shown in Fig. 5. It implies the popularity and interest in
adopting AI-based ChatGPT across diverse occupations.
For RQ1, we identified that the early adopters of ChatGPT
are located in geographically dispersed regions with a diverse
and broad range of professions.
2044
1790
3072
2044
1981
1234
1981
2578
4435
T1 Disruptions for Software…
T2 Entertainment and Exercising…
T3 Natural Language Processing
T4 Impact on Educational Aspects
T5 Chatbot Intelligence
T6 Impact on Business…
T7 Implications for Search Engines
T8 Q&A Testing
T9 Future Career & Opportunite
Number of Tweets
Fig. 6: Distribution of topics based on the number of tweets
captured by Topic Modeling
B. RQ2. ChatGPT Topics - What are the main topics that are
being discussed about ChatGPT in Twitter?
Motivation.
ChatGPT has attained significant attraction
over broad and diverse communities, which includes not only
researchers, managers, and practitioners, but also entertainers,
business analysts, and educationists.
For instance, the registered number of users for ChatGPT
has risen up to 1 Million within 5 days since its beta-
release [4], whereas Facebook, Netflix, and Instagram required
approximately 300, 1200, and 75 days to reach 1 Million users.
An answer to this RQ will identify the most common and
pressing ChatGPT topics that the communities have frequently
encountered while using ChatGPT. This identification will
enable us to understand how diverse communities express their
experience and insights over different domains. We assert that
this understanding is invaluable to assess ChatGPT’s capability,
effectiveness, and usability.
Approach.
We identified the optimal number of topics for
our
ChatGP T
dataset by using LDA as discussed in Section
III. After identifying the optimal number of topics, we examined
the top 20 keywords and randomly selected 30 tweets from
each of the identified topics to select a suitable name, which
provides the best representation for the group of tweets under
that topic. All the authors were involved in the examination
and discussed extensively to reach a consensus on naming the
topics. This approach for naming the topics is common and
widely adopted in the literature [21], [22], [24], [25], [29],
[30].
Results.
We identified 9 topics as shown in Table I that have
been discussed among the early adopters of ChatGPT. Fig. 6
depicts the distribution of these topics based on the number of
tweets captured by Topic Modelling. We describe these results
in detail below.
1) Disruptions to software development: A key area dis-
cussed in the extracted dataset is the potential disruptions
ChatGPT will cause to current software development practices.
For example, users discussed many examples of how ChatGPT
can be used to generate code, assist with debugging, and even
5
TABLE I: Topic Names and top 10 words (lemmatized topic words) for our ChatGP T T weet topics
Sl Topic Name Topic keywords
T1 Disruptions for Software Development code, write, create, program, generate, python, script, developer, error, run
T2 Entertainment and Exercising Creativity write, story, poem, love, fun, short, joke, style, funny, movie
T3 Natural Language Processing model, language, generate, data, text, prompt, human, conversation, learn, response
T4 Impact on Educational Aspects write, student, paper, essay, plan, research, education, school, assignment, teach, homework
T5 Chatbot Intelligence chatbot, intelligence, artificialintelligence, machinelearning, artificial, user, million, robot, security, app
T6 Impact on Business development time, startup, business, company, service, true, idea, control, market, customer
T7 Implications for Search Engines google, search, answer, engine, replace, result, source, StackOverflow, query, internet, information, avlable
T8 Q&A Testing question, answer, wrong, test, response, correct, amp, pretty, simple, solve
T9 Future Careers & Opportunities tool, future, time, people, technology, potential, job, world, change, learn
perform tasks like summarizing and translating code.
2) Entertainment and exercising creativity: Twitter users
were widely using ChatGPT for entertainment purposes, gener-
ating poems, jokes or other humorous write-ups. To generate
entertaining outputs using a ChatGPT, users need to provide
the model with some initial text to work with, such as a prompt
or seed text. A popular use case in this area was efforts to
combine characteristics of different entities (e.g., movie or TV
characters, popular personalities and concepts) in one amusing
write-up.
3) Natural Language Processing: The promising Natural
Language Processing (NLP) capabilities of ChatGPT en-
able it to generate understandable natural human language
that facilitates ChatGPT users in perceiving, comprehending,
and projecting generated text. Moreover, the ChatGPT NLP
capability significantly enhances its usability, utility, and
efficiency. Therefore, we have observed that early adopters
have meticulously discussed the NLP aspect of ChatGPT in
the extracted dataset. For example, users have discussed the
amalgamation of AI and NLP in ChatGPT technology in their
tweets. This enables ChatGPT to generate and understand
human-like texts as an outcome.
