The Application of Affective Measures in Text-Based Emotion Aware

Recommender Systems

John Kalung Leung

, Igor Griva

, William G. Kennedy

, Jason M. Kinser

Sohyun Park

and Seo Young Lee

Computational Sciences and Informatics, Computational and Data Sciences Department, George Mason University Korea,

119-4 Songdomunhwa-ro, Yeonsu-gu, Incheon, 21985, Korea

Department of Mathematical Sciences, George Mason University,4400 University Drive, Fairfax, Virginia 22030, U.S.A.

Center for Social Complexity, Computational and Data Sciences Department, College of Science, George Mason

University, 4400 University Drive, Fairfax, Virginia 22030, U.S.A.

Computational Sciences and Informatics, Computational and Data Sciences Department, College of Science, George

Mason University, 4400 University Drive, Fairfax, Virginia 22030, U.S.A.

Department of Communications, George Mason University Korea, 119-4 Songdomunhwa-ro, Yeonsu-gu, Incheon, 21985,

Korea

Keywords: Emotion Aware Recommender Systems, Affective Computing, Users and Items Emotion Profiles, Text-

Based Emotion Detection and Recognition, Affective Indices and Affective Index Indicators, Emotion

Identification.

Abstract: This paper presents an innovative approach to address the problems researchers face in Emotion Aware

Recommender Systems (EARS): the difficulty and cumbersome collecting voluminously good quality

emotion-tagged datasets and an effective way to protect users’ emotional data privacy. Without enough good-

quality emotion-tagged datasets, researchers cannot conduct repeatable affective computing research in EARS

that generates personalized recommendations based on users’ emotional preferences. Similarly, if we fail to

protect users’ emotional data privacy fully, users could resist engaging with EARS services. This paper

introduced a method that detects affective features in subjective passages using the Generative Pre-trained

Transformer Technology, forming the basis of the affective index and Affective Index Indicator (AII).

Eliminate the need for users to build an affective feature detection mechanism. The paper advocates for a

Separation of Responsibility approach where users protect their emotional profile data while EARS service

providers refrain from retaining or storing it. Service providers can update users’ affective indices in memory

without saving their privacy data, providing affective-aware recommendations without compromising user

privacy. This paper offers a solution to the subjectivity and variability of emotions, data privacy concerns,

and evaluation metrics and benchmarks, paving the way for future EARS research.

1 INTRODUCTION

The Emotion Aware Recommender System (EARS)

aims to provide personalized recommendations by

considering users' affective preferences and opinions

from similar users. However, researching EARS

https://orcid.org/0000-0003-0216-1134

https://orcid.org/0000-0002-2291-233X

https://orcid.org/0000-0001-9238-1215

https://orcid.org/0000-0003-0078-4899

https://orcid.org/0000-0002-1231-5662

https://orcid.org/0000-0001-5867-0726

faces several challenges due to human emotions'

subjective and variable nature (Qian et al., 2019).

These challenges involve the development of

accurate models for emotion detection, classification,

and prediction, as well as collecting sufficient

emotion-tagged datasets, which are hindered by the

subjective and contextual aspects of emotions (Schedl

590

Leung, J., Griva, I., Kennedy, W., Kinser, J., Park, S. and Lee, S.

The Application of Affective Measures in Text-Based Emotion Aware Recommender Systems.

DOI: 10.5220/0012143900003541

In Proceedings of the 12th International Conference on Data Science, Technology and Applications (DATA 2023), pages 590-597

ISBN: 978-989-758-664-4; ISSN: 2184-285X

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

et al., 2018). Additionally, the ethical considerations

surrounding user privacy and data protection pose a

significant challenge in EARS research (Bobadilla et

al., 2013). In order to evaluate the effectiveness of

EARS, we require robust evaluation metrics and

benchmarks. However, the definition and

measurement of the impact of emotions on user

behavior and satisfaction are complex and

multifaceted (Mohammad, 2016). Therefore, this

paper proposes an innovative approach to address the

following challenges faced by EARS researchers: the

difficulty of collecting voluminous, high-quality

emotion-tagged datasets to enhance affective

computing and the need for an effective method to

safeguard users' emotional data privacy.

