An Innovative Approach to Develop Persona from Application Reviews

Dylan Clements, Elysia Giannis, Fi Crowe, Mike Balapitiya, Jason Marshall, Paul Papadopoulos and

Tanjila Kanij

Department of Software Systems and Cybersecurity, Monash University, Clayton, Melbourne, Australia

Keywords:

Persona, E-commerce, Application Reviews.

Abstract:

Software end users are diverse by nature and their different facets inﬂuence the way they use software. An un-

derstanding of the users and their needs are achieved by engaging with the users during requirement engineer-

ing. However, sometimes recruiting users during requirement engineering phase can be very challenging. An

accessible way to understand a user’s perspective and traits is through user application reviews. This research

paper proposes an innovative approach to develop user personas from a data set of e-commerce application

user reviews by using GPT-3 and PATHY. This enables the development teams to see different demographic

data, as well as overall frustrations and expectations that users of their platform possess, so developers know

how to enhance their software solutions. This is also helpful to developers of new e-commerce applications.

1 INTRODUCTION

Most software solutions provide an inadequate user

experience due to the developers lacking the under-

standing of enduser needs, as Mathews et al. states

“end user diversity is not sufﬁciently contemplated”

(Mathews et al., ). This ultimately limits the suc-

cession of these platforms. As such understanding

the endusers of a software is essential. Generally

this is done through rigorous user research during

the requirement engineering phase. However, the

enduser characteristics, as well as their preferences,

keep changing and software needs to take these into

account throughout the development as well as after

deployment.

Understanding end users through rigorous re-

search, is most of the times time, cost and effort-

consuming. Another major challenge is to establish

communication with the end users and get valuable

insight from them. To address these challenges, we

propose an innovative approach to understanding the

end user of software from an alternative source of in-

formation - application reviews provided by the end

users. By collecting a large volume of application re-

views and evaluating a range of characteristic traits of

users, a chance to visualise and assimilate what a par-

ticular user of software would resemble, is presented.

This is what a user persona entails. A persona is a

https://orcid.org/0000-0002-5293-1718

description of a ﬁctional character who will use the

software (Cooper, 1999). By collating particular user

facets and outlining their frustrations, we are able to

develop insightful personas which describe the key

user information and their feelings towards the soft-

ware. This then helps software developers understand

their users better.

To develop the proposed framework we selected

e-commerce as our application domain. It has been

shown that age bias is found in various e-commerce

software solutions, meaning it is harder for people

of certain ages to proﬁciently use these applications

(McIntosh et al., 2021; El Shamy and Hassanein,

2018). Yet, it is unclear which other facets beyond

age affect the way people use e-commerce platforms.

E-commerce is a domain where people of different

demographics purchase goods and services and these

users can also share their experiences of the applica-

tion being used (Obie et al., 2021). This information

can be leveraged, and act as a valuable source of prob-

lem statements, ideas and requests from users, which

could ultimately help development teams understand

potential issues with their software, as well as ﬁnd

ways in which they can be mitigated. Additionally,

e-commerce application usage has risen vastly due

to the evolution of online marketing and changes in

economic and environmental factors (Grundy et al.,

2018). For instance, the COVID-19 pandemic forced

many people to use these platforms out of obligation.

With the population’s increasing interaction with e-

Clements, D., Giannis, E., Crowe, F., Balapitiya, M., Marshall, J., Papadopoulos, P. and Kanij, T.

An Innovative Approach to Develop Persona from Application Reviews.

DOI: 10.5220/0011996000003464

In Proceedings of the 18th International Conference on Evaluation of Novel Approaches to Software Engineering (ENASE 2023), pages 701-708

ISBN: 978-989-758-647-7; ISSN: 2184-4895

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

701

commerce applications, it is imperative that the user

experience is critically considered when developing

and improving these applications to reduce any nega-

tive impacts of insufﬁciently understanding users.