4) Impact on Educational Aspects: The early adopters of
ChatGPT are considering ChatGPT as one of the technologies
to change the traditional way of education. Users are discussing
how ChatGPT can be used for different purposes in the
education domain. For example, users discussed the use of
ChatGPT for early childhood learning, developing syllabi,
scientific literature review, and crisis management learning,
such as safety plans for thoughts of suicide.
5) Chatbot Intelligence: Twitter data reveals that the intelli-
gence of ChatGPT has become a prominent discussion point
among early adopters. The AI capability of ChatGPT enables
it to understand and respond to users’ queries in a meaningful
manner. With the help of AI, ChatGPT can entertain quite
complex and challenging problems that traditional chatbots fail
to understand accurately. For example, ChatGPT can write and
debug complex codes, solve large-scale optimization problems,
and provide accurate responses to complicated queries.
6) Impacts on Business Development: The application of
ChatGPT in business analysis is being widely discussed on
Twitter. Users are discussing the use of ChatGPT for developing
startup pitches, generating business use cases, and creating
business plans. Users are also asking ChatGPT to provide
financial advice.
7) Implications for Search Engines: One of the widely
discussed topics is using ChatGPT as a search engine to query
and retrieve a wide range of information. Exceeding the existing
search engine (e.g., Google Search, Bing) functionalities,
ChatGPT has provided users with a novel experience of
presenting information conveniently by selecting the most
appropriate information and explaining it in simple terms.
8) Q&A Testing: Q&A Testing is the third largest topic in
our
ChatGP T T w eet
dataset. In this topic, adopters typically
use ChatGPT to learn, compare, and verify answers for different
academic subjects (e.g., physics, mathematics, chemistry),
or/and conceptual subjects (e.g., philosophy, religion), among
others. We identified adopters who also ask open-ended and
analytical questions to ChatGPT in order to understand the
capability of ChatGPT. Furthermore, we observed several
questions on complex technical and emerging subjects as well.
9) Future Careers & Opportunities: This topic is the largest
discussed topic for ChatGPT, where early adopters share
their insights and ideas on how this advent AI technology
can influence future career opportunities. Early adopters also
discussed what we require in terms of skills, knowledge,
values, and behaviour to cope-up with this advanced technology.
Besides, we identified early adopters’ significant insight on
how the industry and research organizations can collaborate to
stay up-to-date on the latest developments in AI and ensure
that sustainable growth and progress are achieved and remain
relevant to the social, emotional, and cognitive aspects of
humans.
C. RQ3. ChatGPT Sentiments - What are the sentiments that
are being expressed about ChatGPT topics on Twitter?
Motivation.
Earlier research efforts [9], [15] in analyzing
Twitter sentiments (i.e., positive, negative, or neutral opinion)
asserted that the sentiment analysis is extremely helpful in
determining the public or user perception towards a product,
service, or technology. Sentiment analysis is also significant
since it can impact the longevity of a product, service, and
technology [31]–[33]. As a highly trending chatbot technology,
user communities expressed their sentiments while using
ChatGPT on Twitter. Answers to this RQ will help the ChatGPT
decision makers, where they need to act promptly, as Twitter
sentiments typically provide emotion-rich information.
Approach.
We followed the approach described in Section
III-E.
6
81
92
83
52
78
75
54
38
75
6
1
14
32
20
5
15
40
16
13
7
3
16
2
20
31
22
9
010 20 30 40 50 60 70 80 90 100
T1 Disruptions for Software Development
T2 Entertainment and Exercising Creativity
T3 Natural Language Processing
T4 Impact on Educational Aspects
T5 Chatbot Intelligence
T6 Impact on Business Development
T7 Implications for Search Engines
T8 Q&A Testing
T9 Future Career & Opportunite
Sentiment Polarity (%)
Positive Negitive Neutral
Fig. 7: Results of the qualitative sentiment analysis per topic
Results.
Fig. 7 depicts the summary of the percentage values
(Positive, Negative and Neutral) for each topic. Below, we
describe the results in detail for each topic.
1) Disruptions to software development: The sentiment
analysis returned 81% positive for this topic. Early adopters
were particularly impressed by ChatGPT’s abilities to assist in
coding tasks. We captured many tweets of users giving specific
examples of how ChatGPT assisted them with their software
development activities.