In our envisioned world, affective metric

signatures known as affective indices assign objects

derived from subjective descriptions. Unless the

subjective descriptions are revised, these affective

indices remain static. However, users' affective

indices are dynamic and evolve through their

interactions and consumption of objects.

Our research introduces a novel approach to

tackle EARS challenges and enhance their

effectiveness. We address the difficulties associated

with collecting sufficient high-quality emotion-

tagged datasets and ensuring privacy, which impede

research in affective computing for personalized

recommendations based on users' emotions.

We leverage Generative Pre-trained Transformer

(GPT) technology to eliminate users' need to develop

affective feature detection mechanisms. GPT

effortlessly detects affective features in subjective

passages.

We advocate employing the affective index and

Affective Index Indicator (AII) as the foundation for

detecting affective features and measuring emotions.

We reckon that preserving the privacy of emotional

data is crucial to prevent user resistance. We advocate

for a Separation of Responsibility (SoR) approach,

where users are responsible for protecting their

emotional profile data while service providers refrain

from storing it. The service providers can ensure

efficient and personalized recommendations by

updating users' affective indices in memory without

compromising privacy.

Our research provides solutions to address

subjectivity, variability of emotions, data privacy

concerns, and evaluation metrics. These solutions

contribute to the advancement of EARS research and

its practical implementation.

2 RELATED WORKS

The main challenge in Text-based Emotion Aware

Recommender Systems (EARS) research is obtaining

high-quality, emotion-tagged datasets necessary for

machine learning processing. To address this

challenge, Guo (Guo, 2022) illustrated a deep

learning-assisted semantic text analysis approach that

involves defining emotion keywords, identifying data

sources, developing a collection plan, cleaning and

pre-processing data, and evaluating and refining the

dataset. However, researchers still need more high-

quality emotion-labeled datasets, which are required

to train the emotion prediction model for

classification. While standards Recent advances in

transformer-based models, such as Generative Pre-

trained Transformer (GPT) technology, have yet to

establish benchmarking datasets for generating,

PaLM, GPT-3, ChatGPT, BERT, ELMO, RoBERTa,

and Transformer-XL offer promising new approaches

to dataset collection (Ethayarajh, n.d.). These models

have been trained on large amounts of text data to

generate emotion labels (Kusal et al., 2022). They

may provide valuable resources for researchers in this

field.

2.1 Affective Tagged Datasets

EARS needs voluminously good quality emotional

tags datasets for model training of making

personalized recommendations. EARS requires

Emotional tags to refer to labeling data, such as

Ekman’s six basic emotions: happiness, sadness,

anger, fear, surprise, and disgust (Humintell, 2020).

These tags are crucial in developing EARS

technology that can provision users’ and items’

emotional profiles. However, collecting accurate and

sufficient emotional tags can be extremely

challenging due to subjective and variable emotions

(Russell, 2003), lack of standard tagging methods,

time and resource intensity (Lo et al., 2017), privacy

issues (Kompan et al., 2015), limited diversity (Yang

et al., 2021), and contextual factors (Kanjo et al.,

2015). Therefore, collecting emotional tags for EARS

research demands thoughtful planning and

consideration of these factors (Mauss & Robinson,

2009a) in data features engineering.

2.2 Challenges of Affective Index

Modeling in Personalized

Recommender Systems

Recommender systems are crucial in modeling users'

emotional profiles and item preferences. The

The Application of Affective Measures in Text-Based Emotion Aware Recommender Systems

591

affective index, quantifying human emotion metrics

using probabilistic values, has emerged as a valuable

tool to achieve this goal (Acheampong et al., n.d.). In

recent studies, Leung et al. (Leung et al., 2021)

proposed the Affective Index Indicator (AII), which

employs the Cosine Similarity metrics to generate a

list of peer-wise numerical similarity values.