Given the volume of reviews sent to app stores,

many developers do not have the capacity to extract

and analyse meaningful insights from all of them

(Mathews et al., ). Therefore, an automated machine

learning tool that can be used to accurately capture

the user’s experience to get a clearer understanding of

the platform’s end users would be advantageous to de-

velopment teams. This would enable software devel-

opers to focus more on implementing improvements

for the application, rather than manually analysing

their platforms’ reviews (Mao et al., 2005). To assist

e-commerce developers in understanding their end

users more extensively, we can extract certain facets

from the user application reviews of the platforms be-

ing assessed using GPT-3, and segment the responses

based on their characteristics to investigate possible

trends between the sentiment and user facets. This

will help developers discern what types of users are

avidly or apathetically using their application, which

can lead to further investigation on how the sentiment

for users that show lower interest to the platform can

be restored. This also increases the potential for ac-

commodating more diverse users.

The remainder of this paper is structured as fol-

lows: Section 2 details the review of related research

and background of this research. Section 3 describes

our approach in detail. Section 4 presents the results

of analysing e-commerce app reviews and developing

persona. Section 5 illustrates the results of the evalua-

tion of our proposed method, Section 6 demonstrates

the possible threats to the validity of the research. Fi-

nally, section 8 concludes our ﬁndings and outlines

any future research.

2 REVIEW OF LITERATURE

2.1 Automation

Gao et al. attempted to analyse application reviews to

identify existing issues (Gao et al., 2018) and found

that “noise words” (e.g., misspelled words) made it

challenging to identify new app issues through au-

tomation. They derived an approach information ex-

traction removing the noisy words. Similarly, Mal-

gaonkar et al. reduced the burden of human involve-

ment by using natural language processing, feature

engineering, and word sense disambiguation, to au-

tomatically generate the taxonomy(Malgaonkar et al.,

2022). Mcilroy et al. analysed large app reviews and

found that up to 30% of the reviews raise various is-

sues in a single review, including “feature requests”

and “bug reports” (Mcilroy et al., ).

2.2 Facet Research

Grundy outlines that age and gender are some of the

characteristics that need to be better incorporated into

requirements engineering and design (Grundy, 2021).

Tekin and Sebnem support Grundy’s analysis of the

characteristics that inﬂuence one’s use of e-commerce

and amount of online expenditure, with research dis-

playing the effects of gender and age on users (Kose

and Arslan, 2020). An initial study has looked at the

facets of risk aversion, technical proﬁciency, visual

impairment, and attention among people of different

ages while interacting with e-commerce (McIntosh

et al., 2021). Middle-aged individuals were found

to be more likely to engage in e-commerce activities,

compared to relatively younger and older individuals

by Tekin and Sebnem (Kose and Arslan, 2020).

2.2.1 GPT-3 Overview

Generative Pre-Trained Model 3 (GPT-3) by Ope-

nAI is the largest language model constructed to date

(Dale, 2021). The model is trained on 499 billion to-

kens of web content, including all of Wikipedia, a

variety of books, and a large portion of web pages

(Singh et al., 2021). GPT-3 consists of 12 layers of

transformer decoder blocks with 175 billion trainable

parameters (Acheampong et al., 2021). Text weight

embeddings and positional embeddings are passed

as inputs into 96 attention layers, then into a feed-

forward layer, which then outputs a probability distri-

bution (Acheampong et al., 2021).

Advantages. Although GPT-3’s main application

is in automated text generation (Acheampong et al.,

2021; Dale, 2022), GPT-3 can be used in a range of

scenarios, including text classiﬁcation (or sentiment

analysis) (Mathews et al., ). The extensive training

data set for GPT-3 (Mathews et al., ) and its computa-

tional power (Alexandridis et al., 2021) alone makes

it a prime candidate for our application of NLP on

bulk user review data sets.

Limitations. Despite GPT-3’s computational power

and access to a large training set, the model does

have some limitations. GPT-3’s large training set

means that text classiﬁcation is very costly to run and

requires signiﬁcant processing power (Dale, 2021).

This is why GPT-3 is currently only available via API

(Alexandridis et al., 2021; Ashraf and Perez, 2020), as

the typical machine could not handle the large & com-

plex processes of GPT-3 (Acheampong et al., 2021).

In addition to this, given that the training data set

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

702

is comprised of a wide collection of books, all of

Wikipedia, and an abundance of web pages, the model

will be slightly biased, and reﬂect predictions similar

to that of the training data (Mathews et al., ; Dale,

2021; Ashraf and Perez, 2020).