If you are a programmer like me (kind of lazy one!)
and not using GitHub #copilot, just try #ChatGPT.
This chatty lady wrote an API in #Reactjs in seconds!
Yes, I’m a programmer now. Thanks to #AI and
#OpenAI.
I think this is the most disruptive tech [..].
Another popular use case with positive sentiments was the
assistance this technology provides with regard to debugging
and error handling. Users expressed their satisfaction with how
ChatGPT provided them with specific assistance with sorting
out their coding-related troubleshooting as compared with the
other available services.
ChatGPT is a game changer for programmers and
coders! It can help with everything from debug-
ging and troubleshooting to providing guidance and
suggestions for code improvements. I’m loving the
assistance it provides. #chatgpt #python #R #HTML
#SQL #Java #JavaScript #C #Swift #PHP
Only a very less amount of users (i.e., 6) expressed concerns
about the disruptions ChatGPT will cause to the software
development practices. Some users noted that users need to be
mindful of how they use ChatGPT for specific development
assistance.
As many have said, #chatGPT runs the risk of spitting
out plausible-sounding but wrong answers in some
cases. It’s especially bad for generating code that
uses external libraries that change over time, as one
might expect. Simple logic is good tho
We categorized this tweet as neutral as it points out both
positive and negative aspects of ChatGPT use. Overall, users
are particularly positive that this technology has the potential
to save developers time and effort in their day-to-day tasks
drastically.
2) Entertainment and exercising creativity: Naturally, the
sentiment analysis for this topic turned out to be overwhelm-
ingly positive (i.e., 92%). Twitter users were widely using
ChatGPT for entertainment purposes, generating poems, jokes
or other humorous write-ups. To generate entertaining outputs
using a ChatGPT, users need to provide the model with some
initial text to work with, such as a prompt or seed text. A
popular use in this area was efforts to combine characteristics
of different entities (e.g., movie or TV characters, popular
personalities and concepts) in one amusing write-up.
I asked #ChatGPT “In the style of a Shakespearean
play, write a scene from the sitcom Friends, in which
Chandler accuses Joey of stealing his banana which
he had been saving to eat for his lunch.”
Users also attempted to exercise their creativity by generating
interesting ideas, such as short stories or poems.
7
#ChatGPT write a short story about a dog named
Baxter who has died and is in heaven enjoying
cheeseburgers and beaches, who would like to solve
the mystery of the stolen Christmas dinner ham, in
the style of Arthur Conan Doyle.
We observed that a screenshot of the ChatGPT query output
is usually accompanied with these tweets. By analyzing the
selected tweets for sentiment analysis on this topic on Twitter
itself, we recognised that the ChatGPT-generated output was
very positively received as they had a high engagement. This
might have contributed to the drastic rise in the popularity of
this technology, as evidenced by the large initial number of
subscribers.
3) Natural Language Processing.: We observed diverse
sentiments (i.e., positive, negative, and neutral) of early adopters
of ChatGPT technology in our Twitter dataset. While most
tweets (i.e., 83%) imply satisfaction with ChatGPT, some
tweets (14%) have expressed their concerns. Only 3% of the
tweets reported a neutral viewpoint. For example, early adopters
seem happy with the amalgam of NLP and AI in ChatGPT
technology.
I’m really impressed with the natural language
processing capabilities of ChatGPT’s AI-powered
chatbot. It’s a great example of the power of #NLP
#AI technology. #chatgpt
Early adopters were also surprised by the realistic human-like
text generation of ChatGPT.
#ChatGPT is a prototype dialogue-based #AI chatbot
capable of understanding natural human language
and generating impressively detailed human-like
written text.
The use of human AI trainers and supervised fine-tuning
improves the ChatGPT text generation and perception. However,
some users have shown their concerns about the quality
of the generated text. They argue that ChatGPT provides
misinformation due to the lack of critical-thinking skills,
nuance, and ethical decision-making ability of ChatGPT.
I can always spot generated text - it lacks the depth
and complexity of human thought
In summary, though early adopters raised some negative
sentiments, the majority of ChatGPT users showed positive
responses. We believe ChatGPT is a promising stepping stone
to the development of an absolute human-like chatbot.
4) Chatbot Intelligence.: In the extracted dataset, we have
identified both the positive and negative sentiments of early
adopters regarding the intelligence capability of ChatGPT. As
per our quantitative analysis, 78% tweets showed a favourable
sentiment toward ChatGPT intelligence, 20% tweets raised
harmful impacts of AI for ChatGPT, and 2% tweets had a
neutral opinion. This shows that the majority of early adopters
are in favour of ChatGPT intelligence capability.