𝐼𝑛𝑛𝑒𝑟



𝑥,𝑦



=𝑥



𝑦



⟨

𝑥,𝑦

⟩



𝐶𝑜𝑠𝑆𝑖𝑚



𝑥,𝑦



∑

𝑥



𝑦





∑

𝑥









∑

𝑦







⟨

𝑥,𝑦

⟩



𝑥



𝑦







This computation enables the measurement of

affective indices between an active user or item and

its peers (Leung et al., 2020b), utilizing either the

Nearest Neighbor algorithm (Keller et al., 1985) or

alternative methodologies.

Over time, AII approaches have undergone

significant advancements. Initially based on simple

word lists (Kratzwald et al., 2018), they have evolved

to incorporate more sophisticated techniques such as

machine learning (Nasir et al., 2020) and lexicon-

based methods (Hajek et al., 2020). Notably, recent

research efforts have aimed to combine multiple

techniques, resulting in a more robust AII (Naseem et

al., 2020). However, the utilization of AII in

recommender systems poses challenges regarding

reliability, interpretability, and ethical considerations,

particularly regarding privacy, stereotypes, and

biases (Mauss & Robinson, 2009b).

Ongoing research focuses on developing more

accurate and effective methods, identified by Leung

et al. (Leung et al., 2020a), for measuring emotions in

the text to address these challenges. These

advancements can significantly enhance the

relevancy and accuracy of personalized EARS. By

leveraging the advances in measuring emotions,

recommender systems can offer highly tailored

recommendations, improving user satisfaction and

engagement.

3 METHODOLOGY

3.1 Generate Emotion Tagged Data in

EARS Research Through GPT

This paper proposes leveraging the extensive GPT

database to obtain emotion-tagged data for EARS

research through short prompting conversational

dialogue. Researchers can query GPT for the

affective indices of subjective texts, utilizing massive

domain information when available. Using GPT in

this way can easily extract affective indices from

object descriptive passages and potentially become

the standard method for gathering emotional labels.

Emotion Aware Recommender Systems (EARS)

employ affective indices and Affective Index

Indicators (AII) to measure emotions in users and

items. AII calculates the similarity between a source

object and target objects based on Ekman's six basic

human emotions (Leung et al., 2020b). However,

using AII poses challenges regarding reliability,

interpretability, and avoiding biases. Further research

is needed to enhance AII. Artificial intelligence

language models like ChatGPT have shown promise

in sentiment and affective analysis of a text. This

study shows how the researchers prompt ChatGPT to

build an affective index from the inspirational text.

To analyze sentiment in subjective passages, one

must perform pre-processing, which involves noise

removal, tokenization, and polarity assignment.

Various sentiment analysis libraries and tools are

available for this purpose. By aggregating the polarity

scores, we can obtain an overall AII that we can

utilize in downstream applications, such as user

filtering and personalized recommendations.

ChatGPT can estimate the probabilistic values of

basic emotions in subjective passages using sentiment

analysis techniques and lexicons or machine learning

models (Munn et al., 2023). The scores are

normalized to create an affective index. While

sentiment analysis is imperfect, our method provides

a helpful guide for assessing emotional content

(Lauriola et al., 2022). The AII and affective index

offer valuable tools for recommendation systems and

emotional filtering.

3.2 The Separation of Responsibility

Framework Principles

We strongly advocate for protecting users' emotional

data privacy through a Separation of Responsibility

(SoR) framework. Here is how it works:

• Each object in the system is assigned an

emotion ID (eID) based on its affective indices.

• The eID of an item object remains static, while

a user's eID is dynamic, evolving with their

interactions.

• Users are responsible for managing their eID

under the SoR framework.

• Service Providers manage the eIDs of items

without storing users' eIDs.

• Users can choose to provide their eID for

EARS' affective services.

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

592

• To maintain privacy, users can download an

app to rerank EARS' top-N recommendations

using the app's Affective Index Indicator

algorithm.