2.3 Persona Development

The fundamental idea behind persona development

is gathering information about the users and group-

ing those into personas. Guo and Ma use a tech-

nique that focuses mainly on three pillars of a per-

sona: biographic background; archetype and person-

ality (Guo and Ma, 2018). Ferreira et al. and Marr

summarise that the key steps to developing personas

for e-commerce platforms are to gather demographic

data, general attitudes toward the application, main

goals and frustrations they bear, and ﬁnally common-

alities between the respective facets so they can be

grouped into personas (Ferreira et al., 2017)(Marr,

2020). All these techniques require rigorous user re-

search.

PATHY is another technique of persona develop-

ment, that is based on empathy mapping (Ferreira

et al., 2016). The PATHY technique bridges the gap

between empathy maps and personas themselves and

elicits theoretical user requirements for an applica-

tion (Ferreira et al., 2016). To improve the support

provided to software developers, a second component

was added in the PATHY technique to deal with issues

related to identifying application features and charac-

teristics (Ferreira et al., 2016). An advantage of the

PATHY technique is that it found more potential ap-

plication requirements than Acu

na et al.’s technique

(Acu

na et al., 2012). Given its clear strengths, we

have chosen to use the PATHY technique for this re-

search.

3 METHODOLOGY

Application reviews were extracted with an auto-

mated script written in Python, from 25 different e-

commerce applications. The applications domains in-

cluded: departmental stores, supermarket, pharmacy,

retailer, food, fashion, and so on. A variety of types

of domains ensures a variation of reviews and subse-

quently a variation of personas.

3.1 Development Environment

A development environment was conﬁgured to al-

low for collaborative execution and near-full automa-

tion of the execution of the research methodology.

This environment included a private GitHub repos-

itory, Python installations (including dependencies)

and JSON ﬁles. The OpenAI python API (OpenAI,

) was used to communicate with GPT-3 to perform

the analysis.

3.2 Facet Extraction

In order to extract user facets and identify their age,

gender, expectations and frustrations from the ex-

tracted reviews, we used Open AI’s deep learning

platform, Generative Pre-trained Transformer 3, or

GPT-3.

For setting up the environment we turned the tem-

perature to zero that ensures most deterministic out-

put is produced by GPT -3. We also selected “text-

davinci-2” algorithm since this was the most ad-

vanced algorithm implemented by GPT-3 at that time.

The following code snippet shows the queries we ex-

ecuted against each review.

prompts = [

”Rate the author’s sentiment on a scale of -1 to 1: ”,

”Predict the author’s age: ”,

”Predict the author’s gender: ”,

”What are the author’s expectations?: ”,

”What are the author’s frustrations?: ”

]

For example review “I’m incredibly unhappy with

how slow this app runs. I can barely use it.”, if we

run the following code “Rate the author’s sentiment

on a scale of -1 to 1: ”, we get output as “-1”, which

indicated negative sentiment.

Three of the facets we intended to capture were to

be expressed quantitatively in our review characteris-

tics data set. We found that age, gender and sentiment

were most of the times expressed quantitatively how-

ever sometimes those were expressed qualitatively.

e.g.: for age, instead of “30-40”, the output would

be “The author’s age is likely to be between 30 and

40”. This was highly unusual despite setting GPT-3’s

temperature to 0, which was said to produce entirely

deterministic results.

To work around this challenge, we developed

intermediate analysis scripts using natural language

processing tools such as NLTK. The intermediate

scripts were used to extract the key points from each

of the qualitative responses. Where any qualitative re-

sponses lacked keywords such as age, gender or sen-

timent scores, it was assumed that there were none

found for that review.

Another challenge was that, some facets review

characteristics were empty. This is most likely due to

two main reasons. The ﬁrst being that requests to the

OpenAI API could have failed, despite the back-off

An Innovative Approach to Develop Persona from Application Reviews

703

Figure 1: Persona development Diagram.

mechanism working successfully. The second, being

that GPT-3 outputted a result that could not be cap-

tured as a valid review characteristic. For example,

GPT-3 could output “There is not enough informa-

tion to determine the author’s gender”. As a result,

the author’s gender would be marked as null and dis-

counted from further analysis. 3.28% of review facets

analysed reported null values.