[...] ChatGPT could be a big help and see the
amazing wonders of AI (Artificial Intelligence) in
the future. @OpenAI I typed "Create React App"
and it provided me a clear to read instructions on
creating the application. #chatGPT
On the other hand, early adopters have also raised their
negative sentiments about AI involvement in ChatGPT. For
example, users were concerned about the ominous impacts of
AI-based ChatGPT on society such as an increase in terrorism,
hacking, and unemployment.
ChatGPT artificial intelligence developed by OpenAI
explains how to bomb and theft
In conclusion, ChatGPT’s intelligence capability provides
both positive and negative impacts on society. However, we
recommend an effective policy/regulation development for the
usage of ChatGPT. This mitigates the negative impacts of Chat-
GPT intelligence and hence improves ChatGPT acceptability
among its users in the future.
5) Impact on Educational Aspects: Unlike the other topics
(e.g., software development) where ChatGPT is accepted
overwhelmingly, the adoption of ChatGPT for educational
purposes raised both positive and negative perceptions among
the users. Our analysis identified 52% positive, 32% negative,
and 16% neutral views on the use of ChatGPT for educational
purposes.
Users are accepting ChatGPT for diverse application areas
in the education domain. For example, users discussed many
examples of how ChatGPT can be used for grading papers,
assessing students’ learning, and preparing syllabi. ChatGPT is
also considered as a good personal teacher for a student. Users
are considering ChatGPT as one of the many technologies
that will revolutionise the traditional way of teaching and
assessment.
#ChatGPT will change education as we know it. I am
hopeful. Perhaps this is what will springboard age-
old assessment practices to ones that authentically
assess student learning
On the contrary, the use of ChatGPT by students for writing
essays, preparing assignments, and home-works was raised
as a common public concern on Twitter. Users are concerned
about how the use of ChatGPT for preparing assignments
by students can hinder their learning process. Other concerns
include plagiarism detection for the students’ assignments and
shallow answers to ChatGPT in response to research questions.
With the rise of the amazing #ChatGPT, I am sure
many students will use it to write essays confidently
yet they learn nothing. Also plagiarism software
cannot detect this, even this may not be considered
as plagiarism at all
In summary, though there are some negative perceptions
regarding the adoption ChatGPT for preparing assignments
by students, many users are arguing against this concern as
they believe that ChatGPT is exceptionally good at identifying
text produced by AI, which enables it to mitigate students’
plagiarism.
ChatGPT is amazing, but it will be neither an
effective way for students to plagiarize nor replace
humanistic education.
8
6) Impact on Business Development: The adoption of
ChatGPT in business is being discussed positively with a ray
of hope for future possibilities. Our analysis identified 75%
positive, 5% negative, and 20% neutral views on the use of
ChatGPT in business.
Users are amazed at the response of ChatGPT to build an
elevator pitch for a new idea to investors which can significantly
support the start-up companies. For example, startup people
are hopeful to imagine ChatGPT as a technical co-founder.
ChatGPT is also expected to strongly support business decision-
making.
With ChatGPT and AI-assisted coding, tech busi-
nesses will become smaller and flatter. ChatGPT will
help good coders become massively more productive.
AND it will help management at all levels make better
decisions.
Users are also discussing the use of ChatGPT to create a
business plan for app development, write trading strategies, and
generate business-specific use cases. Besides, users of ChatGPT
are discussing diverse potential application areas of ChatGPT
from a business perspective. For example, the use of ChatGPT
to provide technical customer support and financial advice.
Imagine using ChatGPT as a technical support
specialist at Apple - personalized and accurate
responses to customer inquiries in seconds! This
is just one example of how ChatGPT can improve
customer service for any company.
Our analysis identified that ChatGPT is positively accepted
to have a promising future in the business domain with its
positive support in all business aspects from business plan
development to customer service support.
7) Implications for Search Engines: Our analysis identified
54% positive, 15% negative and 31% neutral sentiments on this
topic. It implies that most of the tweets present either a positive
or neutral sentiment over negative sentiments. For example,
some users have mentioned ChatGPT as a #googlekiller,
indicating that ChatGPT poses a significant threat to existing
search engines. This indicates that the users tend to consider
ChatGPT as a replacement for the current search engines.