• By implementing the SoR framework, we

protect users' data privacy, ensuring the

confidentiality and security of their emotional

information.

This SoR approach safeguards emotional data privacy

and empowers users with greater control over their

information within the EARS system.

4 IMPLEMENTATION

This section outlines a methodology for computing

affective indices for users and objects, utilizing

MovieLens datasets and movie content obtained from

The Movie Database (TMDb) and the Internet Movie

Database (IMDb) through web scraping. Our

approach extends our previous work (Leung et al.,

2021), which detailed the intricate process involved

in this computation. To demonstrate the efficacy of

our method, we utilize the affective profile of User

400 from our prior study. Here is a snapshot of User

400's affective profile at a specific moment.

Table 1: User 400 Affective Indices.

Happiness Ange

Sadness

0.08874 0.11934 0.12709

Fea

Sur

rise Dis

ust

0.20332 0.13918 0.15881

In addition, we introduce a novel and

straightforward approach to obtaining affective

indices for objects using ChatGPT short prompting.

Initially, we acquired the "Top 100 Movies" list from

IMDb (IMDb, 2020) and utilized ChatGPT to

estimate the affective indices of each movie based on

their respective generated movie plots (OpenAI,

2023).

Table 2: Affective Indices of IMDb Top 100 Movies.

Rank Particle List of

IMDb Top 100

Movies

Affective Indices

Happiness, Anger, Sadness,

Fear, Surprise, Disgust

1 The Godfather 0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

2 The Shawshank

Redemption

0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

3 Schindler's List 0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

4 Raging Bull 0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

5 Casablanca 0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

6 Citizen Kane 0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

7 Gone with the

Win

0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

8 The Wizard of

0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

9 One Flew Over

the Cuckoo's

Nest

0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

10 Lawrence of

Arabia

0.28792, 0.12743, 0.14890,

0.17025, 0.11302, 0.15248

96 Rear Window 0.13660, 0.11584, 0.13170,

0.14871, 0.12121, 0.34594

97 The Third Man 0.13902, 0.11464, 0.13434,

0.15794, 0.11726, 0.33780

98 Rebel Without

a Cause

0.13791, 0.11650, 0.13319,

0.15038, 0.12095, 0.34107

99 North by

Northwest

0.13979, 0.11310, 0.13589,

0.15855, 0.11611, 0.33656

100 Yankee Doodle

Dandy

0.13189, 0.12007, 0.12796,

0.14716, 0.12588, 0.34714

4.1 Estimating Probabilistic Affective

Indices with ChatGPT

We use GPT Short Prompting to gather emotion-

tagged datasets effectively (OpenAI, 2023). Here is

an example dialogue:

• Does the user ask GPT for the movie plot

"The Godfather I"?

• GPT response and listed the found movie

plot:

• "The Godfather" (1972) is a renowned crime

drama film directed by Francis Ford

Coppola, based on Mario Puzo's...”

• The user then asks GPT to estimate Ekman's

six basic emotions in probabilistic values to

a four-significant digit accuracy for the

movie plot.

• GPT listed the estimated Ekman's six basic

human emotions in probabilistic values of

"The Godfather" (1972), with the user’s

specified precision:

The Application of Affective Measures in Text-Based Emotion Aware Recommender Systems

593

Figure 1: Godfather I Movie Plot Affective Indices.

The affective index expresses the probabilistic values

of detected emotions. We use OpenAI’s GPT-3 API

(OpenAI, 2023) to analyze the plot’s emotions and

extract the scores for each primary emotion as

suggested (Dale, 2021). We asked GPT-3 to

normalize the scores and compute the probabilistic

values

of Ekman’s six basic human emotions scores

(OpenAI, 2023). Our example shows a high intensity

of anger, followed by fear. Although the film

contained moderate-intensity happiness scenes, it

also exhibited high levels of disgust, surprise, and

sadness in the movie plot. However, the affective

indices may vary based on the algorithm employed

for the prediction model and other factors.