3.3 Persona Development

Given that various persona development techniques

are complex to implement and do not explicitly guide

designers in identifying development relevant infor-

mation, we decided to utilise a more methodical tech-

nique - PATHY. PATHY utilises empathy map check-

list questions to create customer segment proﬁles, and

a template to simplify its implementation. An em-

pathy map reveals the rationale underlying users’ ac-

tions, decisions and choices; therefore it helps in de-

signing for users’ real needs (Ferreira et al., 2016).

Matthews (Gray et al., 2010) proposed four dif-

ferent areas that should be covered when creating an

empathy map: What does the person hear? What does

the person think and feel? What does the person see?

What does the person say and do? Bratsberg, H.M.

(Bratsberg, 2012) mentioned Pain and Gain as impor-

tant areas to look for. Based on these we adopted the

following approach for persona development:

1. Found the most common age and gender within

each e-commerce category and added this infor-

mation to a persona.

2. Identiﬁed the pains and gains of said persona from

the most common expectation and frustrations in

user reviews.

3. Given these facets, the team prepared answers to

Matthews’ questions which contributed to their

biography and character understanding.

4 RESULTS

We extracted 4999 app reviews from open source app

store and play store. Among those, 4931 reviews were

analysed. 68 reviews were discarded due to extrane-

ous errors. Figures 2 outline the number of reviews

analysed per category.

Figure 2: Number of reviews per category.

4.1 Facet Extraction

Three of the facets we intended to capture were to be

expressed quantitatively in our review characteristics

data set. We found that age, gender and sentiment

were most of the times expressed quantitatively how-

ever sometimes those were expressed qualitatively.

e.g.: for age, instead of “30-40”, the output would

be “The author’s age is likely to be between 30 and

40”. This was highly unusual despite setting GPT-3’s

temperature to 0, which was said to produce entirely

deterministic results.

To work around this challenge, we developed

intermediate analysis scripts using natural language

processing tools such as NLTK. The intermediate

scripts were used to extract the key points from each

of the qualitative responses. Where any qualitative

responses lacked key words such as age, gender or

sentiment scores, it was assumed that there were none

found for that review.

Another challenge was that, some facets review

characteristics were empty. This is most likely due to

two main reasons. The ﬁrst being that requests to the

OpenAI API could have failed, despite the back-off

mechanism working successfully. The second, being

that GPT-3 outputted a result that could not be cap-

tured as a valid review characteristic. For example,

GPT-3 could output “There is not enough informa-

tion to determine the author’s gender”. As a result,

the author’s gender would be marked as null and dis-

counted from further analysis. 3.28% of review facets

analysed reported null values.

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

704

4.2 Persona Development

We developed eight personas for the eight categories

of e-commerce applications. In order to develop the

Pharmacy persona in Figure 3, we began by identify-

ing the most common age and gender of reviewers of

the pharmacy mobile applications that we analysed.

We found that this was 40-year-old men. Hence, our

persona was given this gender and age.

To create this persona’s pains and gains, we se-

lected common expectations and frustrations which

were generated from relevant apps. This included that

the pharmacy should stock NDSS products and that

they should be able to add a favourite store to get the

product.

Given this information, we were able to determine

what this person hears, thinks, feels, sees, says and

does. This included giving him a medical condition

(diabetes) that would require NDSS products. Addi-

tionally, we outlined his occupation which might lead

to him expecting certain features such as the ability to

add a favourite store. Also, we believe that in order to

maintain a job and have a condition like this, one most

likely would be quite stoic, think logically, and be in-

active and introverted. While these were assumptions

made by us, and may not be precise, they are guided

by the information we extracted and add to the char-

acter of the developed persona. Consequently, the de-

veloped persona has increased empathy from devel-

opers and allows them to gain a deeper understanding

of the end user.

Due to space limitations, all the personas are not

presented in this article. Another example persona is

the departmental store user persona - Makaylah 4.