Just relaized I have been using google to access
#ChatGPT the apparent #googlekiller.
Furthermore, some ChatGPT users have stated that ChatGPT
performs better in search speed and accuracy than the current
search engines and knowledge-sharing platforms, respectively.
This implies the potential future risks not only for search engine
services but also for other knowledge-sharing related platforms
such as Stack Overflow, Quora and Wikipedia.
On information side #ChatGPT is like: - Google but
more exact - Quora but faster [...].
We came across several instances of negative sentiments
concerning ChatGPT’s ability to provide accurate information
as it is leveraging the internet as its primary data source.
Furthermore, its also been remarked that some users are yet
to understand these limitations of ChatGPT. Thus, users are
encouraged to be critical and verify ChatGPT results from
other sources before using them in their intended applications.
It has become evident that you cannot blindly trust
#ChatGPT; data sourced from the internet is not
completely "verified" or accurate. Most people are
able to recognize this fact for ChatGPT, but not for
the internet as a whole. Be critical of the information
you find online.
Interestingly, several solutions (e.g., Google Chrome exten-
sion) have already emerged to address the above-identified
concerns. These plugins can retrieve and present search results
from ChatGPT and existing search engines, such as Google,
in a single view, facilitating the user to make better-informed
decisions. Therefore, we foresee ChatGPT significantly revolu-
tionizing and enhancing the user experience through improved
user interaction models in future search engines and knowledge-
sharing platforms.
8) Q&A Testing: Our analysis identified 38% positive, 40%
neutral, and 22% negative sentiments while adopting ChatGPT
for testing its capability in terms of questions and answering.
The positive sentiments for Q&A Testing topics are evolved
due to the quality, human-friendly interaction, fast responses,
and the provision of reasons for the generated answers. For
instance, we observed positive sentiments for providing users
with quality answers, which enables us to get the expected
value from ChatGPT.
I recently used #Chatgpt to help me better understand
conditional types in English, and I was impressed
by the quality of the exercises it provided. I found
that by giving it more precise and well-structured
questions, I was able to get even more value out of
it.
Besides, adopters are enthusiastic about the way ChatGPT
interacts with the users while answering their questions.
This is really spot on. I played 20 questions with
#ChatGPT yesterday. First, it guessing correctly who
I and my daughter was thinking about (Rapunzel).
And then us guessing that it was thinking about Billie
Eilish. Really felt like an interaction with a new kind
of entity.
Another cause for the positive sentiments for this topic is the
capability of ChatGPT for providing the reasons behind the
answer.
Nurturing curiosity in children is the best thing par-
ents, grandparents, teachers, caretakers, etc. could
do. when the eventual 1st, 2nd...n "why" question
hits you, might as well use #ChatGPT instead of the
catastrophic "because I told you so".
The negative sentiments for Q&A Testing topics are evolved
due to the wrong, incorrect, and sometimes invalid answers
for their questions.
"#OpenAI #ChatGPT I used this AI to find my
chem[istry] answers but got the wrong answers. for
example, Answer has to be in gms, and it’s giving
me in moles...
9
Besides, we identified some adopters have justified their
sentiments in this topic by mentioning that Stack Overflow has
banned to post answers generated by ChatGPT.
Stack Overflow has temporarily banned responses
generated by OpenAI’s #ChatGPT AI, citing “a high
rate" of incorrect answers. ChatGPT, can answer
questions about coding problems but often produces
“plausible-sounding but incorrect or nonsensical an-
swers,".
One significant concern raised by the early adopters of ChatGPT
in Q&A topic is the confidence shown for wrong or incorrect
answers, which may hamper the ChatGPT’s wider adoptability.
The primary problem is the answers which #ChatGPT
produces have a high rate of being incorrect...The
scary part was just how confidently incorrect it was.
9) Future Careers & Opportunities: Our analysis identified
75% positive, 9% neutral, and 16% negative sentiments while
early adopters were expressing their opinions regarding future
careers and opportunities. The positive sentiments for Future
Careers & Opportunities topics are evolved due to the fast and
effective solution provided by the ChatGPT, which can be a
crucial factor for successful careers.
AI coding and decision support systems will enable
very small teams to create massively valuable prod-
ucts, services, and companies.
Besides, we observed positive sentiments for ChatGPT’s
innovation, learning, and adaptation to make the job tasks
convenient and easy.