4.2 Affective Index Indicator in EARS

Affective Index Indicator (AII) is a metric used in

Emotion Aware Recommender Systems (EARS) to

measure the emotional content of textual data. It

reflects the intensity of emotion words expressed in a

text, such as happiness, sadness, anger, fear, surprise,

and disgust. Various natural language processing

techniques analyze the sentiment and emotion in a

text to calculate the AII (Tsytsarau & Palpanas,

2012). By considering the emotional preferences of

users and items, recommender systems can provide

personalized recommendations that better-fit users’

needs and preferences (Chang & Hsing, 2021). AII is

valuable for building EARS because it allows the

system to consider the emotional profile of items and

an active user’s affective preferences in making

recommendations. However, AII is one of many

approaches to building EARS, and its effectiveness

may vary depending on the specific application and

user context. When designing EARS, designers must

consider ethical and privacy considerations to ensure

the system does not perpetuate biases or stereotypes

related to emotions or personal characteristics and to

secure user data.

4.3 Key Principles and Implementation

Guidelines of the Separation of

Responsibility Framework

Recommender systems employ various methods such

as Collaborative Filtering, Content Filtering, Hybrid

approaches, and more to generate top-N

recommendations. It is essential to highlight that

users receive these recommendations through push or

pull services.

Furthermore, it is worth noting that all

recommender systems operate as hybrid filters,

combining multiple techniques to provide the best

possible recommendations for users. Among these

approaches, Collaborative Filtering is the dominant

method in the field, accounting for approximately

80% of recommender systems implementations.

Collaborative Filtering-based recommender systems

often rely on rating systems, such as the 5-star rating

system used by Amazon or the 10-point scale

employed by IMdB for movie ratings. Additionally,

The up-sales feature often influences

recommendations, enhancing the relevance and

personalization of the suggestions with phrases like

"Customers who bought this item also bought that."

The critical points of the Separation of Responsibility

(SoR) framework are as follows:

Figure 2: Separation of Responsibility Framework.

• Every object has its emotion ID (eID).

• Item object’s eID is a statistic.

• The user object’s eID is dynamic.

• SoR user self-manages eID.

• The Service Provider (SP) manages items’ eID.

• Users can give eID for EARS.

• SP does not store users’ eIDs.

• SoR protects user data privacy by not keeping

user data in its operations.

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

594

5 RESULTS

5.1 Detecting Affective Features Using

GPT-NLP Database and Affective

Index and Affective Index Indicator

Method for EARS

This study introduces a new method for detecting

affective features in subjective writing using a

Generative Pre-trained Transformer (GPT) Natural

Language Processing (NLP) database. The affective

index method relies on GPT to create a profile of an

item’s affective features. To derive the user’s

affective profile, one can average the affective indices

of all the consumed objects (Leung et al., 2020c). The

affective index plays a crucial role in recommender

systems, where subjective descriptions are utilized, as

it can determine the similarity between items and

users.

In this section, we present a comprehensive

methodology for computing affective indices for

users and objects using MovieLens datasets and

movie content obtained through web scraping from

The Movie Database (TMDb) and the Internet Movie

Database (IMDb). Our approach builds upon the

work of (Leung et al., 2021), where we extensively

described the laborious and demanding process

involved in this computation. To illustrate the

effectiveness of our method, we leverage the affective

profile of User400 from our previous study.

By guiding ChatGPT (OpenAI, 2023), we

successfully computed User400's Affective Index

Indicator by reordering the Top 100 Movies' affective

indices according to User400's preferences. These

results unequivocally demonstrate the seamless

capabilities of ChatGPT in performing these tasks.

Notably, users are relieved from the burden of

programming, downloading, or scraping webpages,

making our proposed Top-N recommendation's

affective-aware reranking capability easily integrable

into various applications.

Overall, our methodology showcases a

streamlined approach to computing affective indices

for both users and objects, empowered by the

capabilities of ChatGPT. By eliminating the need for

extensive programming or data acquisition efforts,

our method offers a user-friendly and efficient

solution for incorporating affective awareness into

recommendation systems.