5 EVALUATION

Assessing the accuracy of our results, and hence the

accuracy of our subsequently developed personas re-

quires consideration of two aspects. Firstly, we must

assess the accuracy of GPT-3. To do so we must de-

termine the precision and accuracy of GPT-3’s Natu-

ral Language Processing model. Secondly, we need

to ensure that our extracted reviews have enough data

in them to extract the information we need, and the

quality of discussion within these reviews needs to be

high enough to add value to our persona.

GPT-3 has extracted four facets for us, therefore

we must assess the accuracy of each of these aspects.

The four aspects were the reviewers age, gender, sen-

timent, frustrations, and expectations. Thankfully

these tests have already been conducted which ver-

ify the accuracy of transformer-based models. Since

Figure 3: A user persona of Joseph.

Figure 4: A user persona of Makaylah.

GPT-3 is the largest and most advanced transformer-

based model at the time of writing we can assume

the ﬁndings of previous studies would carry over. In

particular we can look at the PAN Author Proﬁling

task. As discussed earlier this competition proved,

through 2016-2019 (Rangel et al., 2016; Rangel et al.,

2017; Rangel et al., 2018; Rangel and Rosso, 2019),

that age, gender and sentiment could be determined

from twitter tweets. Therefore, we know that we are

able to determine implicit information such as age

An Innovative Approach to Develop Persona from Application Reviews

705

and gender from short term text. Hence, GPT-3 can

be utilized to estimate a reviewer’s demographics and

sentiment. Next is determining the explicit informa-

tion of the reviewer’s expectations and frustrations.

Since this information is explicit, to determine GPT-

3’s accuracy, we would simply have to compare the

expectation/frustration against the review and deter-

mine whether that aspect was present. Therefore, our

veriﬁcation for this process involved sampling 100 re-

views and manually identifying their expectations and

frustrations. We then analysed the expectations and

frustrations derived by GPT-3. We found that in 87%

of cases, the expectations and frustrations manually

identiﬁed aligned with those identiﬁed by GPT-3.

Apart from the accuracy of GPT-3, we also had

to ensure that there was enough detail in our reviews

to afford GPT-3 the highest probability of extracting

accurate information. Therefore, to assess the qual-

ity of our reviews, we used information power. In

doing so we would assess our extracted reviews, be-

fore being passed into GPT-3, against ﬁve criteria

items: “study aim, sample speciﬁcity, use of estab-

lished theory, quality of dialogue, and analysis strat-

egy” (Malterud et al., 2016). Firstly, the study aim in

this case was to develop user personas from app re-

views. As a persona consists a user’s demographics

and expectation/frustration and the nature of a review

is for a user to express their expectation/frustration,

and we have determined that we can accurately ﬁnd

their demographics, hence the information power is

higher. Secondly, our sample speciﬁcity was sparse

as situations occur where insufﬁcient data was pro-

vided in the review, hence this lowers our information

power. Thirdly, for established theory, we considered

the quality of dialog or in other words how detailed

the reviews were. In many cases we found that re-

views were one sentence at maximum, hence we had

a very low quality of dialog. This severely impacted

the information power. Lastly for analysis strategy,

considering all of the above, we opted to conduct

cross-case studies. This involved collating all of the

data in each app into one or two identiﬁable personas.

Again, this reduced our information power. There-

fore, with these considerations, identifying the most

prevalent age, gender, sentiment, frustrations, and ex-

pectations, and creating larger, overarching categories

for the apps was the strategy we followed. In doing

so, our developed personas will more accurately rep-

resent the most prominent aspects we could extract.

To further access the accuracy of our developed

personas, we can analyse how well it resembles ex-

isting personas that developers have already devel-

oped. We collected some e-commerce personas and

compared those with the ones we developed. For ex-

ample, the retail customer persona of “Suggestible

Sally” (biz, 2015) shared many commonalities with

our developed Makaylah persona, seen in Figure 4,

who represented a department store shopper. Some

key items were the positive response to marketing and

engagement with support staff. Therefore, we can

have higher conﬁdence in the information we have ex-

tracted from our reviews.

6 THREATS TO VALIDITY

Although our methods produced reasonable results,

there are a few threats to the validity of our ﬁndings.