AI is the future! With its ability to learn and adapt,
it’s no surprise that more and more companies are
turning to this technology to improve their operations
and better serve their customers. Exciting times
ahead! #AI #innovation #ChatGPT #OpenAI
Besides, adopters are enthusiastic about the customization
feature of ChatGPT to provide personalized messages, services,
or answers, which adopters believed to help them to be more
engaged in future with their customers.
Imagine using #ChatGPT as a technical support spe-
cialist at [...] - personalized and accurate responses
to customer inquiries in seconds! This is just one
example of how ChatGPT can improve customer
service for any company.
The negative sentiments for Future Careers & Opportunities
topics are evolved due to fear among people of losing their
jobs to ChatGPT.
#ChatGPT ready to replace Product Managers...
Besides, we identified some adopters who have expressed to
lose their software programmer jobs as ChatGPT can perform
a magnitude of software development activities within a very
short time, which may take a long time for a human software
programmer.
More importantly, GPT will create a huge turmoil in
the IT industry. Many coding jobs will be taken over
by this AI, as it is able to do a multitude of coding
jobs within seconds (which takes humans many weeks
to do). This is a watershed moment in tech history.
V. IMPLICATIONS
Considering the increasing popularity of ChatGPT technol-
ogy, we conducted this study to identify and analyze topics with
regard to the sentiments of ChatGPT early adopters. The topic
modelling and sentiment analysis results provide an idea of
how ChatGPT technology is being perceived among early users,
and it also indicates the potential for ChatGPT’s acceptance in
the future. In this section, we discuss implications for users
and researchers based on our results.
A. Implications for Users
Our study identifies several benefits for users of ChatGPT.
For example, as described in
IV-B
1, ChatGPT has the potential
to change traditional ways of software development, which
shows that software developers can utilize ChatGPT while
creating software and ensuring an effective software develop-
ment process. Similarly, as described in
IV-B
6, ChatGPT can
help business developers create their feasibility reports and
business cases to enable smooth operations and other business
development-related tasks. For example, a user might query
business ideas for a specific context, and the model could
generate a list of possible ideas which can be used as a starting
point for brainstorming.
Our study has recognized several pitfalls that users must
consider and weigh when integrating ChatGPT into their
workflows and applications. For example, since ChatGPT
results are not verified or fact-checked by any established
authority, users must not solely depend on ChatGPT results
to perform any critical task. Therefore, users must critically
analyze ChatGPT results and consult other verified data sources
to make sensible decisions. Similarly, users must be cautious
about sharing their personal and confidential data with ChatGPT,
as adversarial attacks on ChatGPT language models can lead
to possible data breaches, leaving ChatGPT users vulnerable. It
is also essential to highlight the responsible usage of ChatGPT
from the users’ end, as integrating ChatGPT’s creations and
compositions into commercial and academic works can lead
to legal and ethical concerns.
B. Implications for Researchers
Our study provides potential future research directions to
researchers for exploring concerns related to the ethical use
of ChatGPT and other practical implications. For example,
users raised concerns about this technology being misused
in writing or completing educational activities such as essays
or assignments, which would hamper students’ learning. This
issue was captured in the Impacts on Educational Aspects topic.
Further, the topics Q&A Testing captured concerns related to
misinformation or inaccurate outputs being generated from
ChatGPT. These issues indicate that researchers should explore
means of mitigating these ethical and practical implications.
For example, effective protocols for ChatGPT’s usage can be
developed to ensure its ethical usage across different domains.
10
VI. TH RE ATS TO VALIDITY
Firstly, our study focuses on tweets from Twitter as a
representative of the sentiments of early adopters of ChatGPT.
Twitter was the 9th most visited website globally in 2021
[34], which is an enriched knowledge base for analyzing
human sentiment. However, future research can focus on
other resources, such as Stack Overflow and blogs, to further
generalize our findings.
Secondly, manual topic-wise sentiment analysis can be
subject to human judgment bias. To mitigate this threat to
validity, each sample dataset of a specific topic was labelled by
two researchers, and any disagreements were resolved through
discussion.
Thirdly, while threat modelling is useful in handling a large
amount of data, its usage introduces some threats. For example,
identifying the optimal value for a number of topics
N
is a
potential threat. To mitigate it, we followed the commonly
used method of experimentation with a broad range of values
for N[21], [35].
VII. CONCLUSION
In conclusion, our study of early adopters’ sentiments
about ChatGPT revealed overwhelming excitement and limited
concerns about this application of a large language model.