Table 3: User 400 Rerank IMDb Top 100 Movie.

Rank Particle List of

IMDb Top 100

Movies

User 400 Rerank with

Affective Index Indicator

1 The Godfathe

The Godfather, 0.95679

2 The Shawshank

Redemption

Pulp Fiction, 0.93456

3 Schindler's List The Shawshank

Redemption, 0.89234

4 Raging Bull Fight Club, 0.87591

5 Casablanca The Dark Knight, 0.86543

6 Citizen Kane The Matrix, 0.84672

7 Gone with the

Win

Inception, 0.82549

8 The Wizard of

Forrest Gump, 0.81236

9 One Flew Over

the Cuckoo's

Nest

Goodfellas, 0.79825

10 Lawrence of

Arabia

The Lord of the Rings: The

Fellowship of the Ring,

0.78219

96 Rear Window The Con

urin

2, 0.49018

97 The Third Man The Hobbit: The

Desolation of Smaug,

0.48949

98 Rebel Without

a Cause

The Jungle Book, 0.48883

99 North by

Northwest

The Help, 0.48820

100 Yankee Doodle

Dandy

The Hobbit: The Battle of

the Five Armies, 0.48760

5.2 Balancing Personalized Services

and User Privacy Data Protection

Through the Separation of

Responsibility Framework

User affective profiles have gained popularity in

applications like personalized marketing and mental

health monitoring, but privacy concerns arise.

Balancing personalized services and privacy

protection is crucial, especially in Emotion Aware

Recommender Systems (EARS). To address these

challenges, we propose a novel approach: a

separation of responsibility (SOR) framework

involving four parties - human users (Us), affective

aware service providers (AASP), product

manufacturers and service providers (PMSP), and

profiles service authority (PSA).

We suggest assigning an emotion ID to all objects,

allowing personalized services while protecting user

privacy. PMSPs obtain emotion IDs for their

The Application of Affective Measures in Text-Based Emotion Aware Recommender Systems

595

products/services through the PSA or partners. Users'

emotion IDs are dynamic based on interactions, while

object emotion IDs remain constant unless subjective

descriptions change.

PMSPs store all objects' emotion IDs for

personalized recommendations without storing users'

emotion IDs. Users retain ownership and control over

their emotion IDs, providing them when requesting

recommendations from AASPs. PMSPs offer

appropriate affective personalized recommendations,

reporting aggregated emotion IDs to the PSA. The

PSA periodically updates and shares the emotion ID

with the user.

Users can also handle affective aware

recommendations through an app without sending

their emotion ID to AASPs. When receiving a top-N

recommendation list from AASP, a user can rerank

the list on their computing device using the app,

ensuring the privacy of the emotion ID data.

Our separation of responsibility framework

ensures users safeguard their private data while

service providers offer personalized services based on

affective profiles.

6 FUTURE WORK

We can explore several options to perform future

studies with other organizations' language models and

conversational agents. Google's Meena is a

transformer-based neural network known for

generating human-like responses. Microsoft's

XiaoIce, on the other hand, emulates a teenage girl's

conversational style and personality. Facebook's

Blender combines rule-based and machine-learning

approaches to generate natural language responses.

Additionally, Amazon's Alexa and Apple's Siri utilize

natural language processing and machine learning to

understand and respond to user requests. Each of

these conversational agents has strengths and

weaknesses, and their effectiveness may vary

depending on the use case.

Other comparable tools include OpenAI's GPT-3,

which surpasses ChatGPT in size and potency.

Google's BERT excels at understanding the

contextual meaning of words and phrases. The Allen

Institute for Artificial Intelligence's ELMO generates

contextualized word embeddings. Facebook's

RoBERTa, on the other hand, is pre-trained on a

larger dataset. Lastly, Carnegie Mellon University

and Google's Transformer-XL are adept at handling

longer text sequences and generating more accurate

predictions for text completion tasks. Although all

these language models have been trained on extensive

text data and employ variations of the transformer

architecture, each model possesses its strengths and

weaknesses, making them suitable for specific tasks

or use cases.