6.1 Internal

Due to the training data set used for GPT-3, the re-

sults will contain some bias. Given that GPT-3 is

trained on almost all of the public internet, it will con-

tain the biases found in public web content (Mathews

et al., ; Dale, 2021; Ashraf and Perez, 2020). Further-

more, we were only able to analyse a sample set of

4931 reviews from various categories of mobile apps.

This implies that our ﬁndings may not apply to all e-

commerce applications. Moreover, some of the GPT-

3 analysis has required further manual analysis of the

data set. This has caused potential for further internal

errors on the data set due to incorrect classiﬁcation

of information. As we are using the PATHY persona

development technique (Ferreira et al., 2016), there

is some subjective analysis of information. However,

we have taken steps to ensure these risks are mini-

mized by automating as much analysis as possible.

Moreover, we have also ensured that all personas were

reviewed and agreed upon by all researchers.

6.2 External

In terms of the external validity of our research, this

method is not directly transferable to be used on re-

views written in languages other than English. How-

ever, the tooling we used to extract reviews can be

used to obtain reviews from app stores across the

world, and GPT-3 is capable of identifying the lan-

guage in which a text is written (Chiriatti, 2020).

Therefore the tooling we used can be extended to

work with reviews in different languages, however, it

would require changes to the methodology.

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

706

7 DISCUSSION

Implication for Software Development. To improve

e-commerce applications and ensure that end users

are satisﬁed with their experience, it is beneﬁcial to

understand the characteristics and pain points of users

that might effect their use of the platform. Grundy

states that characteristic factors of users should be in-

corporated in engineering and design requirements at

an early stage to ensure the success of the product af-

ter its development life cycle (Grundy, 2021). It is

also supported by Kose and Arslan that when these

characteristics are taken into account, it can directly

impact the use of e-commerce and online expendi-

ture, showing how particular facets had a blunt cor-

relation to them (Kose and Arslan, 2020). To assist

e-commerce developers in understanding their end

users more extensively, we could extract certain facets

from the user application reviews of the platforms be-

ing assessed using GPT-3, and segment the responses

based on their characteristics to investigate possible

trends between the sentiment and user facets. We de-

veloped eight personas based on the ﬁndings. We be-

lieve this will help developers discern what types of

users are avidly or apathetically using their applica-

tion, which can lead to further investigation on how

the sentiment for users that show lower interest in the

platform can be restored. This will also be beneﬁcial

for developers of new e-commerce platforms to un-

derstand the diverse user facets and their needs.

Implication for Research. The research presents an

important perspective of collecting enduser data and

effectively using those for software development. We

conducted a small-scale research with one selected

domain such as - e-commerce. The ﬁndings of this ap-

proach indicate a number of things - ﬁrstly, the use of

existing machine learning tools (eg. GPT-3) make it

easy, efﬁcient and less time consuming for analysing

small-scale app reviews in order to understand end

users’ facets, frustrations and expectations. The initial

promising results indicate that a customised machine

learning algorithm can be also developed for this pur-

pose if needed. Secondly, it was apparent that the

app reviews contain a lot of information about end

users, especially their expectations and frustrations

about the software. This can be an excellent source

of information about the end users. This framework

can be further investigated and developed as a fully

automated tool that can give developers information

about their end users.

8 CONCLUSIONS

This research has shown that it is possible to de-

velop user personas from e-commerce mobile appli-

cation reviews. By using the PATHY persona devel-

opment technique and the GPT-3 machine learning

model, we were able to extract user facets, expecta-

tions and frustrations from a data set of 4931 Aus-

tralian e-commerce app reviews. These ﬁndings can

be used by software developers to better understand

the users of their platforms. However, there are some

threats to the validity of these ﬁndings, which should

be considered in future work.

There are a number of ways in which this research

can be extended. Firstly, by amending the methodol-

ogy, we could extend this model to work with reviews

written in different languages. This ability would al-

low us to work with a more diverse data set, resulting

in more widely-application ﬁndings. Secondly, with

a larger sample size, the validity of the research can

be improved. If more reviews from different mobile

applications are used to develop personas, it would

lead to stronger results. Thirdly, our research takes

a holistic approach to developing personas from mo-

bile application reviews. More speciﬁc research can

be conducted by sorting results by each application.