The majority of users were impressed by the performance
of ChatGPT and the potential of large language models to
assist with tasks related to several domains (e.g., Software
development, Business initiatives and analysis, NLP). However,
there are also important ethical implications that need to be
considered in ChatGPT use and further development. For
example, some users were concerned about the negative
effect it would have on the education industry activities such
as take-home assignments and essay writing for students.
Overall, our study provides valuable insights into the sentiments
of early adopters of ChatGPT and highlights the need for
continued research and dialogue to develop best practices for
the responsible use of large language models.
REFERENCES
[1]
S. Lock. (2022) What is ai chatbot phenomenon chatgpt and
could it replace humans? Accessed December 10, 2022. [On-
line]. Available: https://www.theguardian.com/technology/2022/dec/05/
what-is- ai-chatbot- phenomenon-chatgpt- and-could- it-replace- humans
[2]
(2022) Chatgpt: Optimizing language models for dialogue. Accessed
December 10, 2022. [Online]. Available: https://openai.com/blog/chatgpt/
[3]
K. Naidu. (2022) Chatgpt is a new ai chatbot that can find errors in
your code and write you a story. Accessed December 10, 2022. [Online].
Available: https://crazyengineer.in/chatgpt-is-a- new-ai-chatbot/
[4]
J. Sier. (2022) Chatgpt takes the internet by
storm, bad poetry and all. Accessed December 10,
2022. [Online]. Available: https://www.afr.com/technology/
chatgpt-takes- the-internet- by-storm- bad-poetry- and-all- 20221207-p5c4hv
[5]
A. Friedman, “Ai chatbot predicted to replace google in a couple of
years,” Dec 2022. [Online]. Available: https://www.phonearena.com/
news/ai-chatbot-could-replace-google-soon_id144120
[6]
A. Hern. (2022) Ai bot chatgpt stuns academics with essay-
writing skills and usability. Accessed December 10, 2022. [On-
line]. Available: https://www.theguardian.com/technology/2022/dec/04/
ai-bot- chatgpt-stuns- academics-with- essay-writing- skills-and- usability
[7]
H. Anber, A. Salah, and A. Abd El-Aziz, “A literature review on
twitter data analysis,” International Journal of Computer and Electrical
Engineering, vol. 8, no. 3, p. 241, 2016.
[8]
V. Sahayak, V. Shete, and A. Pathan, “Sentiment analysis on twitter data,”
International Journal of Innovative Research in Advanced Engineering
(IJIRAE), vol. 2, no. 1, pp. 178–183, 2015.
[9]
J. Bian, K. Yoshigoe, A. Hicks, J. Yuan, Z. He, M. Xie, Y. Guo,
M. Prosperi, R. Salloum, and F. Modave, “Mining twitter to assess
the public perception of the “internet of things”,” PloS one, vol. 11,
no. 7, p. e0158450, 2016.
[10]
J.-W. Guo, C. L. Radloff, S. E. Wawrzynski, and K. G. Cloyes, “Mining
twitter to explore the emergence of covid-19 symptoms,” Public Health
Nursing, vol. 37, no. 6, pp. 934–940, 2020.
[11]
J. Bian, U. Topaloglu, and F. Yu, “Towards large-scale twitter mining for
drug-related adverse events,” in Proceedings of the 2012 international
workshop on Smart health and wellbeing, 2012, pp. 25–32.
[12]
W. Wang, I. Hernandez, D. A. Newman, J. He, and J. Bian, “Twitter
analysis: Studying us weekly trends in work stress and emotion,” Applied
Psychology, vol. 65, no. 2, pp. 355–378, 2016.
[13]
M. Paul and M. Dredze, “You are what you tweet: Analyzing twitter for
public health,” in Proceedings of the International AAAI Conference on
Web and Social Media, vol. 5, no. 1, 2011, pp. 265–272.
[14]
E. Aramaki, S. Maskawa, and M. Morita, “Twitter catches the flu:
detecting influenza epidemics using twitter,” in Proceedings of the 2011
Conference on empirical methods in natural language processing, 2011,
pp. 1568–1576.
[15]
S. Trivedi and N. Patel, “Mining public opinion about hybrid working
with roberta,” Empirical Quests for Management Essences, vol. 2, no. 1,
pp. 31–44, 2022.
[16] B. Alhijawi and A. Awajan, “Prediction of movie success using twitter
temporal mining,” in Proceedings of Sixth International Congress on
Information and Communication Technology. Springer, 2022, pp. 105–
116.