7 CONCLUSION

In conclusion, this paper:

• Presents an innovative approach to address

problems in Emotion Aware Recommender

Systems (EARS).

• Problems include difficulty collecting good

quality emotion-tagged datasets and protecting

users’ emotional data privacy.

• Insufficient datasets hinder affective

computing research for personalized

recommendations based on users’ emotional

states and preferences.

• Introduces method using Generative Pre-

trained Transformer Technology to detect

affective features in subjective passages.

• Using GPT technology eliminates the need for

users to build an affective feature detection

mechanism.

• Introduces affective index and Affective Index

Indicator (AII) as the basis for detecting

affective features and measures.

• Failure to protect emotional data privacy can

lead to user resistance to engaging in affective

services offered by EARS.

• We advocate for separation of responsibility

approach, with users protecting emotional

profile data and service providers refraining

from storing it.

• Service providers can update users' affective

indices in memory without compromising

users’ emotional data privacy.

• We offer solutions to subjectivity and

variability of emotions, data privacy concerns,

and evaluation metrics and benchmarks.

• Paves the way for future EARS research.

REFERENCES

Acheampong, F. A., Nunoo-Mensah, H., & Chen, W. (n.d.).

Transformer models for text-based emotion detection:

A review of BERT-based approaches. Artificial

Intelligence Review, 1--41.

Bobadilla, J., Ortega, F., Hernando, A., & Gutiérrez, A.

(2013). Recommender systems survey. Knowledge-

Based Systems, 46, 109–132.

DATA 2023 - 12th International Conference on Data Science, Technology and Applications

596

Chang, Y.-C., & Hsing, Y.-C. (2021). Emotion-infused

deep neural network for emotionally resonant

conversation. Applied Soft Computing, 113, 107861.

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

Language Engineering, 27(1), 113--118.

Ethayarajh, K. (n.d.). How contextual are contextualized

word representations? Comparing the geometry of

BERT, ELMo, and GPT-2 embeddings. ArXiv Preprint

ArXiv:1909.00512.

Guo, J. (2022). Deep learning approach to text analysis for

human emotion detection from big data. Journal of

Intelligent Systems, 31(1), 113–126.

https://doi.org/10.1515/jisys-2022-0001

Hajek, P., Barushka, A., & Munk, M. (2020). Fake

consumer review detection using deep neural networks

integrating word embeddings and emotion mining.

Neural Computing and Applications, 32, 17259--

17274.

Humintell. (2020). Ekman 7 Emotion Facial Recognition.

https://www.humintell.com/

IMDb. (2020). Top 100 Greatest Movies of All Time (The

Ultimate List). http://www.imdb.com/list/ls055592025/

Kanjo, E., Al-Husain, L., & Chamberlain, A. (2015).

Emotions in context: Examining pervasive affective

sensing systems, applications, and analyses. Personal

and Ubiquitous Computing, 19, 1197--1212.

Keller, J. M., Lyu, M. R., & Bourgeois, J. A. (1985). A

fuzzy k-nearest neighbor algorithm. IEEE Transactions

on Systems, Man, and Cybernetics, 4, 580--585.

Kompan, M., Matz, S. C., Gosling, S. D., Popov, V., &

Stillwell, D. (2015). Facebook as a research tool for the

social sciences: Opportunities, challenges, ethical

considerations, and practical guidelines. American

Psychologist, 70(6), 543.

Kratzwald, B., Ilić, S., Kraus, M., Feuerriegel, S., &

Prendinger, H. (2018). Deep learning for affective

computing: Text-based emotion recognition in decision

support. Decision Support Systems, 115, 24–35.

Kusal, S., Patil, S., Choudrie, J., Kotecha, K., Mishra, S., &

Abraham, A. (2022). AI-based Conversational Agents:

A Scoping Review from Technologies to Future

Directions. IEEE Access.