This would allow us to investigate how personas re-

late to speciﬁc applications, and potentially see trends

in personas from similar application categories. Fi-

nally, while we focused on e-commerce mobile ap-

plication reviews, this model could be extended to

process and analyse reviews from different platforms

such as web applications.

ACKNOWLEDGEMENTS.

Kanij is supported by ARC Laureate Fellowship

FL190100035.

REFERENCES

(2015). The eight personas of retail customer bases.

Acheampong, F. A., Nunoo-Mensah, H., and Chen, W.

(2021). Transformer models for text-based emotion

detection: a review of bert-based approaches. The Ar-

tiﬁcial intelligence review, 54(8):5789–5829.

Acu

na, S. T., Castro, J. W., and Juzgado, N. J. (2012). A hci

technique for improving requirements elicitation. Inf.

Softw. Technol., 54:1357–1375.

Alexandridis, G., Varlamis, I., Korovesis, K., Caridakis, G.,

and Tsantilas, P. (2021). A survey on sentiment anal-

ysis and opinion mining in greek social media. Infor-

mation (Basel), 12(8):331.

An Innovative Approach to Develop Persona from Application Reviews

707

Ashraf, Muhammad Adnan, R. M. A. N. F. N. D. P. V. S. and

Perez, F. (2020). A study of deep learning methods for

same-genre and cross-genre author proﬁling. Journal

of Intelligent & Fuzzy Systems, 39(15):2353–2363.

Bratsberg, H. M. (2012). Empathy maps of the foursight

preferences.

Chiriatti, L. F. . M. (2020). Gpt-3: Its nature, scope, limits,

and consequences. Minds & Machines, 30(1):681–

694.

Cooper, A. (1999). The inmates are running the asylum:

Why high-tech products drive us crazy and how to re-

store the sanity pearson education.

Dale, R. (2021). Gpt-3: What’s it good for? Natural lan-

guage engineering, 27(1):113–118.

Dale, R. (2022). Nlp: How to spend a billion dollars. Nat-

ural Language Engineering, 28(1):125–136.

El Shamy, N. and Hassanein, K. (2018). The Impact of Age

and Cognitive Style on E-Commerce Decisions: The

Role of Cognitive Bias Susceptibility. In Davis, F. D.,

Riedl, R., vom Brocke, J., L

eger, P.-M., and Ran-

dolph, A. B., editors, Information Systems and Neu-

roscience, Lecture Notes in Information Systems and

Organisation, pages 73–83, Cham. Springer Interna-

tional Publishing.

Ferreira, B., Santos, G., and Conte, T. (2017). Identify-

ing Possible Requirements using Personas - A Quali-

tative Study:. In Proceedings of the 19th International

Conference on Enterprise Information Systems, pages

64–75, Porto, Portugal. SCITEPRESS - Science and

Technology Publications.

Ferreira, B. M., Barbosa, S. D., and Conte, T. (2016). Pa-

thy: Using empathy with personas to design applica-

tions that meet the users’ needs. In International Con-

ference on Human-Computer Interaction, pages 153–

165. Springer.

Gao, C., Zeng, J., Lyu, M. R., and King, I. (2018). On-

line app review analysis for identifying emerging is-

sues. In Proceedings of the 40th International Confer-

ence on Software Engineering, ICSE ’18, pages 48–

58, New York, NY, USA. Association for Computing

Machinery.

Gray, D., Brown, S., and Macanufo, J. (2010). Gamestorm-

ing. O’Reilly Media, Sebastopol, CA.

Grundy, J. (2021). Impact of End User Human Aspects

on Software Engineering:. In Proceedings of the 16th

International Conference on Evaluation of Novel Ap-

proaches to Software Engineering, pages 9–20, On-

line Streaming, — Select a Country —. SCITEPRESS

- Science and Technology Publications.

Grundy, J., Mouzakis, K., Vasa, R., Cain, A., Curums-

ing, M., Abdelrazek, M., and Fernando, N. (2018).

Supporting Diverse Challenges of Ageing with Digital

Enhanced Living Solutions. Telehealth for our Ageing

Society, pages 75–90. Publisher: IOS Press.

Guo, A. and Ma, J. (2018). Archetype-Based Modeling

of Persona for Comprehensive Personality Computing

from Personal Big Data. Sensors, 18(3):684.