[17]
A. Nawaz, T. Ali, Y. Hafeez, M. R. Rashid et al., “Mining public opinion:
a sentiment based forecasting for democratic elections of pakistan,”
Spatial Information Research, vol. 30, no. 1, pp. 169–181, 2022.
[18]
E. K. Steven Bird and E. Loper, Natural Language Processing with
Python. O’Reilly Media Inc, 2009.
[19]
Z. T. Sworna, C. Islam, and M. A. Babar, “Apiro: A framework for
automated security tools api recommendation,” ACM Transactions on
Software Engineering and Methodology, 2022.
[20]
D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent dirichlet allocation,”
Journal of machine Learning research, vol. 3, no. Jan, pp. 993–1022,
2003.
[21]
M. U. Haque, L. H. Iwaya, and M. A. Babar, “Challenges in docker
development: A large-scale study using stack overflow,” in Proceedings
of the 14th ACM/IEEE International Symposium on Empirical Software
Engineering and Measurement (ESEM), 2020, pp. 1–11.
[22]
M. Zahedi, R. N. Rajapakse, and M. A. Babar, “Mining questions asked
about continuous software engineering: A case study of stack overflow,”
in Proceedings of the evaluation and assessment in software engineering,
2020, pp. 41–50.
[23]
A. K. McCallum, “Mallet: A machine learning for language toolkit,”
2002, http://mallet.cs.umass.edu.
[24]
S. Ahmed and M. Bagherzadeh, “What do concurrency developers ask
about? a large-scale study using stack overflow,” in Proceedings of
the 12th ACM/IEEE International Symposium on Empirical Software
Engineering and Measurement (ESEM), 2018, pp. 1–10.
[25]
M. Bagherzadeh and R. Khatchadourian, “Going big: a large-scale study
on what big data developers ask,” in Proceedings of the 2019 27th
ACM Joint Meeting on European Software Engineering Conference and
Symposium on the Foundations of Software Engineering, 2019, pp. 432–
442.
[26]
M. Röder, A. Both, and A. Hinneburg, “Exploring the space of topic
coherence measures,” in Proceedings of the eighth ACM international
conference on Web search and data mining, 2015, pp. 399–408.
[27]
R. N. Rajapakse, M. Zahedi, and M. A. Babar, “Collaborative application
security testing for devsecops: An empirical analysis of challenges, best
practices and tool support,” arXiv preprint arXiv:2211.06953, 2022.
[28]
A. Strauss and J. M. Corbin, Grounded theory in practice. Sage, 1997.
[29]
Z. Wan, X. Xia, and A. E. Hassan, “What is discussed about blockchain?
a case study on the use of balanced lda and the reference architecture
of a domain to capture online discussions about blockchain platforms
across the stack exchange communities,” IEEE Transactions on Software
Engineering, no. 01, pp. 1–1, 2019.
[30]
A. Abdellatif, D. Costa, K. Badran, R. Abdalkareem, and E. Shihab,
“Challenges in chatbot development: A study of stack overflow posts,”
11
in Proceedings of the 17th international conference on mining software
repositories, 2020, pp. 174–185.
[31]
A. Pak and P. Paroubek, “Twitter as a corpus for sentiment analysis and
opinion mining,” in Proceedings of the Seventh International Conference
on Language Resources and Evaluation (LREC’10), 2010.
[32]
B. Gokulakrishnan, P. Priyanthan, T. Ragavan, N. Prasath, and A. Perera,
“Opinion mining and sentiment analysis on a twitter data stream,” in
International conference on advances in ICT for emerging regions
(ICTer2012). IEEE, 2012, pp. 182–188.
[33]
S. Wakade, C. Shekar, K. J. Liszka, and C.-C. Chan, “Text mining for
sentiment analysis of twitter data,” in Proceedings of the International
Conference on Information and Knowledge Engineering (IKE). The
Steering Committee of The World Congress in Computer Science,
Computer . . . , 2012, p. 1.
[34]
C. Beveridge. (2022) 33 twitter stats that matter to marketers
in 2023. Accessed December 10, 2022. [Online]. Available: https:
//blog.hootsuite.com/twitter-statistics/
[35]
K. Stevens, P. Kegelmeyer, D. Andrzejewski, and D. Buttler, “Exploring
topic coherence over many models and many topics,” in Proceedings of
the 2012 Joint Conference on Empirical Methods in Natural Language
Processing and Computational Natural Language Learning. Association
for Computational Linguistics, 2012, pp. 952–961.
12