Lauriola, I., Lavelli, A., & Aiolli, F. (2022). An

introduction to deep learning in natural language

processing: Models, techniques, and tools.

Neurocomputing, 470, 443--456.

Leung, J. K., Griva, I., & Kennedy, W. G. (2020a). An

Affective Aware Pseudo Association Method to

Connect Disjoint Users Across Multiple Datasets – an

enhanced validation method for Text-based Emotion

Aware Recommender. International Journal on

Natural Language Computing (IJNLC) Vol, 9(4).

https://doi.org/10.5121/ijnlc.2020.9402

Leung, J. K., Griva, I., & Kennedy, W. G. (2020b). Making

Use of Affective Features from Media Content

Metadata for Better Movie Recommendation Making.

ArXiv Preprint ArXiv:2007.00636.

Leung, J. K., Griva, I., & Kennedy, W. G. (2020c). Text-

based Emotion Aware Recommender. Proceedings of

International Conference on Natural Language

Computing and AI (NLCAI 2020), 10, 101–114.

https://doi.org/10.5121/csit.2020.101009

Leung, J. K., Griva, I., & Kennedy, W. G. (2021). Applying

the Affective Aware Pseudo Association Method to

Enhance the Top-N Recommendations Distribution to

Users in Group Emotion Recommender Systems.

International Journal on Natural Language Computing

(IJNLC), 10, 1–20.

https://doi.org/10.5121/ijnlc.2021.10101

Lo, S. L., Cambria, E., Chiong, R., & Cornforth, D. (2017).

Multilingual sentiment analysis: From formal to

informal and scarce resource languages. Artificial

Intelligence Review, 48, 499--527.

Mauss, I. B., & Robinson, M. D. (2009a). Measures of

emotion: A review. Cognition & Emotion, 23(2), 209–

237. https://doi.org/10.1080/02699930802204677

Mauss, I. B., & Robinson, M. D. (2009b). Measures of

emotion: A review. Cognition and Emotion, 23(2), 209-

-237.

Mohammad, S. M. (2016). Sentiment analysis: Detecting

valence, emotions, and other affectual states from text.

In Emotion measurement (pp. 201–237). Elsevier.

Munn, L., Magee, L., & Arora, V. (2023). Truth Machines:

Synthesizing Veracity in AI Language Models. ArXiv

Preprint ArXiv:2301.12066.

Naseem, U., Razzak, I., Musial, K., & Imran, M. (2020).

Transformer based deep intelligent contextual

embedding for twitter sentiment analysis. Future

Generation Computer Systems, 113, 58--69.

Nasir, A. F. A., Nee, E. S., Choong, C. S., Ghani, A. S. A.,

Majeed, A. P. A., Adam, A., & Furqan, M. (2020). Text-

based emotion prediction system using machine

learning approach (Vol. 769). IOP Publishing.

OpenAI. (2023). ChatGPT (Mar 14 version). Large

Language Model. https://chat.openai.com/chat

Qian, Y., Zhang, Y., Ma, X., Yu, H., & Peng, L. (2019).

EARS: Emotion-aware recommender system based on

hybrid information fusion. Information Fusion, 46,

141–146.

Russell, J. A. (2003). Core affect and the psychological

construction of emotion. Psychological Review, 110(1),

145.

Schedl, M., Zamani, H., Chen, C.-W., Deldjoo, Y., & Elahi,

M. (2018). Current challenges and visions in music

recommender systems research. International Journal

of Multimedia Information Retrieval, 7, 95--116.

Tsytsarau, M., & Palpanas, T. (2012). Survey on mining

subjective data on the web. Data Mining and

Knowledge Discovery, 24, 478--514.

Yang, K., Wang, C., Sarsenbayeva, Z., Tag, B., Dingler, T.,

Wadley, G., & Goncalves, J. (2021). Benchmarking

commercial emotion detection systems using realistic

distortions of facial image datasets. The Visual

Computer, 37, 1447--1466.

The Application of Affective Measures in Text-Based Emotion Aware Recommender Systems

597