Kose, T. and Arslan, S. (2020). Turkish consumer participa-

tion in e-commerce. Journal of electronic commerce

in organizations, 18(4):30–50.

Malgaonkar, S., Licorish, S. A., and Savarimuthu, B.

T. R. (2022). Automatically generating taxonomy for

grouping app reviews — a study of three apps. Soft-

ware Quality Journal, 30(2):483–512.

Malterud, K., Siersma, V. D., and Guassora, A. D. (2016).

Sample size in qualitative interview studies: Guided

by information power. Qualitative Health Research,

26(13):1753–1760. PMID: 26613970.

Mao, J.-Y., Vredenburg, K., Smith, P. W., and Carey, T.

(2005). The state of user-centered design practice.

Communications of the ACM, 48(3):105–109.

Marr, K. (2020). How to Create Buyer Personas for Your

Online Store.

Mathews, C., Ye, K., Grozdanovski, J., Marinelli, M.,

Zhong, K., Khalajzadeh, H., Obie, H., and Grundy,

J. AH-CID: A Tool to Automatically Detect Human-

Centric Issues in App Reviews. page 12.

Mcilroy, S., Ali, N., Khalid, H., and Hassan, A. E. Analyz-

ing and automatically labelling the types of user issues

that are raised in mobile app reviews | SpringerLink.

McIntosh, J., Du, X., Wu, Z., Truong, G., Ly, Q., How,

R., Viswanathan, S., and Kanij, T. (2021). Evaluating

Age Bias In E-commerce. In 2021 IEEE/ACM 13th

International Workshop on Cooperative and Human

Aspects of Software Engineering (CHASE), pages 31–

40, Madrid, Spain. IEEE.

Obie, H. O., Hussain, W., Xia, X., Grundy, J., Li, L.,

Turhan, B., Whittle, J., and Shahin, M. (2021). A First

Look at Human Values-Violation in App Reviews. In

2021 IEEE/ACM 43rd International Conference on

Software Engineering: Software Engineering in So-

ciety (ICSE-SEIS), pages 29–38, Madrid, ES. IEEE.

OpenAI. openai: Python client library for the OpenAI API.

Rangel, F., Montes-y-G

omez, M., Potthast, M., and Stein,

B. (2018). Overview of the 6th Author Proﬁling

Task at PAN 2018: Cross-domain Authorship Attri-

bution and Style Change Detection. In Cappellato, L.,

Ferro, N., Nie, J.-Y., and Soulier, L., editors, CLEF

2018 Evaluation Labs and Workshop – Working Notes

Papers, 10-14 September, Avignon, France. CEUR-

WS.org.

Rangel, F. and Rosso, P. (2019). Overview of the 7th Author

Proﬁling Task at PAN 2019: Bots and Gender Pro-

ﬁling. In Cappellato, L., Ferro, N., Losada, D., and

uller, H., editors, CLEF 2019 Labs and Workshops,

Notebook Papers. CEUR-WS.org.

Rangel, F., Rosso, P., Potthast, M., and Stein, B. (2017).

Overview of the 5th Author Proﬁling Task at PAN

2017: Gender and Language Variety Identiﬁcation in

Twitter. In Cappellato, L., Ferro, N., Goeuriot, L., and

Mandl, T., editors, Working Notes Papers of the CLEF

2017 Evaluation Labs, volume 1866 of CEUR Work-

shop Proceedings.

Rangel, F., Rosso, P., Verhoeven, B., Daelemans, W., Pot-

thast, M., and Stein, B. (2016). Overview of the

4th Author Proﬁling Task at PAN 2016: Cross-Genre

Evaluations. In Balog, K., Cappellato, L., Ferro, N.,

and Macdonald, C., editors, Working Notes Papers of

the CLEF 2016 Evaluation Labs, volume 1609 of Lec-

ture Notes in Computer Science.

Singh, R., Garg, V., and GPT-3 (2021). Human Factors in

NDE 4.0 Development Decisions. Journal of Nonde-

structive Evaluation, 40(3):71.

ENASE 2023 - 18th International Conference on Evaluation of Novel Approaches to Software Engineering

708